Concat overlapping videos with ffmpeg using python3

Question

I have a number of videos that overlap in time that were downloaded from an online video stream: 0m00s - 15m10s 14m50s - 30m20s ... The first few seconds (usually around 20 seconds) of a video segment duplicate last few seconds of the previous segