MP4ファイルを保存中に、「なんだこれ。時間かかるな。」と思ってcontrol + Cで中断してしまいました。動画の記録自体は終わっていたので、m2tsファイルみたいにぶった切っても平気だと思っていたのです。完全に愚かでした。
ffmpegに食わせてみると“moov atom not found”と表示されます。調べてみると、MP4ファイルは動画と音声が細切れにごちゃ混ぜに保存されており、それを読み解くテーブルが別に記録されているそうです。私はそのテーブル作成中にアボートしてしまったので、動画自体は残っていますが、どこが動画でどこが音声かわからない代物が出来上がってしまったわけです。
検索の結果、似たような事象に対してPythonスクリプトで解読して復元に成功した例がありました。
とはいえ、そのようなデータを扱うにはフォーマットについて勉強が必要なので、しばらく考えないことにしました。最近、ようやく重い腰を上げて取り組んだところ、バイナリデータに不慣れなため時間はかかりましたが成功したので、後続の方のためにコードを公開します。
AACに絶望
上の例(finsta360.py)では、H.264はNAL Unit、AACはADTSフォーマットで格納されていました。つまり、それぞれのヘッダを見つけることでパースできます。しかしながら、私の場合はH.264がNAL Unitというのは同じでしたが、AACがADTSヘッダのついていない生データ(RDB: Raw Data Block)でした（しかもH.264の間は1ブロックとは限らない）。RDBには以下のようにヘッダらしきものはついていますが、AACはブロックが可変長のため、ブロック境界の判別は簡単な条件分岐では行えません。
[ID Syn Ele (3 bit)][Ele Instance Tag (4)][Bitstream][ID TERM (3)]
RDBの構成（参考）
例えば、ステレオの場合は0x20 or 0x21で始まりますが、その後はいろんなパターンがあり、高々8 bitの目印ではビットストリーム中に容易に現れるため、単純にはいきません。
チャンクの判別
いったんH.264チャンク、AACチャンクの判別について整理します。（説明の順番が変ですね。わかりにくいかも。すみません。）
finsta360.pyでは、ADTSヘッダのあるなしでH.264、AACどちらかを判別しています。私の動画の場合、ADTSヘッダは判別に使えませんが、幸運にもH.264チャンクに特徴的なパターンがありました。
H.264データは以下のようなNAL Unitに格納されています。
[Bitstreamのサイズ(4 Byte)][Bitstream]
このUnitの塊（チャンク）は必ず[0x00000002][0x09F0]で始まっていました！ というわけで以下の戦略で判別を行います。
- 6バイト読み込み、0x209F0だったらH.264チャンク、そうでなかったらAACチャンク
- H.264チャンクの場合、次のNAL Unitまでどんどん飛ぶ。飛んだ先が0x20 or 0x21だったらAACチャンクなので打ち切り、AACチャンクへ。
- AACチャンクの場合、0x209F0まで探す。打ち切り、H.264チャンクへ。
AACチャンクは1チャンクに1サンプル（＝ブロック）とは限りませんが、その判別は後で考えることにします。（finsta360.pyをもとに）コードにしたものが以下です。
recover_sample_tables_from_mdat_fast()改
def recover_sample_tables_from_mdat_fast(filename, verbose=False): mov_table = [] aac_table = [] with open(filename, 'rb') as f_in: # look for 'mdat' src_cur = 0 while True: f_in.seek(src_cur) if f_in.tell() != src_cur: raise ValueError(f'seek failed? {f_in.tell()} != {src_cur}') n, atom_type = read_atom_head(f_in) if atom_type == 'mdat': break src_cur += n # 'mdat' is found mdat_start = src_cur if n == 0: # mdat from impcomplete mp4 file f_in.seek(0, 2) mdat_end = f_in.tell() # seek the data_start position # 8 bytes for the header PLUS 8 bytes for the reserved space of the size # f_in.seek(src_cur + 16) f_in.seek(src_cur) buf = f_in.read(4) head = struct.unpack('>I', buf)[0] if head == 1: f_in.seek(src_cur + 16) else: f_in.seek(src_cur + 8) else: mdat_end = src_cur + n n = 0 while True: cur = f_in.tell() if cur >= mdat_end: break # buf = f_in.read(4) buf = f_in.read(6) # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: # h264 chunk # frame_length = struct.unpack('>I', buf)[0] + 4 # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(4) if (buf[0] & 0b11111110) == 0x20: break frame_length += struct.unpack('>I', buf)[0] + 4 if cur+frame_length >= mdat_end: break f_in.seek(cur + frame_length) if verbose: print(f'{n}: [mov] {cur}, {frame_length}') mov_table.append((cur, frame_length)) f_in.seek(cur+frame_length) else: # buf_2 = f_in.read(2) # from https://wiki.multimedia.cx/index.php/ADTS # AAAAAAAA AAAABCCD EEFFFFGH HHIJKLMM MMMMMMMM MMMOOOOO OOOOOOPP (QQQQQQQQ QQQQQQQQ) # 0th-byte 1st 2nd 3rd 4th 5th 6th (7th 8th ) # 0xFF 0xF1 0x4C 0X80 -- typical case for Insta360 ONE-X # M 13 frame length, this value must include 7 or 9 bytes of header length # FrameLength = (ProtectionAbsent == 1 ? 7 : 9) + size(AACFrame) # frame_length = ((buf[3] & 0b11) << 11) | (buf_2[0] << 3) | (buf_2[1] >> 5) # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(6) # if is_aac_header(buf, frame_length): break if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: break frame_length += 1 if cur + frame_length >= mdat_end: break f_in.seek(cur + frame_length) # if verbose: print(f'{n}: [aac] {cur}, {frame_length}') aac_table.append((cur, frame_length)) f_in.seek(cur+frame_length) n += 1 return mov_table, aac_table
サンプルとは、MP4のstscに記録されるデータ単位のことです。既存の動画を解析したところ、H.264はチャンクにNAL Unitが複数あっても1サンプルみたいですが、AACはブロックそれ自体が1サンプルのようです。（MP4の構造について大変有用なページが以下です。）
また、BoxについてはAppleのAtomのページも役に立ちます。
AACブロック境界の判別
AACも複数のブロックを1サンプルとしても（つまりstscを(1, 1, 1)とする）再生できたりするのでは？と思いました。しかしながら、sample_durationが整数のため、stscの変更に辻褄を合わせた数字（小数になりました）にしても音ズレしてしまうでしょう。
なんとか判別できないかと調べていたら、壊れたAACファイルを修復する製品のページに詳しい資料がありました。AACブロックには6種類のパターンがあるようです。
それをもとに以下のコードを書いてみました。
is_aac_header()
def is_aac_header(buf, frame_length): cpe = (buf[0] & 0b11100000) >> 5 element_instance_tag = (buf[0] & 0b00011110) >> 1 common_window = (buf[0] & 0b00000001) if cpe != 0b001 or element_instance_tag != 0b0000: # not stereo # print('not stereo') return False elif common_window == 1: window_sequence = (buf[1] & 0b01100000) >> 5 if window_sequence != 0b10: # common case always_0 = (buf[1] & 0b10000000) >> 7 maxsfb = ((buf[1] & 0b00001111) << 2) | ((buf[2] & 0b11000000) >> 6) predictor = (buf[2] & 0b00100000) >> 5 gain = ((buf[2] & 0b00000111) << 5) | ((buf[3] & 0b11111000) >> 3) inc = (buf[4] & 0b01111100) >> 2 if always_0 == 0 and 40 <= maxsfb and maxsfb <= 51 and predictor == 0 and 100 <= gain and gain <= 228 and inc != 0: print('common case') return True else: # print('not common case') return False else: # eight-short-sequence always_10 = (buf[1] & 0b01100000) >> 5 maxsfb = (buf[1] & 0b00001111) gain = ((buf[3] & 0b01111111) << 1) | ((buf[4] & 0b10000000) >> 7) if always_10 == 0b10 and maxsfb >= 8 and 100 <= gain and gain <= 200: print('eight-short-sequence') return True else: # print('not eight-short-sequence') return False else: pulse_data = (buf[3] & 0b00001000) >> 3 tns_data = (buf[3] & 0b00000100) >> 2 gain_control_data = (buf[3] & 0b00000010) >> 1 if pulse_data != 0 or tns_data != 0 or gain_control_data != 0: maxsfb = ((buf[2] & 0b00001111) << 2) | ((buf[3] & 0b11000000) >> 6) if maxsfb != 0: gain = buf[1] always_0 = (buf[2] & 0b10000000) >> 7 if 100 <= gain and gain <= 200 and always_0 == 0: window_sequence = (buf[2] & 0b01100000) >> 5 if window_sequence != 0b10: # multiple windows if 48 <= maxsfb and maxsfb <= 51: print('multiple windows') return True else: # print('not multiple windows') return False else: # multiple windows and eight-short-sequence maxsfb = (buf[2] & 0b00001111) if maxsfb >= 8: print('multiple windows and eight-short-sequence') return True else: # print('not multiple windows and eight-short-sequence') return False else: # print('not multiple windows type') return False else: # degenerated block if frame_length == 7: print('degenerated block') return True else: # print('not degenerated block') return False else: # degenerated block with padding # gain = ((buf[1] & 0b01111111) << 1) | ((buf[2] & 0b10000000) >> 7) maxsfb = ((buf[2] & 0b00000111) << 3) | ((buf[3] & 0b11100000) >> 5) maxsfb2 = (buf[5] & 0b00011111) if maxsfb == 0 and maxsfb2 == 0: print('degenerated block with padding') return True else: # print('not degenerated block with padding') return False
ところが、資料を正しく読み解けていないかもしれませんが、どうしてもパターン判別の誤爆が免れませんでした。詰んだなと思いつつ、壊れたAACの復元について調べていたらFAAD2というデコーダーがraw AACファイルをADTSフォーマットのAACファイルに変換できるとのこと！ AACのテーブル作成は諦め、H.264、AACそれぞれを抽出・復元し、あとでffmpegなりを使ってmuxする方針に切り替えることにしました。
AACの復元
というわけで、mdatからH.264チャンクを取り除いたものをファイルヘッダもなしにそのままバイナリに垂れ流すコードを書いてみました（付録）。FAAD2に食わせてみると神！ 復元できました（ADTSフォーマットの.aac）。音程が変でしたが、サンプリングレートを指定するオプションをつけてみたら解消しました。
H.264の復元
H.264の復元には苦労しました。Boxの内容は複数のファイルをリファレンスに、ファイル固有のところだけ修正し、あとはコピペで済ませました。それでも、finsta360.pyの元記事にもあるように、親ボックスのサイズを間違えると全然内容がズレちゃうので大変でした。
Boxの構成もInsta360 ONE Xと違っていたので修正しています。mvhd_durationとか、その他の部分も（私の思うように）直してあります。
というわけで、やっとffmpegに怒られない（video onlyの）ファイルが出来上がりました。
復元結果
上で復元したH.264とAACをmuxしたところ、ほぼ完璧に元動画が復元できました！ 別々に処理したせいか、少し音ズレしています（2時間くらい再生すると目立つ程度）。また、数か所画像が緑に乱れるところがあります。疲れたのでそれらは追いませんでした。
感想
AACのフレーム境界の判別に躍起になっていたため、時間がかかってしまいました。たぶんさらに時間をかけても私には無理だったと思います。早くFAAD2を知っておきたかった。
付録
こちらはGPL 3.0を継承し、その下で公開します。
mov.py
#!/usr/bin/env python # coding: utf-8 # finsta360 - python script to finalize incomplete MP4 of Insta360 ONE-X # https://github.com/kichiki/finsta360 import sys import os.path import struct from datetime import datetime, timedelta import gc #from tqdm import tqdm # ## parsing mp4 def parse_mvhd(buf): # Movie Header Atoms version = buf[0] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] time_scale = struct.unpack('>I', buf[12:16])[0] duration = struct.unpack('>I', buf[16:20])[0] preferred_rate = struct.unpack('>I', buf[20:24])[0] preferred_volume = struct.unpack('>H', buf[24:26])[0] # next 10 bytes are reserved matrix_structure = struct.unpack('>IIIIIIIII', buf[36:72]) preview_time = struct.unpack('>I', buf[72:76])[0] preview_duration = struct.unpack('>I', buf[76:80])[0] poster_time = struct.unpack('>I', buf[80:84])[0] selection_time = struct.unpack('>I', buf[84:88])[0] selection_duration = struct.unpack('>I', buf[88:92])[0] current_time = struct.unpack('>I', buf[92:96])[0] next_track_id = struct.unpack('>I', buf[96:100])[0] print(f'version : {version}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'time scale : {time_scale}') print(f'duration : {duration} / {duration/time_scale} sec / {duration/time_scale/60} min') print(f'preferred_rate : {preferred_rate}') print(f'preferred_volume : {preferred_volume}') print(f'matrix_structure : {matrix_structure}') print(f'preview_time : {preview_time}') print(f'preview_duration : {preview_duration}') print(f'poster_time : {poster_time}') print(f'selection_time : {selection_time}') print(f'selection_duration : {selection_duration}') print(f'current_time : {current_time}') print(f'next_track_id : {next_track_id}') def parse_tkhd(buf): #Track Header Atoms version = buf[0] flags = buf[1:4] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] track_id = struct.unpack('>I', buf[12:16])[0] # next 4 bytes are reserved duration = struct.unpack('>I', buf[20:24])[0] # next 8 bytes is reserved layer = struct.unpack('>H', buf[32:34])[0] alternate_group = struct.unpack('>H', buf[34:36])[0] volume = struct.unpack('>H', buf[36:38])[0] # next 2 bytes are reserved matrix_structure = struct.unpack('>IIIIIIIII', buf[40:76]) track_width = struct.unpack('>I', buf[76:80])[0] track_height = struct.unpack('>I', buf[80:84])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'track_id : {track_id}') print(f'duration : {duration}') print(f'layer : {layer}') print(f'alternate_group : {alternate_group}') print(f'volume : {volume}') print(f'matrix_structure : {matrix_structure}') print(f'track_width : {track_width}') print(f'track_height : {track_height}') def parse_mdhd(buf): #Media Header Atoms version = buf[0] flags = buf[1:4] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] time_scale = struct.unpack('>I', buf[12:16])[0] duration = struct.unpack('>I', buf[16:20])[0] language = struct.unpack('>H', buf[20:22])[0] quality = struct.unpack('>H', buf[22:24])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'time scale : {time_scale}') print(f'duration : {duration} / {duration/time_scale} sec / {duration/time_scale/60} min') print(f'language : {language}') print(f'quality : {quality}') def parse_stsd(buf): #Sample Description Atoms print('DATA:') print_binaries(buf) print(f'size of buf: {len(buf)}') version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') sample_description_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break sample_description_size = struct.unpack('>I', buf[i0:i0+4])[0] data_format = str(buf[i0+4:i0+8], 'utf-8') data_reference_index = struct.unpack('>H', buf[i0+14:i0+16])[0] sample_description_table.append( (sample_description_size, data_format, data_reference_index)) print('%d: size: 0x%X, format: %s, ref_index: 0x%X' % ( i, sample_description_size, data_format, data_reference_index)) def parse_stsz(buf): #Sample Size Atoms version = buf[0] flags = buf[1:4] sample_size = struct.unpack('>I', buf[4:8])[0] n_entries = struct.unpack('>I', buf[8:12])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'sample_size : {sample_size}') print(f'number of entries : {n_entries}') sizes = [] for i in range(n_entries): i0 = 12 + i*4 i1 = i0 + 4 if len(buf) < i1: break size = struct.unpack('>I', buf[i0:i1])[0] sizes.append(size) print(f' {i}: {size}') def parse_stsc(buf): #Sample-to-Chunk Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') stoc = [] for i in range(n_entries): i0 = 8 + i*12 i1 = i0 + 12 if len(buf) < i1: break first_chunk = struct.unpack('>I', buf[i0:i0+4])[0] samples_per_chunk = struct.unpack('>I', buf[i0+4:i0+8])[0] sample_desc_id = struct.unpack('>I', buf[i0+8:i0+12])[0] stoc.append((first_chunk, samples_per_chunk, sample_desc_id)) print(f' {i}: {(first_chunk, samples_per_chunk, sample_desc_id)}') def parse_stco(buf): #Chunk Offset Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') chunk_offset_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break offset = struct.unpack('>I', buf[i0:i1])[0] chunk_offset_table.append(offset) print(f' {i}: {offset}') def parse_co64(buf): #64-bit chunk offset atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') chunk_offset_table = [] for i in range(n_entries): i0 = 8 + i*8 i1 = i0 + 8 if len(buf) < i1: break offset = struct.unpack('>Q', buf[i0:i1])[0] chunk_offset_table.append(offset) print(f' {i}: {offset}') def parse_stts(buf): #Time-to-Sample Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') time_to_sample_table = [] for i in range(n_entries): i0 = 8 + i*8 i1 = i0 + 8 if len(buf) < i1: break sample_count = struct.unpack('>I', buf[i0:i0+4])[0] sample_duration = struct.unpack('>I', buf[i0+4:i0+8])[0] time_to_sample_table.append((sample_count, sample_duration)) print(f' {i}: {(sample_count, sample_duration)}') def parse_stss(buf): #Sync Sample Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') sync_sample_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break sample = struct.unpack('>I', buf[i0:i1])[0] sync_sample_table.append(sample) print(f' {i}: {sample}') def parse_uuid(buf): print_binaries(buf[:16]) print('%s' % str(buf[16:], 'utf-8')) def print_binaries(buf, cur=None): if cur is None: cur = 0 for i in range(0, len(buf), 8): print('%010X : ' % (i+cur), end='') j = min(i+8, len(buf)) buf_ = buf[i:j] print(' '.join(['%02X'%(b) for b in buf_]), end='') print(' : ', end='') print(''.join(['%c'%(b) for b in buf_])) def print_atom_headers(f, verbose=False, pre_label=''): atom_start = f.tell() buf = f.read(8) n = struct.unpack('>I', buf[:4])[0] atom_type = str(buf[4:], 'utf-8') if n == 1: # decode 64-bit size buf = f.read(8) n = struct.unpack('>Q', buf)[0] #elif n == 0: # raise ValueError('not implemented yet') #print(f'{atom_type} (size: {n})') if not pre_label is None: print('%s%s (size: 0x%X)' % (pre_label, atom_type, n)) else: print('%s (size: 0x%X)' % (atom_type, n)) data_start = f.tell() if verbose: print_binaries(buf, atom_start) if not atom_type in ('moov', 'trak', 'mdia', 'minf', 'edts', 'dinf', 'stbl'): if n > 8: if atom_type == 'uuid': n_ = n else: n_ = min(n, 128) buf = f.read(n_-8) if atom_type == 'mvhd': parse_mvhd(buf) elif atom_type == 'tkhd': parse_tkhd(buf) elif atom_type == 'mdhd': parse_mdhd(buf) elif atom_type == 'stsd': parse_stsd(buf) elif atom_type == 'stsz': parse_stsz(buf) elif atom_type == 'stsc': parse_stsc(buf) elif atom_type == 'stco': parse_stco(buf) elif atom_type == 'co64': parse_co64(buf) elif atom_type == 'stts': parse_stts(buf) elif atom_type == 'stss': parse_stss(buf) elif atom_type == 'uuid': parse_uuid(buf) else: print('DATA:') print_binaries(buf, cur=data_start) else: # sub Atoms sub_end = atom_start + n sub_cur = data_start while True: f.seek(sub_cur) if f.tell() != sub_cur: raise ValueError(f'seek failed? {f.tell()} != {sub_cur}') sub_n, sub_type = print_atom_headers(f, verbose=False, pre_label=pre_label+atom_type+ ' / ') if sub_n == 0: break sub_cur += sub_n if sub_cur >= sub_end: break print('') return n, atom_type def print_atoms(filename, verbose=False): with open(filename, 'rb') as f: f.seek(0, 2) file_size = f.tell() print('file size : 0x%010X' % (file_size)) print('') cur = 0 while True: f.seek(cur) if f.tell() != cur: raise ValueError(f'seek failed? {f.tell()} != {cur}') n, _ = print_atom_headers(f, verbose=verbose) print('size : 0x%X' % (n)) if n == 0: break cur += n if cur >= file_size: break print('') # ## extracting `moov` as a reference def read_atom_head(f): cur = f.tell() buf = f.read(8) n = struct.unpack('>I', buf[:4])[0] atom_type = str(buf[4:], 'utf-8') buf2 = None if n == 1: # decode 64-bit size buf2 = f.read(8) n = struct.unpack('>Q', buf2)[0] del buf del buf2 gc.collect() return n, atom_type def extract_moov(src_filename, dst_filename, n_chunk=65536, verbose=False): with open(src_filename, 'rb') as f_src, open(dst_filename, 'wb') as f_dst: f_src.seek(0, 2) src_end = f_src.tell() # look for 'moov' src_cur = 0 while True: f_src.seek(src_cur) if f_src.tell() != src_cur: raise ValueError(f'seek failed? {f_src.tell()} != {src_cur}') n, atom_type = read_atom_head(f_src) if atom_type == 'moov': break src_cur += n # 'moov' is found moov_start = src_cur # copy moov f_src.seek(moov_start) if f_src.tell() != moov_start: raise ValueError(f'seek failed? {f_src.tell()} != {moov_start}') if verbose: it_moov = tqdm(range(moov_start, src_end, n_chunk)) else: it_moov = range(moov_start, src_end, n_chunk) #for src_cur in tqdm(range(moov_start, src_end, n_chunk)): for src_cur in it_moov: f_dst.write(f_src.read(n_chunk)) if src_end - src_cur > 0: f_dst.write(f_src.read(src_end - src_cur)) # ## regenerating sample tables from `mdat` def is_aac_header(buf, frame_length): cpe = (buf[0] & 0b11100000) >> 5 element_instance_tag = (buf[0] & 0b00011110) >> 1 common_window = (buf[0] & 0b00000001) if cpe != 0b001 or element_instance_tag != 0b0000: # not stereo # print('not stereo') return False elif common_window == 1: window_sequence = (buf[1] & 0b01100000) >> 5 if window_sequence != 0b10: # common case always_0 = (buf[1] & 0b10000000) >> 7 maxsfb = ((buf[1] & 0b00001111) << 2) | ((buf[2] & 0b11000000) >> 6) predictor = (buf[2] & 0b00100000) >> 5 gain = ((buf[2] & 0b00000111) << 5) | ((buf[3] & 0b11111000) >> 3) inc = (buf[4] & 0b01111100) >> 2 if always_0 == 0 and 40 <= maxsfb and maxsfb <= 51 and predictor == 0 and 100 <= gain and gain <= 228 and inc != 0: print('common case') return True else: # print('not common case') return False else: # eight-short-sequence always_10 = (buf[1] & 0b01100000) >> 5 maxsfb = (buf[1] & 0b00001111) gain = ((buf[3] & 0b01111111) << 1) | ((buf[4] & 0b10000000) >> 7) if always_10 == 0b10 and maxsfb >= 8 and 100 <= gain and gain <= 200: print('eight-short-sequence') return True else: # print('not eight-short-sequence') return False else: pulse_data = (buf[3] & 0b00001000) >> 3 tns_data = (buf[3] & 0b00000100) >> 2 gain_control_data = (buf[3] & 0b00000010) >> 1 if pulse_data != 0 or tns_data != 0 or gain_control_data != 0: maxsfb = ((buf[2] & 0b00001111) << 2) | ((buf[3] & 0b11000000) >> 6) if maxsfb != 0: gain = buf[1] always_0 = (buf[2] & 0b10000000) >> 7 if 100 <= gain and gain <= 200 and always_0 == 0: window_sequence = (buf[2] & 0b01100000) >> 5 if window_sequence != 0b10: # multiple windows if 48 <= maxsfb and maxsfb <= 51: print('multiple windows') return True else: # print('not multiple windows') return False else: # multiple windows and eight-short-sequence maxsfb = (buf[2] & 0b00001111) if maxsfb >= 8: print('multiple windows and eight-short-sequence') return True else: # print('not multiple windows and eight-short-sequence') return False else: # print('not multiple windows type') return False else: # degenerated block if frame_length == 7: print('degenerated block') return True else: # print('not degenerated block') return False else: # degenerated block with padding # gain = ((buf[1] & 0b01111111) << 1) | ((buf[2] & 0b10000000) >> 7) maxsfb = ((buf[2] & 0b00000111) << 3) | ((buf[3] & 0b11100000) >> 5) maxsfb2 = (buf[5] & 0b00011111) if maxsfb == 0 and maxsfb2 == 0: print('degenerated block with padding') return True else: # print('not degenerated block with padding') return False def recover_sample_tables_from_mdat_fast(filename, verbose=False): mov_table = [] aac_table = [] with open(filename, 'rb') as f_in: # look for 'mdat' src_cur = 0 while True: f_in.seek(src_cur) if f_in.tell() != src_cur: raise ValueError(f'seek failed? {f_in.tell()} != {src_cur}') n, atom_type = read_atom_head(f_in) if atom_type == 'mdat': break src_cur += n # 'mdat' is found mdat_start = src_cur if n == 0: # mdat from impcomplete mp4 file f_in.seek(0, 2) mdat_end = f_in.tell() # seek the data_start position # 8 bytes for the header PLUS 8 bytes for the reserved space of the size # f_in.seek(src_cur + 16) f_in.seek(src_cur) buf = f_in.read(4) head = struct.unpack('>I', buf)[0] if head == 1: f_in.seek(src_cur + 16) else: f_in.seek(src_cur + 8) else: mdat_end = src_cur + n # n = 0 # while True: # cur = f_in.tell() # if cur >= mdat_end: break # # buf = f_in.read(4) # buf = f_in.read(6) # # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: # # if buf[0] != 0x21 or (buf[1] & 0b11101111) != 0x0A or (buf[2] & 0b11111110) != 0x14: # if buf[0] != 0x21: # # h264 chunk # frame_length = struct.unpack('>I', buf)[0] + 4 # if cur+frame_length >= mdat_end: break # if verbose: print(f'{n}: [mov] {cur}, {frame_length}') # mov_table.append((cur, frame_length)) # f_in.seek(cur+frame_length) # else: # # buf_2 = f_in.read(2) # # from https://wiki.multimedia.cx/index.php/ADTS # # AAAAAAAA AAAABCCD EEFFFFGH HHIJKLMM MMMMMMMM MMMOOOOO OOOOOOPP (QQQQQQQQ QQQQQQQQ) # # 0th-byte 1st 2nd 3rd 4th 5th 6th (7th 8th ) # # 0xFF 0xF1 0x4C 0X80 -- typical case for Insta360 ONE-X # # M 13 frame length, this value must include 7 or 9 bytes of header length # # FrameLength = (ProtectionAbsent == 1 ? 7 : 9) + size(AACFrame) # # frame_length = ((buf[3] & 0b11) << 11) | (buf_2[0] << 3) | (buf_2[1] >> 5) # # if cur+frame_length >= mdat_end: break # frame_length = 0 # while True: # if cur + frame_length >= mdat_end: # break # frame_length_2 = struct.unpack('>I', buf)[0] + 4 # if frame_length_2 < 1048576 and cur + frame_length + frame_length_2 <= mdat_end: # f_in.seek(cur + frame_length + frame_length_2) # buf_2 = f_in.read(3) # if buf_2[0] == 0x21:# and (buf_2[1] & 0b11101111) == 0x0A and (buf_2[2] & 0b11111110) == 0x14: # # frame_length -= 6 # break # # flag = (buf[0] == 0x21)# and (buf[1] & 0b11101111) == 0x0A and (buf[2] & 0b11111110) == 0x14) # # if frame_length > 0 and flag: # # break # frame_length += 1 # f_in.seek(cur + frame_length) # buf = f_in.read(4) # if verbose: print(f'{n}: [aac] {cur}, {frame_length}') # aac_table.append((cur, frame_length)) # f_in.seek(cur+frame_length) # n += 1 # # frame_length = struct.unpack('>I', buf)[0] + 4 # # if cur + frame_length > mdat_end: # # break # # f_in.seek(cur + frame_length) # # buf_2 = f_in.read(4) # # if buf_2[0] == 0x21: # # if verbose: # # print(f'{n}: [mov] {cur}, {frame_length}') # # mov_table.append((cur, frame_length)) # # if verbose: # # print(f'{n}: [aac] {cur}, {frame_length}') # # aac_table.append((cur, frame_length_2)) # # n += 1 # # frame_length_2 = 0 # # else: # # while True: # # cur += 1 # # if cur > mdat_end or # # frame_length_2 += 1 # # frame_length_2 = struct.unpack('>I', buf_2)[0] + 4 # # if cur + frame_length + frame_length_2 > mdat_end: # # break # # f_in.seek(cur + frame_length + frame_length_2) # # buf_3 = f_in.read(4) # # if buf_3 = f_in.read(1) # # frame_length_2 += 1 # # f_in.seek(cur + 1) n = 0 while True: cur = f_in.tell() if cur >= mdat_end: break # buf = f_in.read(4) buf = f_in.read(6) # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: # h264 chunk # frame_length = struct.unpack('>I', buf)[0] + 4 # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(4) if (buf[0] & 0b11111110) == 0x20: break frame_length += struct.unpack('>I', buf)[0] + 4 if cur+frame_length >= mdat_end: break f_in.seek(cur + frame_length) if verbose: print(f'{n}: [mov] {cur}, {frame_length}') mov_table.append((cur, frame_length)) f_in.seek(cur+frame_length) else: # buf_2 = f_in.read(2) # from https://wiki.multimedia.cx/index.php/ADTS # AAAAAAAA AAAABCCD EEFFFFGH HHIJKLMM MMMMMMMM MMMOOOOO OOOOOOPP (QQQQQQQQ QQQQQQQQ) # 0th-byte 1st 2nd 3rd 4th 5th 6th (7th 8th ) # 0xFF 0xF1 0x4C 0X80 -- typical case for Insta360 ONE-X # M 13 frame length, this value must include 7 or 9 bytes of header length # FrameLength = (ProtectionAbsent == 1 ? 7 : 9) + size(AACFrame) # frame_length = ((buf[3] & 0b11) << 11) | (buf_2[0] << 3) | (buf_2[1] >> 5) # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(6) # if is_aac_header(buf, frame_length): break if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: break frame_length += 1 if cur + frame_length >= mdat_end: break f_in.seek(cur + frame_length) # if verbose: print(f'{n}: [aac] {cur}, {frame_length}') aac_table.append((cur, frame_length)) f_in.seek(cur+frame_length) n += 1 return mov_table, aac_table # ## rebuilding `moov` from sample tables def copy_atom_box(target_type, target_size, f_src, f_dst, only_header=True): src_size, atom_type = read_atom_head(f_src) if atom_type != target_type: raise ValueError(f'{target_type} not found but {atom_type}') if target_size is None: target_size = src_size f_dst.write(struct.pack('>I', target_size)) f_dst.write(target_type.encode('utf-8')) if not only_header: f_dst.write(f_src.read(target_size-8)) return src_size def recover_moov_from_sample_tables( moov_const, ref_filename, dst_filename, mov_table, aac_table, full_copy=True, n_chunk=65536, verbose=False, ): # constants # mov_sample_duration = 1001 # aac_sample_duration = 1024 # mvhd_timescale = 48000 # mov_timescale = 30000 # aac_timescale = 48000 mov_sample_duration = moov_const[0] aac_sample_duration = moov_const[1] mvhd_timescale = moov_const[2] mov_timescale = moov_const[3] aac_timescale = moov_const[4] n_mov_table = len(mov_table) n_aac_table = len(aac_table) mov_mdhd_duration = n_mov_table * mov_sample_duration aac_mdhd_duration = n_aac_table * aac_sample_duration mov_tkhd_duration = int(mov_mdhd_duration * mvhd_timescale / mov_timescale) aac_tkhd_duration = int(aac_mdhd_duration * mvhd_timescale / aac_timescale) if mov_tkhd_duration >= aac_tkhd_duration: mvhd_duration = mov_tkhd_duration else: mvhd_duration = aac_tkhd_duration sample_size_tables = [] sample_size_tables.append([s for o, s in mov_table]) sample_size_tables.append([s for o, s in aac_table]) chunk_offset_tables = [] chunk_offset_tables.append([o for o, s in mov_table]) chunk_offset_tables.append([o for o, s in aac_table]) # moov structure is assumed to be in the fixed format (for now) mov_stsz_size = len(sample_size_tables[0])* 4 + 20 aac_stsz_size = len(sample_size_tables[1])* 4 + 20 # mov_co64_size = len(chunk_offset_tables[0])* 8 + 16 # aac_co64_size = len(chunk_offset_tables[1])* 8 + 16 mov_stco_size = len(chunk_offset_tables[0])* 4 + 16 aac_stco_size = len(chunk_offset_tables[1])* 4 + 16 mov_stss_size = ((len(sample_size_tables[0])-1)//150 + 1)* 4 + 16 # mov_stbl_size = 8 + 0x141 + 0x18 + 0x1C + mov_stsz_size + mov_co64_size + mov_stss_size # aac_stbl_size = 8 + 0x82 + 0x18 + 0x1C + aac_stsz_size + aac_co64_size mov_stbl_size = 8 + 0xAB + 0x18 + 0x1C + mov_stsz_size + mov_stco_size + mov_stss_size aac_stbl_size = 8 + 0x67 + 0x18 + 0x14B75C + aac_stsz_size + aac_stco_size + 0x1A + 0x1C mov_minf_size = 8 + 0x14 + 0x24 + mov_stbl_size aac_minf_size = 8 + 0x10 + 0x24 + aac_stbl_size # mov_mdia_size = 8 + 0x20 + 0x2E + mov_minf_size # aac_mdia_size = 8 + 0x20 + 0x2E + aac_minf_size mov_mdia_size = 8 + 0x20 + 0x2D + mov_minf_size aac_mdia_size = 8 + 0x20 + 0x2D + aac_minf_size # mov_trak_size = 8 + 0x5C + 0x24 + mov_mdia_size + 0x618 # aac_trak_size = 8 + 0x5C + 0x24 + aac_mdia_size mov_trak_size = 8 + 0x5C + 0x24 + mov_mdia_size aac_trak_size = 8 + 0x5C + 0x30 + aac_mdia_size # moov_size = 8 + 0x6C + 0x73 + mov_trak_size + aac_trak_size # moov_size = 8 + 0x6C + 0x62 + mov_trak_size + aac_trak_size moov_size = 8 + 0x6C + 0x62 + mov_trak_size with open(ref_filename, 'rb') as f_moov, open(dst_filename, 'wb') as f_dst: f_moov.seek(0, 2) file_size = f_moov.tell() cur = 0 f_moov.seek(cur) if f_moov.tell() != cur: raise ValueError(f'seek failed? {f_moov.tell()} != {cur}') # moov copy_atom_box('moov', moov_size, f_moov, f_dst, only_header=True) #copy_atom_box('mvhd', None, f_moov, f_dst, only_header=False) # mvhd : duration = mov_tkhd_duration n = copy_atom_box('mvhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:12]) f_dst.write(struct.pack('>I', mvhd_timescale)) #duration = struct.unpack('>I', buf[16:20])[0] # f_dst.write(struct.pack('>I', mov_tkhd_duration)) f_dst.write(struct.pack('>I', mvhd_duration)) # the rest is unchanged f_dst.write(buf[20:]) #... #next_track_id = struct.unpack('>I', buf[96:100])[0] if n != (100+8): raise ValueError(f'ERROR: mov tkhd box size is not 108 but {n}') #copy_atom_box('udta', None, f_moov, f_dst, only_header=False) # movie track # trak copy_atom_box('trak', mov_trak_size, f_moov, f_dst, only_header=True) #copy_atom_box('tkhd', None, f_moov, f_dst, only_header=False) # tkhd : duration = mov_tkhd_duration n = copy_atom_box('tkhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:20]) #duration = struct.unpack('>I', buf[20:24])[0] f_dst.write(struct.pack('>I', mov_tkhd_duration)) # the rest is unchanged f_dst.write(buf[24:76]) #... f_dst.write(struct.pack('>I', 83886080)) f_dst.write(struct.pack('>I', 47185920)) #track_height = struct.unpack('>I', buf[80:84])[0] if n != (84+8): raise ValueError(f'ERROR: mov tkhd box size is not 92 but {n}') # copy_atom_box('edts', None, f_moov, f_dst, only_header=False) copy_atom_box('edts', None, f_moov, f_dst, only_header=True) n = copy_atom_box('elst', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:8]) f_dst.write(struct.pack('>I', mov_tkhd_duration)) f_dst.write(buf[12:]) if n != (20+8): raise ValueError(f'ERROR: mov elst box size is not 28 but {n}') # mdia copy_atom_box('mdia', mov_mdia_size, f_moov, f_dst, only_header=True) #copy_atom_box('mdhd', None, f_moov, f_dst, only_header=False) # mdhd : duration = mov_mdhd_duration n = copy_atom_box('mdhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:16]) #duration = struct.unpack('>I', buf[16:20])[0] f_dst.write(struct.pack('>I', mov_mdhd_duration)) # the rest is unchanged f_dst.write(buf[20:]) #... #quality = struct.unpack('>H', buf[22:24])[0] if n != (24+8): raise ValueError(f'ERROR: mov mdhd box size is not 32 but {n}') copy_atom_box('hdlr', None, f_moov, f_dst, only_header=False) # minf copy_atom_box('minf', mov_minf_size, f_moov, f_dst, only_header=True) copy_atom_box('vmhd', None, f_moov, f_dst, only_header=False) copy_atom_box('dinf', None, f_moov, f_dst, only_header=False) # stbl copy_atom_box('stbl', mov_stbl_size, f_moov, f_dst, only_header=True) copy_atom_box('stsd', None, f_moov, f_dst, only_header=False) #copy_atom_box('stts', None, f_moov, f_dst, only_header=False) # stts : sample_count = n_mov_table n = copy_atom_box('stts', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 1)) # n_entries f_dst.write(struct.pack('>I', n_mov_table)) # sample_count f_dst.write(struct.pack('>I', mov_sample_duration)) # sample_duration # stss n = copy_atom_box('stss', mov_stss_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags mov_stss_entries = (len(sample_size_tables[0])-1)//150 + 1 f_dst.write(struct.pack('>I', mov_stss_entries)) # n_entries ss = 1 for i_ss in range(mov_stss_entries): f_dst.write(struct.pack('>I', ss)) ss += 150 copy_atom_box('stsc', None, f_moov, f_dst, only_header=False) # stsz n = copy_atom_box('stsz', mov_stsz_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 0)) # sample_size f_dst.write(struct.pack('>I', len(sample_size_tables[0]))) # n_entries for sz in sample_size_tables[0]: f_dst.write(struct.pack('>I', sz)) # co64 # n = copy_atom_box('co64', mov_co64_size, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', len(chunk_offset_tables[0]))) # n_entries # for co in chunk_offset_tables[0]: # f_dst.write(struct.pack('>Q', co)) n = copy_atom_box('stco', mov_stco_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', len(chunk_offset_tables[0]))) # n_entries for co in chunk_offset_tables[0]: f_dst.write(struct.pack('>I', co)) # uuid # copy_atom_box('uuid', None, f_moov, f_dst, only_header=False) # # audio track # # trak # copy_atom_box('trak', aac_trak_size, f_moov, f_dst, only_header=True) # #copy_atom_box('tkhd', None, f_moov, f_dst, only_header=False) # # tkhd : duration = aac_mdhd_duration # n = copy_atom_box('tkhd', None, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # # the following is unchanged # f_dst.write(buf[:20]) # #duration = struct.unpack('>I', buf[20:24])[0] # f_dst.write(struct.pack('>I', aac_mdhd_duration)) # # the rest is unchanged # f_dst.write(buf[24:]) # #... # #track_height = struct.unpack('>I', buf[80:84])[0] # if n != (84+8): raise ValueError(f'ERROR: audio tkhd box size is not 92 but {n}') # # copy_atom_box('edts', None, f_moov, f_dst, only_header=False) # copy_atom_box('edts', None, f_moov, f_dst, only_header=True) # n = copy_atom_box('elst', None, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:12]) # f_dst.write(struct.pack('>I', aac_tkhd_duration)) # f_dst.write(buf[16:]) # if n != (32+8): raise ValueError(f'ERROR: mov elst box size is not 40 but {n}') # # mdia # copy_atom_box('mdia', aac_mdia_size, f_moov, f_dst, only_header=True) # #copy_atom_box('mdhd', None, f_moov, f_dst, only_header=False) # # mdhd : duration = aac_mdhd_duration # n = copy_atom_box('mdhd', None, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # # the following is unchanged # f_dst.write(buf[:16]) # #duration = struct.unpack('>I', buf[16:20])[0] # f_dst.write(struct.pack('>I', aac_mdhd_duration)) # # the rest is unchanged # f_dst.write(buf[20:]) # #... # #quality = struct.unpack('>H', buf[22:24])[0] # if n != (24+8): raise ValueError(f'ERROR: audio mdhd box size is not 32 but {n}') # copy_atom_box('hdlr', None, f_moov, f_dst, only_header=False) # # minf # copy_atom_box('minf', aac_minf_size, f_moov, f_dst, only_header=True) # copy_atom_box('smhd', None, f_moov, f_dst, only_header=False) # copy_atom_box('dinf', None, f_moov, f_dst, only_header=False) # # stbl # copy_atom_box('stbl', aac_stbl_size, f_moov, f_dst, only_header=True) # copy_atom_box('stsd', None, f_moov, f_dst, only_header=False) # #copy_atom_box('stts', None, f_moov, f_dst, only_header=False) # # stts : sample_count = n_aac_table # n = copy_atom_box('stts', None, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', 1)) # n_entries # f_dst.write(struct.pack('>I', n_aac_table)) # sample_count # f_dst.write(struct.pack('>I', aac_sample_duration)) # sample_duration # # copy_atom_box('stsc', None, f_moov, f_dst, only_header=False) # cur = f_dst.tell() # n = copy_atom_box('stsc', None, f_moov, f_dst, only_header=True) # f_dst.seek(cur) # f_dst.write(struct.pack('>I', 28)) # f_dst.seek(cur+8) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', 1)) # n_entries # f_dst.write(struct.pack('>I', 1)) # f_dst.write(struct.pack('>I', 1)) # f_dst.write(struct.pack('>I', 1)) # # stsz # n = copy_atom_box('stsz', aac_stsz_size, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', 0)) # sample_size # f_dst.write(struct.pack('>I', len(sample_size_tables[1]))) # n_entries # for sz in sample_size_tables[1]: # f_dst.write(struct.pack('>I', sz)) # # co64 # # n = copy_atom_box('co64', aac_co64_size, f_moov, f_dst, only_header=True) # # buf = f_moov.read(n-8) # # f_dst.write(buf[:4]) # version + flags # # f_dst.write(struct.pack('>I', len(chunk_offset_tables[1]))) # n_entries # # for co in chunk_offset_tables[1]: # # f_dst.write(struct.pack('>Q', co)) # n = copy_atom_box('stco', aac_stco_size, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', len(chunk_offset_tables[1]))) # n_entries # for co in chunk_offset_tables[1]: # f_dst.write(struct.pack('>I', co)) # copy_atom_box('sgpd', None, f_moov, f_dst, only_header=False) # copy_atom_box('sbgp', None, f_moov, f_dst, only_header=False) # to skip the audio track print_atom_headers(f_moov) copy_atom_box('udta', None, f_moov, f_dst, only_header=False) if not full_copy: return # just copy the rest of reference moov file moov_cur = f_moov.tell() f_moov.seek(0, 2) moov_size = f_moov.tell() f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') if verbose: it_moov = tqdm(range(moov_cur, moov_size, n_chunk)) else: it_moov = range(moov_cur, moov_size, n_chunk) #for moov_cur in tqdm(range(moov_cur, moov_size, n_chunk)): for moov_cur in it_moov: f_dst.write(f_moov.read(n_chunk)) if moov_size - moov_cur > 0: f_dst.write(f_moov.read(moov_size - moov_cur)) # ## merging the recovered `moov` def merge_moov( src_filename, moov_filename, dst_filename, n_chunk=65536, verbose=False): with open(src_filename, 'rb') as f_src, open(moov_filename, 'rb') as f_moov, open(dst_filename, 'wb') as f_dst: f_src.seek(0, 2) file_size = f_src.tell() cur = 0 f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') # ftyp n, atom_type = read_atom_head(f_src) if atom_type != 'ftyp': raise ValueError('ftyp not found') f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') buf = f_src.read(n) f_dst.write(buf) if verbose: print_binaries(buf) cur += n # cur += 16 # free n, atom_type = read_atom_head(f_src) if atom_type != 'free': raise ValueError('ftyp not found') f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') buf = f_src.read(n) f_dst.write(buf) if verbose: print_binaries(buf) cur += n # mdat n, atom_type = read_atom_head(f_src) if atom_type != 'mdat': raise ValueError('mdat not found') # if n != 0: raise ValueError('size would be zero...') # # fixed mdat header # if verbose: # print_binaries(struct.pack('>Icccc', 1, b'm', b'd', b'a', b't')) # print_binaries(struct.pack('>Q', file_size-0x20)) # f_dst.write(struct.pack('>Icccc', 1, b'm', b'd', b'a', b't')) # f_dst.write(struct.pack('>Q', file_size-0x20)) # cur += 16 f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') if verbose: it_cur = tqdm(range(cur, file_size, n_chunk)) else: it_cur = range(cur, file_size, n_chunk) #for cur in tqdm(range(cur, file_size, n_chunk)): for cur in it_cur: f_dst.write(f_src.read(n_chunk)) if file_size - cur > 0: f_dst.write(f_src.read(file_size - cur)) print('') temp = f_dst.tell() f_dst.seek(40) f_dst.write(struct.pack('>Icccc', file_size - 0x20, b'm', b'd', b'a', b't')) f_dst.seek(temp + 8) # search moov f_moov.seek(0, 2) moov_size = f_moov.tell() if verbose: print(f'moov_size: {moov_size}') moov_cur = 0 f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') n, atom_type = read_atom_head(f_moov) if atom_type != 'moov': raise ValueError(f'something is wrong...') # copy moov f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') if verbose: it_moov = tqdm(range(moov_cur, moov_size, n_chunk)) else: it_moov = range(moov_cur, moov_size, n_chunk) #for moov_cur in tqdm(range(moov_cur, moov_size, n_chunk)): for moov_cur in it_moov: f_dst.write(f_moov.read(n_chunk)) if moov_size - moov_cur > 0: f_dst.write(f_moov.read(moov_size - moov_cur)) # # main program to recover corrupted MP4 def finsta360( moov_const, src_filename, ref_filename=None, dst_filename=None, keep_temp=False, verbose=False): if ref_filename is None: # check mode print_atoms(src_filename) return # temporary files ref_moov_filename = 'finsta360_ref.moov' new_moov_filename = 'finsta360_new.moov' # 1) extract reference moov print('') print('########################################') print(f'# 1) extracting reference moov from\n\t{ref_filename}') extract_moov(ref_filename, ref_moov_filename) if verbose: print_atoms(ref_moov_filename) # 2) regenerate sample tables from mdat print('') print('########################################') print(f'# 2) regenerate sample tables from mdat in\n\t{src_filename}') mov_table, aac_table = recover_sample_tables_from_mdat_fast( src_filename, verbose=verbose) if verbose: print(f'number of samples (movie) : {len(mov_table)}') print(f'number of samples (audio) : {len(aac_table)}') # constants # mov_sample_duration = 1001 # aac_sample_duration = 1024 # mvhd_timescale = 48000 # mov_timescale = 30000 # aac_timescale = 48000 mov_sample_duration = moov_const[0] aac_sample_duration = moov_const[1] mvhd_timescale = moov_const[2] mov_timescale = moov_const[3] aac_timescale = moov_const[4] n_mov_table = len(mov_table) n_aac_table = len(aac_table) mov_mdhd_duration = n_mov_table * mov_sample_duration aac_mdhd_duration = n_aac_table * aac_sample_duration mov_tkhd_duration = int(mov_mdhd_duration * mvhd_timescale / mov_timescale) # mvhd mvhd_duration_sec = mov_tkhd_duration / mvhd_timescale print(f'mvhd duration : {mvhd_duration_sec} sec / {mvhd_duration_sec/60} min') # movie mdhd mov_duration_sec = mov_mdhd_duration / mov_timescale print(f'movie duration : {mov_duration_sec} sec / {mov_duration_sec/60} min') # audio mdhd aac_duration_sec = aac_mdhd_duration / aac_timescale print(f'audio duration : {aac_duration_sec} sec / {aac_duration_sec/60} min') # 3) rebuilding moov from the sample tables print('') print('########################################') print(f'# 3) rebuilding moov from the sample tables') recover_moov_from_sample_tables( moov_const, ref_moov_filename, new_moov_filename, mov_table, aac_table, full_copy=True, ) if verbose: print_atoms(new_moov_filename) if dst_filename is None: # test mode if not keep_temp: os.remove(ref_moov_filename) os.remove(new_moov_filename) return # 4) merging the rebuilt moov into the source print('') print('########################################') print(f'# 4) merging the rebuilt moov into\n\t{src_filename}\nas\n\t{dst_filename}') merge_moov( src_filename, new_moov_filename, dst_filename, ) if not keep_temp: os.remove(ref_moov_filename) os.remove(new_moov_filename) def usage(): print('finsta360.py : to finalize incomplete MP4 of Insta360 ONE-X') print('https://github.com/kichiki/finsta360') print('USAGE: finsta360.py [options]') print('\t-s file : source file, that is, corrupted mp4 (insv) file') print('\t-r file : complete mp4 (insv) file as a reference') print('\t-o file : output recovered mp4 (insv) file') print('\t-v : to set verbose mode') print('\t-k : to keep temporary files') print('\t (reference and recovered moov files, finsta360*.moov)') print('If you provide only source file (-s), program prints the metadata') print('If you dont provide output file (-o), program just runs without writing') sys.exit () if __name__ == '__main__': src_filename = None ref_filename = None dst_filename = None verbose = False keep_temp = False i = 1 while i < len(sys.argv): if sys.argv[i] == '-s': src_filename = sys.argv[i+1] i += 2 elif sys.argv[i] == '-r': ref_filename = sys.argv[i+1] i += 2 elif sys.argv[i] == '-o': dst_filename = sys.argv[i+1] i += 2 elif sys.argv[i] == '-v': verbose = True i += 1 elif sys.argv[i] == '-k': keep_temp = True i += 1 else: usage() break if src_filename is None: print(f'you must provie source file {src_filename}') usage() sys.exit() if not os.path.exists(src_filename): print(f'source file {src_filename} does not exist') sys.exit() if not ref_filename is None and not os.path.exists(ref_filename): print(f'reference file {ref_filename} does not exist') sys.exit() if not dst_filename is None and os.path.exists(dst_filename): print(f'output file {dst_filename} already exists') sys.exit() # constants mov_sample_duration = 3000 aac_sample_duration = 1024 mvhd_timescale = 90000 mov_timescale = 90000 aac_timescale = 48000 moov_const = (mov_sample_duration, aac_sample_duration, mvhd_timescale, mov_timescale, aac_timescale) # with open('aac.aac', 'rb') as f_in: # while True: # cur = f_in.tell() # buf = f_in.read(16) # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: # hoge = 0 # else: # print_binaries(buf[7:]) # f_in.seek(cur+4) finsta360( moov_const, src_filename, ref_filename, dst_filename, keep_temp, verbose) sys.exit()
rawaac.py
#!/usr/bin/env python # coding: utf-8 # finsta360 - python script to finalize incomplete MP4 of Insta360 ONE-X # https://github.com/kichiki/finsta360 import sys import os.path import struct from datetime import datetime, timedelta import gc #from tqdm import tqdm # ## parsing mp4 def parse_mvhd(buf): # Movie Header Atoms version = buf[0] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] time_scale = struct.unpack('>I', buf[12:16])[0] duration = struct.unpack('>I', buf[16:20])[0] preferred_rate = struct.unpack('>I', buf[20:24])[0] preferred_volume = struct.unpack('>H', buf[24:26])[0] # next 10 bytes are reserved matrix_structure = struct.unpack('>IIIIIIIII', buf[36:72]) preview_time = struct.unpack('>I', buf[72:76])[0] preview_duration = struct.unpack('>I', buf[76:80])[0] poster_time = struct.unpack('>I', buf[80:84])[0] selection_time = struct.unpack('>I', buf[84:88])[0] selection_duration = struct.unpack('>I', buf[88:92])[0] current_time = struct.unpack('>I', buf[92:96])[0] next_track_id = struct.unpack('>I', buf[96:100])[0] print(f'version : {version}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'time scale : {time_scale}') print(f'duration : {duration} / {duration/time_scale} sec / {duration/time_scale/60} min') print(f'preferred_rate : {preferred_rate}') print(f'preferred_volume : {preferred_volume}') print(f'matrix_structure : {matrix_structure}') print(f'preview_time : {preview_time}') print(f'preview_duration : {preview_duration}') print(f'poster_time : {poster_time}') print(f'selection_time : {selection_time}') print(f'selection_duration : {selection_duration}') print(f'current_time : {current_time}') print(f'next_track_id : {next_track_id}') def parse_tkhd(buf): #Track Header Atoms version = buf[0] flags = buf[1:4] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] track_id = struct.unpack('>I', buf[12:16])[0] # next 4 bytes are reserved duration = struct.unpack('>I', buf[20:24])[0] # next 8 bytes is reserved layer = struct.unpack('>H', buf[32:34])[0] alternate_group = struct.unpack('>H', buf[34:36])[0] volume = struct.unpack('>H', buf[36:38])[0] # next 2 bytes are reserved matrix_structure = struct.unpack('>IIIIIIIII', buf[40:76]) track_width = struct.unpack('>I', buf[76:80])[0] track_height = struct.unpack('>I', buf[80:84])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'track_id : {track_id}') print(f'duration : {duration}') print(f'layer : {layer}') print(f'alternate_group : {alternate_group}') print(f'volume : {volume}') print(f'matrix_structure : {matrix_structure}') print(f'track_width : {track_width}') print(f'track_height : {track_height}') def parse_mdhd(buf): #Media Header Atoms version = buf[0] flags = buf[1:4] creation_time = struct.unpack('>I', buf[4:8])[0] modification_time = struct.unpack('>I', buf[8:12])[0] time_scale = struct.unpack('>I', buf[12:16])[0] duration = struct.unpack('>I', buf[16:20])[0] language = struct.unpack('>H', buf[20:22])[0] quality = struct.unpack('>H', buf[22:24])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'creation time : {datetime(1904,1,1) + timedelta(seconds=creation_time)}') print(f'modification_time : {datetime(1904,1,1) + timedelta(seconds=modification_time)}') print(f'time scale : {time_scale}') print(f'duration : {duration} / {duration/time_scale} sec / {duration/time_scale/60} min') print(f'language : {language}') print(f'quality : {quality}') def parse_stsd(buf): #Sample Description Atoms print('DATA:') print_binaries(buf) print(f'size of buf: {len(buf)}') version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') sample_description_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break sample_description_size = struct.unpack('>I', buf[i0:i0+4])[0] data_format = str(buf[i0+4:i0+8], 'utf-8') data_reference_index = struct.unpack('>H', buf[i0+14:i0+16])[0] sample_description_table.append( (sample_description_size, data_format, data_reference_index)) print('%d: size: 0x%X, format: %s, ref_index: 0x%X' % ( i, sample_description_size, data_format, data_reference_index)) def parse_stsz(buf): #Sample Size Atoms version = buf[0] flags = buf[1:4] sample_size = struct.unpack('>I', buf[4:8])[0] n_entries = struct.unpack('>I', buf[8:12])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'sample_size : {sample_size}') print(f'number of entries : {n_entries}') sizes = [] for i in range(n_entries): i0 = 12 + i*4 i1 = i0 + 4 if len(buf) < i1: break size = struct.unpack('>I', buf[i0:i1])[0] sizes.append(size) print(f' {i}: {size}') def parse_stsc(buf): #Sample-to-Chunk Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') stoc = [] for i in range(n_entries): i0 = 8 + i*12 i1 = i0 + 12 if len(buf) < i1: break first_chunk = struct.unpack('>I', buf[i0:i0+4])[0] samples_per_chunk = struct.unpack('>I', buf[i0+4:i0+8])[0] sample_desc_id = struct.unpack('>I', buf[i0+8:i0+12])[0] stoc.append((first_chunk, samples_per_chunk, sample_desc_id)) print(f' {i}: {(first_chunk, samples_per_chunk, sample_desc_id)}') def parse_stco(buf): #Chunk Offset Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') chunk_offset_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break offset = struct.unpack('>I', buf[i0:i1])[0] chunk_offset_table.append(offset) print(f' {i}: {offset}') def parse_co64(buf): #64-bit chunk offset atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') chunk_offset_table = [] for i in range(n_entries): i0 = 8 + i*8 i1 = i0 + 8 if len(buf) < i1: break offset = struct.unpack('>Q', buf[i0:i1])[0] chunk_offset_table.append(offset) print(f' {i}: {offset}') def parse_stts(buf): #Time-to-Sample Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') time_to_sample_table = [] for i in range(n_entries): i0 = 8 + i*8 i1 = i0 + 8 if len(buf) < i1: break sample_count = struct.unpack('>I', buf[i0:i0+4])[0] sample_duration = struct.unpack('>I', buf[i0+4:i0+8])[0] time_to_sample_table.append((sample_count, sample_duration)) print(f' {i}: {(sample_count, sample_duration)}') def parse_stss(buf): #Sync Sample Atoms version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] print(f'version : {version}') print(f'flags : {flags}') print(f'number of entries : {n_entries}') sync_sample_table = [] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break sample = struct.unpack('>I', buf[i0:i1])[0] sync_sample_table.append(sample) print(f' {i}: {sample}') def parse_uuid(buf): print_binaries(buf[:16]) print('%s' % str(buf[16:], 'utf-8')) def print_binaries(buf, cur=None): if cur is None: cur = 0 for i in range(0, len(buf), 8): print('%010X : ' % (i+cur), end='') j = min(i+8, len(buf)) buf_ = buf[i:j] print(' '.join(['%02X'%(b) for b in buf_]), end='') print(' : ', end='') print(''.join(['%c'%(b) for b in buf_])) def print_atom_headers(f, verbose=False, pre_label=''): atom_start = f.tell() buf = f.read(8) n = struct.unpack('>I', buf[:4])[0] atom_type = str(buf[4:], 'utf-8') if n == 1: # decode 64-bit size buf = f.read(8) n = struct.unpack('>Q', buf)[0] #elif n == 0: # raise ValueError('not implemented yet') #print(f'{atom_type} (size: {n})') if not pre_label is None: print('%s%s (size: 0x%X)' % (pre_label, atom_type, n)) else: print('%s (size: 0x%X)' % (atom_type, n)) data_start = f.tell() if verbose: print_binaries(buf, atom_start) if not atom_type in ('moov', 'trak', 'mdia', 'minf', 'edts', 'dinf', 'stbl'): if n > 8: if atom_type == 'uuid': n_ = n else: n_ = min(n, 128) buf = f.read(n_-8) if atom_type == 'mvhd': parse_mvhd(buf) elif atom_type == 'tkhd': parse_tkhd(buf) elif atom_type == 'mdhd': parse_mdhd(buf) elif atom_type == 'stsd': parse_stsd(buf) elif atom_type == 'stsz': parse_stsz(buf) elif atom_type == 'stsc': parse_stsc(buf) elif atom_type == 'stco': parse_stco(buf) elif atom_type == 'co64': parse_co64(buf) elif atom_type == 'stts': parse_stts(buf) elif atom_type == 'stss': parse_stss(buf) elif atom_type == 'uuid': parse_uuid(buf) else: print('DATA:') print_binaries(buf, cur=data_start) else: # sub Atoms sub_end = atom_start + n sub_cur = data_start while True: f.seek(sub_cur) if f.tell() != sub_cur: raise ValueError(f'seek failed? {f.tell()} != {sub_cur}') sub_n, sub_type = print_atom_headers(f, verbose=False, pre_label=pre_label+atom_type+ ' / ') if sub_n == 0: break sub_cur += sub_n if sub_cur >= sub_end: break print('') return n, atom_type def print_atoms(filename, verbose=False): with open(filename, 'rb') as f: f.seek(0, 2) file_size = f.tell() print('file size : 0x%010X' % (file_size)) print('') cur = 0 while True: f.seek(cur) if f.tell() != cur: raise ValueError(f'seek failed? {f.tell()} != {cur}') n, _ = print_atom_headers(f, verbose=verbose) print('size : 0x%X' % (n)) if n == 0: break cur += n if cur >= file_size: break print('') # ## extracting `moov` as a reference def read_atom_head(f): cur = f.tell() buf = f.read(8) n = struct.unpack('>I', buf[:4])[0] atom_type = str(buf[4:], 'utf-8') buf2 = None if n == 1: # decode 64-bit size buf2 = f.read(8) n = struct.unpack('>Q', buf2)[0] del buf del buf2 gc.collect() return n, atom_type def extract_moov(src_filename, dst_filename, n_chunk=65536, verbose=False): with open(src_filename, 'rb') as f_src, open(dst_filename, 'wb') as f_dst: f_src.seek(0, 2) src_end = f_src.tell() # look for 'moov' src_cur = 0 while True: f_src.seek(src_cur) if f_src.tell() != src_cur: raise ValueError(f'seek failed? {f_src.tell()} != {src_cur}') n, atom_type = read_atom_head(f_src) if atom_type == 'moov': break src_cur += n # 'moov' is found moov_start = src_cur # copy moov f_src.seek(moov_start) if f_src.tell() != moov_start: raise ValueError(f'seek failed? {f_src.tell()} != {moov_start}') if verbose: it_moov = tqdm(range(moov_start, src_end, n_chunk)) else: it_moov = range(moov_start, src_end, n_chunk) #for src_cur in tqdm(range(moov_start, src_end, n_chunk)): for src_cur in it_moov: f_dst.write(f_src.read(n_chunk)) if src_end - src_cur > 0: f_dst.write(f_src.read(src_end - src_cur)) # ## regenerating sample tables from `mdat` def is_aac_header(buf, frame_length): cpe = (buf[0] & 0b11100000) >> 5 element_instance_tag = (buf[0] & 0b00011110) >> 1 common_window = (buf[0] & 0b00000001) if cpe != 0b001 or element_instance_tag != 0b0000: # not stereo # print('not stereo') return False elif common_window == 1: window_sequence = (buf[1] & 0b01100000) >> 5 if window_sequence != 0b10: # common case always_0 = (buf[1] & 0b10000000) >> 7 maxsfb = ((buf[1] & 0b00001111) << 2) | ((buf[2] & 0b11000000) >> 6) predictor = (buf[2] & 0b00100000) >> 5 gain = ((buf[2] & 0b00000111) << 5) | ((buf[3] & 0b11111000) >> 3) inc = (buf[4] & 0b01111100) >> 2 if always_0 == 0 and 40 <= maxsfb and maxsfb <= 51 and predictor == 0 and 100 <= gain and gain <= 228 and inc != 0: print('common case') return True else: # print('not common case') return False else: # eight-short-sequence always_10 = (buf[1] & 0b01100000) >> 5 maxsfb = (buf[1] & 0b00001111) gain = ((buf[3] & 0b01111111) << 1) | ((buf[4] & 0b10000000) >> 7) if always_10 == 0b10 and maxsfb >= 8 and 100 <= gain and gain <= 200: print('eight-short-sequence') return True else: # print('not eight-short-sequence') return False else: pulse_data = (buf[3] & 0b00001000) >> 3 tns_data = (buf[3] & 0b00000100) >> 2 gain_control_data = (buf[3] & 0b00000010) >> 1 if pulse_data != 0 or tns_data != 0 or gain_control_data != 0: maxsfb = ((buf[2] & 0b00001111) << 2) | ((buf[3] & 0b11000000) >> 6) if maxsfb != 0: gain = buf[1] always_0 = (buf[2] & 0b10000000) >> 7 if 100 <= gain and gain <= 200 and always_0 == 0: window_sequence = (buf[2] & 0b01100000) >> 5 if window_sequence != 0b10: # multiple windows if 48 <= maxsfb and maxsfb <= 51: print('multiple windows') return True else: # print('not multiple windows') return False else: # multiple windows and eight-short-sequence maxsfb = (buf[2] & 0b00001111) if maxsfb >= 8: print('multiple windows and eight-short-sequence') return True else: # print('not multiple windows and eight-short-sequence') return False else: # print('not multiple windows type') return False else: # degenerated block if frame_length == 7: print('degenerated block') return True else: # print('not degenerated block') return False else: # degenerated block with padding # gain = ((buf[1] & 0b01111111) << 1) | ((buf[2] & 0b10000000) >> 7) maxsfb = ((buf[2] & 0b00000111) << 3) | ((buf[3] & 0b11100000) >> 5) maxsfb2 = (buf[5] & 0b00011111) if maxsfb == 0 and maxsfb2 == 0: print('degenerated block with padding') return True else: # print('not degenerated block with padding') return False def recover_sample_tables_from_mdat_fast(filename, verbose=False): mov_table = [] aac_table = [] with open(filename, 'rb') as f_in: # look for 'mdat' src_cur = 0 while True: f_in.seek(src_cur) if f_in.tell() != src_cur: raise ValueError(f'seek failed? {f_in.tell()} != {src_cur}') n, atom_type = read_atom_head(f_in) if atom_type == 'mdat': break src_cur += n # 'mdat' is found mdat_start = src_cur if n == 0: # mdat from impcomplete mp4 file f_in.seek(0, 2) mdat_end = f_in.tell() # seek the data_start position # 8 bytes for the header PLUS 8 bytes for the reserved space of the size # f_in.seek(src_cur + 16) f_in.seek(src_cur) buf = f_in.read(4) head = struct.unpack('>I', buf)[0] if head == 1: f_in.seek(src_cur + 16) else: f_in.seek(src_cur + 8) else: mdat_end = src_cur + n # n = 0 # while True: # cur = f_in.tell() # if cur >= mdat_end: break # # buf = f_in.read(4) # buf = f_in.read(6) # # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: # # if buf[0] != 0x21 or (buf[1] & 0b11101111) != 0x0A or (buf[2] & 0b11111110) != 0x14: # if buf[0] != 0x21: # # h264 chunk # frame_length = struct.unpack('>I', buf)[0] + 4 # if cur+frame_length >= mdat_end: break # if verbose: print(f'{n}: [mov] {cur}, {frame_length}') # mov_table.append((cur, frame_length)) # f_in.seek(cur+frame_length) # else: # # buf_2 = f_in.read(2) # # from https://wiki.multimedia.cx/index.php/ADTS # # AAAAAAAA AAAABCCD EEFFFFGH HHIJKLMM MMMMMMMM MMMOOOOO OOOOOOPP (QQQQQQQQ QQQQQQQQ) # # 0th-byte 1st 2nd 3rd 4th 5th 6th (7th 8th ) # # 0xFF 0xF1 0x4C 0X80 -- typical case for Insta360 ONE-X # # M 13 frame length, this value must include 7 or 9 bytes of header length # # FrameLength = (ProtectionAbsent == 1 ? 7 : 9) + size(AACFrame) # # frame_length = ((buf[3] & 0b11) << 11) | (buf_2[0] << 3) | (buf_2[1] >> 5) # # if cur+frame_length >= mdat_end: break # frame_length = 0 # while True: # if cur + frame_length >= mdat_end: # break # frame_length_2 = struct.unpack('>I', buf)[0] + 4 # if frame_length_2 < 1048576 and cur + frame_length + frame_length_2 <= mdat_end: # f_in.seek(cur + frame_length + frame_length_2) # buf_2 = f_in.read(3) # if buf_2[0] == 0x21:# and (buf_2[1] & 0b11101111) == 0x0A and (buf_2[2] & 0b11111110) == 0x14: # # frame_length -= 6 # break # # flag = (buf[0] == 0x21)# and (buf[1] & 0b11101111) == 0x0A and (buf[2] & 0b11111110) == 0x14) # # if frame_length > 0 and flag: # # break # frame_length += 1 # f_in.seek(cur + frame_length) # buf = f_in.read(4) # if verbose: print(f'{n}: [aac] {cur}, {frame_length}') # aac_table.append((cur, frame_length)) # f_in.seek(cur+frame_length) # n += 1 # # frame_length = struct.unpack('>I', buf)[0] + 4 # # if cur + frame_length > mdat_end: # # break # # f_in.seek(cur + frame_length) # # buf_2 = f_in.read(4) # # if buf_2[0] == 0x21: # # if verbose: # # print(f'{n}: [mov] {cur}, {frame_length}') # # mov_table.append((cur, frame_length)) # # if verbose: # # print(f'{n}: [aac] {cur}, {frame_length}') # # aac_table.append((cur, frame_length_2)) # # n += 1 # # frame_length_2 = 0 # # else: # # while True: # # cur += 1 # # if cur > mdat_end or # # frame_length_2 += 1 # # frame_length_2 = struct.unpack('>I', buf_2)[0] + 4 # # if cur + frame_length + frame_length_2 > mdat_end: # # break # # f_in.seek(cur + frame_length + frame_length_2) # # buf_3 = f_in.read(4) # # if buf_3 = f_in.read(1) # # frame_length_2 += 1 # # f_in.seek(cur + 1) n = 0 while True: cur = f_in.tell() if cur >= mdat_end: break # buf = f_in.read(4) buf = f_in.read(6) # if buf[0] != 0xFF or buf[1] != 0xF1 or buf[2] != 0x4C or (buf[3] & 0b11111100) != 0x80: if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: # h264 chunk # frame_length = struct.unpack('>I', buf)[0] + 4 # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(4) if (buf[0] & 0b11111110) == 0x20: break frame_length += struct.unpack('>I', buf)[0] + 4 if cur+frame_length >= mdat_end: break f_in.seek(cur + frame_length) # if verbose: print(f'{n}: [mov] {cur}, {frame_length}') mov_table.append((cur, frame_length)) f_in.seek(cur+frame_length) else: # buf_2 = f_in.read(2) # from https://wiki.multimedia.cx/index.php/ADTS # AAAAAAAA AAAABCCD EEFFFFGH HHIJKLMM MMMMMMMM MMMOOOOO OOOOOOPP (QQQQQQQQ QQQQQQQQ) # 0th-byte 1st 2nd 3rd 4th 5th 6th (7th 8th ) # 0xFF 0xF1 0x4C 0X80 -- typical case for Insta360 ONE-X # M 13 frame length, this value must include 7 or 9 bytes of header length # FrameLength = (ProtectionAbsent == 1 ? 7 : 9) + size(AACFrame) # frame_length = ((buf[3] & 0b11) << 11) | (buf_2[0] << 3) | (buf_2[1] >> 5) # if cur+frame_length >= mdat_end: break frame_length = 6 while True: buf = f_in.read(6) # if is_aac_header(buf, frame_length): break if struct.unpack('>I', buf[:4])[0] == 2 and buf[4] == 0x09 and buf[5] == 0xF0: break frame_length += 1 if cur + frame_length >= mdat_end: break f_in.seek(cur + frame_length) if verbose: print(f'{n}: [aac] {cur}, {frame_length}') aac_table.append((cur, frame_length)) f_in.seek(cur+frame_length) n += 1 return mov_table, aac_table # ## rebuilding `moov` from sample tables def copy_atom_box(target_type, target_size, f_src, f_dst, only_header=True): src_size, atom_type = read_atom_head(f_src) if atom_type != target_type: raise ValueError(f'{target_type} not found but {atom_type}') if target_size is None: target_size = src_size f_dst.write(struct.pack('>I', target_size)) f_dst.write(target_type.encode('utf-8')) if not only_header: f_dst.write(f_src.read(target_size-8)) return src_size def recover_moov_from_sample_tables( moov_const, ref_filename, dst_filename, mov_table, aac_table, full_copy=True, n_chunk=65536, verbose=False, ): # constants # mov_sample_duration = 1001 # aac_sample_duration = 1024 # mvhd_timescale = 48000 # mov_timescale = 30000 # aac_timescale = 48000 mov_sample_duration = moov_const[0] aac_sample_duration = moov_const[1] mvhd_timescale = moov_const[2] mov_timescale = moov_const[3] aac_timescale = moov_const[4] n_mov_table = len(mov_table) n_aac_table = len(aac_table) mov_mdhd_duration = n_mov_table * mov_sample_duration aac_mdhd_duration = n_aac_table * aac_sample_duration mov_tkhd_duration = int(mov_mdhd_duration * mvhd_timescale / mov_timescale) aac_tkhd_duration = int(aac_mdhd_duration * mvhd_timescale / aac_timescale) if mov_tkhd_duration >= aac_tkhd_duration: mvhd_duration = mov_tkhd_duration else: mvhd_duration = aac_tkhd_duration sample_size_tables = [] sample_size_tables.append([s for o, s in mov_table]) sample_size_tables.append([s for o, s in aac_table]) chunk_offset_tables = [] chunk_offset_tables.append([o for o, s in mov_table]) chunk_offset_tables.append([o for o, s in aac_table]) # moov structure is assumed to be in the fixed format (for now) mov_stsz_size = len(sample_size_tables[0])* 4 + 20 aac_stsz_size = len(sample_size_tables[1])* 4 + 20 # mov_co64_size = len(chunk_offset_tables[0])* 8 + 16 # aac_co64_size = len(chunk_offset_tables[1])* 8 + 16 mov_stco_size = len(chunk_offset_tables[0])* 4 + 16 aac_stco_size = len(chunk_offset_tables[1])* 4 + 16 mov_stss_size = ((len(sample_size_tables[0])-1)//150 + 1)* 4 + 16 # mov_stbl_size = 8 + 0x141 + 0x18 + 0x1C + mov_stsz_size + mov_co64_size + mov_stss_size # aac_stbl_size = 8 + 0x82 + 0x18 + 0x1C + aac_stsz_size + aac_co64_size mov_stbl_size = 8 + 0xAB + 0x18 + 0x1C + mov_stsz_size + mov_stco_size + mov_stss_size aac_stbl_size = 8 + 0x67 + 0x18 + 0x14B75C + aac_stsz_size + aac_stco_size + 0x1A + 0x1C mov_minf_size = 8 + 0x14 + 0x24 + mov_stbl_size aac_minf_size = 8 + 0x10 + 0x24 + aac_stbl_size # mov_mdia_size = 8 + 0x20 + 0x2E + mov_minf_size # aac_mdia_size = 8 + 0x20 + 0x2E + aac_minf_size mov_mdia_size = 8 + 0x20 + 0x2D + mov_minf_size aac_mdia_size = 8 + 0x20 + 0x2D + aac_minf_size # mov_trak_size = 8 + 0x5C + 0x24 + mov_mdia_size + 0x618 # aac_trak_size = 8 + 0x5C + 0x24 + aac_mdia_size mov_trak_size = 8 + 0x5C + 0x24 + mov_mdia_size aac_trak_size = 8 + 0x5C + 0x30 + aac_mdia_size # moov_size = 8 + 0x6C + 0x73 + mov_trak_size + aac_trak_size moov_size = 8 + 0x6C + 0x62 + mov_trak_size + aac_trak_size with open(ref_filename, 'rb') as f_moov, open(dst_filename, 'wb') as f_dst: f_moov.seek(0, 2) file_size = f_moov.tell() cur = 0 f_moov.seek(cur) if f_moov.tell() != cur: raise ValueError(f'seek failed? {f_moov.tell()} != {cur}') # moov copy_atom_box('moov', moov_size, f_moov, f_dst, only_header=True) #copy_atom_box('mvhd', None, f_moov, f_dst, only_header=False) # mvhd : duration = mov_tkhd_duration n = copy_atom_box('mvhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:12]) f_dst.write(struct.pack('>I', mvhd_timescale)) #duration = struct.unpack('>I', buf[16:20])[0] # f_dst.write(struct.pack('>I', mov_tkhd_duration)) f_dst.write(struct.pack('>I', mvhd_duration)) # the rest is unchanged f_dst.write(buf[20:]) #... #next_track_id = struct.unpack('>I', buf[96:100])[0] if n != (100+8): raise ValueError(f'ERROR: mov tkhd box size is not 108 but {n}') #copy_atom_box('udta', None, f_moov, f_dst, only_header=False) # movie track # trak copy_atom_box('trak', mov_trak_size, f_moov, f_dst, only_header=True) #copy_atom_box('tkhd', None, f_moov, f_dst, only_header=False) # tkhd : duration = mov_tkhd_duration n = copy_atom_box('tkhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:20]) #duration = struct.unpack('>I', buf[20:24])[0] f_dst.write(struct.pack('>I', mov_tkhd_duration)) # the rest is unchanged f_dst.write(buf[24:76]) #... #track_height = struct.unpack('>I', buf[80:84])[0] if n != (84+8): raise ValueError(f'ERROR: mov tkhd box size is not 92 but {n}') # copy_atom_box('edts', None, f_moov, f_dst, only_header=False) copy_atom_box('edts', None, f_moov, f_dst, only_header=True) n = copy_atom_box('elst', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:8]) f_dst.write(struct.pack('>I', mov_tkhd_duration)) f_dst.write(buf[12:]) if n != (20+8): raise ValueError(f'ERROR: mov elst box size is not 28 but {n}') # mdia copy_atom_box('mdia', mov_mdia_size, f_moov, f_dst, only_header=True) #copy_atom_box('mdhd', None, f_moov, f_dst, only_header=False) # mdhd : duration = mov_mdhd_duration n = copy_atom_box('mdhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:16]) #duration = struct.unpack('>I', buf[16:20])[0] f_dst.write(struct.pack('>I', mov_mdhd_duration)) # the rest is unchanged f_dst.write(buf[20:]) #... #quality = struct.unpack('>H', buf[22:24])[0] if n != (24+8): raise ValueError(f'ERROR: mov mdhd box size is not 32 but {n}') copy_atom_box('hdlr', None, f_moov, f_dst, only_header=False) # minf copy_atom_box('minf', mov_minf_size, f_moov, f_dst, only_header=True) copy_atom_box('vmhd', None, f_moov, f_dst, only_header=False) copy_atom_box('dinf', None, f_moov, f_dst, only_header=False) # stbl copy_atom_box('stbl', mov_stbl_size, f_moov, f_dst, only_header=True) copy_atom_box('stsd', None, f_moov, f_dst, only_header=False) #copy_atom_box('stts', None, f_moov, f_dst, only_header=False) # stts : sample_count = n_mov_table n = copy_atom_box('stts', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 1)) # n_entries f_dst.write(struct.pack('>I', n_mov_table)) # sample_count f_dst.write(struct.pack('>I', mov_sample_duration)) # sample_duration # stss n = copy_atom_box('stss', mov_stss_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags mov_stss_entries = (len(sample_size_tables[0])-1)//150 + 1 f_dst.write(struct.pack('>I', mov_stss_entries)) # n_entries ss = 1 for i_ss in range(mov_stss_entries): f_dst.write(struct.pack('>I', ss)) ss += 150 copy_atom_box('stsc', None, f_moov, f_dst, only_header=False) # stsz n = copy_atom_box('stsz', mov_stsz_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 0)) # sample_size f_dst.write(struct.pack('>I', len(sample_size_tables[0]))) # n_entries for sz in sample_size_tables[0]: f_dst.write(struct.pack('>I', sz)) # co64 # n = copy_atom_box('co64', mov_co64_size, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', len(chunk_offset_tables[0]))) # n_entries # for co in chunk_offset_tables[0]: # f_dst.write(struct.pack('>Q', co)) n = copy_atom_box('stco', mov_stco_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', len(chunk_offset_tables[0]))) # n_entries for co in chunk_offset_tables[0]: f_dst.write(struct.pack('>I', co)) # uuid # copy_atom_box('uuid', None, f_moov, f_dst, only_header=False) # audio track # trak copy_atom_box('trak', aac_trak_size, f_moov, f_dst, only_header=True) #copy_atom_box('tkhd', None, f_moov, f_dst, only_header=False) # tkhd : duration = aac_mdhd_duration n = copy_atom_box('tkhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:20]) #duration = struct.unpack('>I', buf[20:24])[0] f_dst.write(struct.pack('>I', aac_mdhd_duration)) # the rest is unchanged f_dst.write(buf[24:]) #... #track_height = struct.unpack('>I', buf[80:84])[0] if n != (84+8): raise ValueError(f'ERROR: audio tkhd box size is not 92 but {n}') # copy_atom_box('edts', None, f_moov, f_dst, only_header=False) copy_atom_box('edts', None, f_moov, f_dst, only_header=True) n = copy_atom_box('elst', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:12]) f_dst.write(struct.pack('>I', aac_tkhd_duration)) f_dst.write(buf[16:]) if n != (32+8): raise ValueError(f'ERROR: mov elst box size is not 40 but {n}') # mdia copy_atom_box('mdia', aac_mdia_size, f_moov, f_dst, only_header=True) #copy_atom_box('mdhd', None, f_moov, f_dst, only_header=False) # mdhd : duration = aac_mdhd_duration n = copy_atom_box('mdhd', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) # the following is unchanged f_dst.write(buf[:16]) #duration = struct.unpack('>I', buf[16:20])[0] f_dst.write(struct.pack('>I', aac_mdhd_duration)) # the rest is unchanged f_dst.write(buf[20:]) #... #quality = struct.unpack('>H', buf[22:24])[0] if n != (24+8): raise ValueError(f'ERROR: audio mdhd box size is not 32 but {n}') copy_atom_box('hdlr', None, f_moov, f_dst, only_header=False) # minf copy_atom_box('minf', aac_minf_size, f_moov, f_dst, only_header=True) copy_atom_box('smhd', None, f_moov, f_dst, only_header=False) copy_atom_box('dinf', None, f_moov, f_dst, only_header=False) # stbl copy_atom_box('stbl', aac_stbl_size, f_moov, f_dst, only_header=True) copy_atom_box('stsd', None, f_moov, f_dst, only_header=False) #copy_atom_box('stts', None, f_moov, f_dst, only_header=False) # stts : sample_count = n_aac_table n = copy_atom_box('stts', None, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 1)) # n_entries f_dst.write(struct.pack('>I', n_aac_table)) # sample_count f_dst.write(struct.pack('>I', aac_sample_duration)) # sample_duration # copy_atom_box('stsc', None, f_moov, f_dst, only_header=False) cur = f_dst.tell() n = copy_atom_box('stsc', None, f_moov, f_dst, only_header=True) f_dst.seek(cur) f_dst.write(struct.pack('>I', 28)) f_dst.seek(cur+8) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 1)) # n_entries f_dst.write(struct.pack('>I', 1)) f_dst.write(struct.pack('>I', 1)) f_dst.write(struct.pack('>I', 1)) # stsz n = copy_atom_box('stsz', aac_stsz_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', 0)) # sample_size f_dst.write(struct.pack('>I', len(sample_size_tables[1]))) # n_entries for sz in sample_size_tables[1]: f_dst.write(struct.pack('>I', sz)) # co64 # n = copy_atom_box('co64', aac_co64_size, f_moov, f_dst, only_header=True) # buf = f_moov.read(n-8) # f_dst.write(buf[:4]) # version + flags # f_dst.write(struct.pack('>I', len(chunk_offset_tables[1]))) # n_entries # for co in chunk_offset_tables[1]: # f_dst.write(struct.pack('>Q', co)) n = copy_atom_box('stco', aac_stco_size, f_moov, f_dst, only_header=True) buf = f_moov.read(n-8) f_dst.write(buf[:4]) # version + flags f_dst.write(struct.pack('>I', len(chunk_offset_tables[1]))) # n_entries for co in chunk_offset_tables[1]: f_dst.write(struct.pack('>I', co)) copy_atom_box('sgpd', None, f_moov, f_dst, only_header=False) copy_atom_box('sbgp', None, f_moov, f_dst, only_header=False) copy_atom_box('udta', None, f_moov, f_dst, only_header=False) if not full_copy: return # just copy the rest of reference moov file moov_cur = f_moov.tell() f_moov.seek(0, 2) moov_size = f_moov.tell() f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') if verbose: it_moov = tqdm(range(moov_cur, moov_size, n_chunk)) else: it_moov = range(moov_cur, moov_size, n_chunk) #for moov_cur in tqdm(range(moov_cur, moov_size, n_chunk)): for moov_cur in it_moov: f_dst.write(f_moov.read(n_chunk)) if moov_size - moov_cur > 0: f_dst.write(f_moov.read(moov_size - moov_cur)) # ## merging the recovered `moov` def merge_moov( src_filename, moov_filename, dst_filename, n_chunk=65536, verbose=False): with open(src_filename, 'rb') as f_src, open(moov_filename, 'rb') as f_moov, open(dst_filename, 'wb') as f_dst: f_src.seek(0, 2) file_size = f_src.tell() cur = 0 f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') # ftyp n, atom_type = read_atom_head(f_src) if atom_type != 'ftyp': raise ValueError('ftyp not found') f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') buf = f_src.read(n) f_dst.write(buf) if verbose: print_binaries(buf) cur += n cur += 16 # mdat n, atom_type = read_atom_head(f_src) if atom_type != 'mdat': raise ValueError('mdat not found') if n != 0: raise ValueError('size would be zero...') # fixed mdat header if verbose: print_binaries(struct.pack('>Icccc', 1, b'm', b'd', b'a', b't')) print_binaries(struct.pack('>Q', file_size-0x20)) f_dst.write(struct.pack('>Icccc', 1, b'm', b'd', b'a', b't')) f_dst.write(struct.pack('>Q', file_size-0x20)) cur += 16 f_src.seek(cur) if f_src.tell() != cur: raise ValueError(f'seek failed? {f_src.tell()} != {cur}') if verbose: it_cur = tqdm(range(cur, file_size, n_chunk)) else: it_cur = range(cur, file_size, n_chunk) #for cur in tqdm(range(cur, file_size, n_chunk)): for cur in it_cur: f_dst.write(f_src.read(n_chunk)) if file_size - cur > 0: f_dst.write(f_src.read(file_size - cur)) print('') # search moov f_moov.seek(0, 2) moov_size = f_moov.tell() if verbose: print(f'moov_size: {moov_size}') moov_cur = 0 f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') n, atom_type = read_atom_head(f_moov) if atom_type != 'moov': raise ValueError(f'something is wrong...') # copy moov f_moov.seek(moov_cur) if f_moov.tell() != moov_cur: raise ValueError(f'seek failed? {f_moov.tell()} != {moov_cur}') if verbose: it_moov = tqdm(range(moov_cur, moov_size, n_chunk)) else: it_moov = range(moov_cur, moov_size, n_chunk) #for moov_cur in tqdm(range(moov_cur, moov_size, n_chunk)): for moov_cur in it_moov: f_dst.write(f_moov.read(n_chunk)) if moov_size - moov_cur > 0: f_dst.write(f_moov.read(moov_size - moov_cur)) # # main program to recover corrupted MP4 def finsta360( moov_const, src_filename, ref_filename=None, dst_filename=None, keep_temp=False, verbose=False): if ref_filename is None: # check mode print_atoms(src_filename) return # temporary files ref_moov_filename = 'finsta360_ref.moov' new_moov_filename = 'finsta360_new.moov' # 1) extract reference moov print('') print('########################################') print(f'# 1) extracting reference moov from\n\t{ref_filename}') extract_moov(ref_filename, ref_moov_filename) if verbose: print_atoms(ref_moov_filename) # 2) regenerate sample tables from mdat print('') print('########################################') print(f'# 2) regenerate sample tables from mdat in\n\t{src_filename}') mov_table, aac_table = recover_sample_tables_from_mdat_fast( src_filename, verbose=True) sys.exit() if verbose: print(f'number of samples (movie) : {len(mov_table)}') print(f'number of samples (audio) : {len(aac_table)}') # constants # mov_sample_duration = 1001 # aac_sample_duration = 1024 # mvhd_timescale = 48000 # mov_timescale = 30000 # aac_timescale = 48000 mov_sample_duration = moov_const[0] aac_sample_duration = moov_const[1] mvhd_timescale = moov_const[2] mov_timescale = moov_const[3] aac_timescale = moov_const[4] n_mov_table = len(mov_table) n_aac_table = len(aac_table) mov_mdhd_duration = n_mov_table * mov_sample_duration aac_mdhd_duration = n_aac_table * aac_sample_duration mov_tkhd_duration = int(mov_mdhd_duration * mvhd_timescale / mov_timescale) # mvhd mvhd_duration_sec = mov_tkhd_duration / mvhd_timescale print(f'mvhd duration : {mvhd_duration_sec} sec / {mvhd_duration_sec/60} min') # movie mdhd mov_duration_sec = mov_mdhd_duration / mov_timescale print(f'movie duration : {mov_duration_sec} sec / {mov_duration_sec/60} min') # audio mdhd aac_duration_sec = aac_mdhd_duration / aac_timescale print(f'audio duration : {aac_duration_sec} sec / {aac_duration_sec/60} min') # 3) rebuilding moov from the sample tables print('') print('########################################') print(f'# 3) rebuilding moov from the sample tables') recover_moov_from_sample_tables( moov_const, ref_moov_filename, new_moov_filename, mov_table, aac_table, full_copy=True, ) if verbose: print_atoms(new_moov_filename) if dst_filename is None: # test mode if not keep_temp: os.remove(ref_moov_filename) os.remove(new_moov_filename) return # 4) merging the rebuilt moov into the source print('') print('########################################') print(f'# 4) merging the rebuilt moov into\n\t{src_filename}\nas\n\t{dst_filename}') merge_moov( src_filename, new_moov_filename, dst_filename, ) if not keep_temp: os.remove(ref_moov_filename) os.remove(new_moov_filename) def usage(): print('finsta360.py : to finalize incomplete MP4 of Insta360 ONE-X') print('https://github.com/kichiki/finsta360') print('USAGE: finsta360.py [options]') print('\t-s file : source file, that is, corrupted mp4 (insv) file') print('\t-r file : complete mp4 (insv) file as a reference') print('\t-o file : output recovered mp4 (insv) file') print('\t-v : to set verbose mode') print('\t-k : to keep temporary files') print('\t (reference and recovered moov files, finsta360*.moov)') print('If you provide only source file (-s), program prints the metadata') print('If you dont provide output file (-o), program just runs without writing') sys.exit () if __name__ == '__main__': # src_filename = None # ref_filename = None # dst_filename = None # verbose = False # keep_temp = False # i = 1 # while i < len(sys.argv): # if sys.argv[i] == '-s': # src_filename = sys.argv[i+1] # i += 2 # elif sys.argv[i] == '-r': # ref_filename = sys.argv[i+1] # i += 2 # elif sys.argv[i] == '-o': # dst_filename = sys.argv[i+1] # i += 2 # elif sys.argv[i] == '-v': # verbose = True # i += 1 # elif sys.argv[i] == '-k': # keep_temp = True # i += 1 # else: # usage() # break # if src_filename is None: # print(f'you must provie source file {src_filename}') # usage() # sys.exit() # if not os.path.exists(src_filename): # print(f'source file {src_filename} does not exist') # sys.exit() # if not ref_filename is None and not os.path.exists(ref_filename): # print(f'reference file {ref_filename} does not exist') # sys.exit() # if not dst_filename is None and os.path.exists(dst_filename): # print(f'output file {dst_filename} already exists') # sys.exit() # # constants # mov_sample_duration = 3000 # aac_sample_duration = 1024 # mvhd_timescale = 1000 # mov_timescale = 90000 # aac_timescale = 48000 # moov_const = (mov_sample_duration, # aac_sample_duration, # mvhd_timescale, # mov_timescale, # aac_timescale) # finsta360( # moov_const, # src_filename, # ref_filename, # dst_filename, # keep_temp, # verbose) # sys.exit() if len(sys.argv) != 3: print(f'Usage: python {sys.argv[0]} in.mp4 out.aac') sys.exit(1) filename_in = sys.argv[1] filename_out = sys.argv[2] _, aac_table = recover_sample_tables_from_mdat_fast(filename_in, verbose=True) with open(filename_in, 'rb') as f_in, open(filename_out, 'wb') as f_out: # f_out.seek(7) # num = 0 for i in range(len(aac_table)): offset, length = aac_table[i] f_in.seek(offset) for j in range(length): f_out.write(f_in.read(1)) # num += 1 # MACHIGAI # frame_length = 7 + num # buf_int = [0] * 7 # buf_int[0] = 0xFF # buf_int[1] = 0xF1 # buf_int[2] = 0x4C # buf_int[3] = 0b10000000 | ((frame_length & 0b11_00000000_000) >> 11) # buf_int[4] = ((frame_length & 0b00_11111111_000) >> 3) # buf_int[5] = 0b00011111 | ((frame_length & 0b00_00000000_111) << 5) # buf_int[6] = 0b11111100 # f_out.seek(0) # f_out.write(bytes(buf_int))
付録2
（正常な）MP4からサンプルの最初の6バイトをビット列で表示するコードです。AACのフレーム境界を判別できないかと作りました。stscを読み解く部分が大変でした。
chunk.py
#!/usr/bin/env python from datetime import datetime from datetime import timedelta import os.path import struct import sys def parse_box(f): buf = f.read(8) box_size = struct.unpack('>I', buf[:4])[0] box_type = str(buf[4:], 'utf-8') if box_size == 1: buf = f.read(8) box_size = struct.unpack('>Q', buf)[0] return box_size, box_type def main(filename_in): containers = ('moov', 'trak', 'edts', 'mdia', 'minf', 'dinf', 'stbl') with open(filename_in, 'rb') as f_in: f_in.seek(0, 2) file_size = f_in.tell() cur = 0 sc_table = [] sz_table = [] co_table = [] while True: f_in.seek(cur) box_size, box_type = parse_box(f_in) if box_size == 0: box_size = file_size - cur if box_type in containers: box_size, box_type = parse_box(f_in) if box_size == 0: box_size = file_size - cur cur += 8 elif box_type == 'stsc': # Sample-to-Chunk Atoms buf = f_in.read(box_size - 8) version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] for i in range(n_entries): i0 = 8 + i*12 i1 = i0 + 12 if len(buf) < i1: break first_chunk = struct.unpack('>I', buf[i0:i0+4])[0] samples_per_chunk = struct.unpack('>I', buf[i0+4:i0+8])[0] sample_desc_id = struct.unpack('>I', buf[i0+8:i0+12])[0] sc_table.append((first_chunk, samples_per_chunk, sample_desc_id)) # print(f'{i:6d}: {(first_chunk, samples_per_chunk, sample_desc_id)}') elif box_type == 'stsz': #Sample Size Atoms buf = f_in.read(box_size - 8) version = buf[0] flags = buf[1:4] sample_size = struct.unpack('>I', buf[4:8])[0] n_entries = struct.unpack('>I', buf[8:12])[0] for i in range(n_entries): i0 = 12 + i*4 i1 = i0 + 4 if len(buf) < i1: break size = struct.unpack('>I', buf[i0:i1])[0] sz_table.append(size) # print(f'{i:6d}: {size}') elif box_type == 'stco': #Chunk Offset Atoms buf = f_in.read(box_size - 8) version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] for i in range(n_entries): i0 = 8 + i*4 i1 = i0 + 4 if len(buf) < i1: break offset = struct.unpack('>I', buf[i0:i1])[0] co_table.append(offset) # print(f'{i:6d}: {offset}') elif box_type == 'co64': #64-bit chunk offset atoms buf = f_in.read(box_size - 8) version = buf[0] flags = buf[1:4] n_entries = struct.unpack('>I', buf[4:8])[0] for i in range(n_entries): i0 = 8 + i*8 i1 = i0 + 8 if len(buf) < i1: break offset = struct.unpack('>Q', buf[i0:i1])[0] co_table.append(offset) # print(f'{i:6d}: {offset}') if len(sc_table) != 0 and len(sz_table) != 0 and len(co_table) != 0: print('######################## ########################') i = 0 l = 0 while True: m0, n, _ = sc_table[i] if i + 1 < len(sc_table): m1 = sc_table[i + 1][0] else: m1 = len(co_table) + 1 j = m0 - 1 while True: offset = co_table[j] k = 0 while True: cur_temp = f_in.tell() f_in.seek(offset) buf = bytes([0x00, 0x00]) + f_in.read(6) f_in.seek(cur_temp) binary = struct.unpack('>Q', buf)[0] mark = ' ' if sz_table[l] < 100: mark = 'v' print(f'{mark}{offset:10d} {sz_table[l]:6d} {binary:059_b}') offset += sz_table[l] l += 1 k += 1 if k >= n: break j += 1 if j >= m1 - 1: break i += 1 if i >= len(sc_table): break print('') sc_table = [] sz_table = [] co_table = [] cur += box_size if cur >= file_size: break if __name__ == '__main__': if len(sys.argv) != 2: print(f'Usage: python {sys.argv[0]} in.mp4') sys.exit(1) filename_in = sys.argv[1] main(filename_in)
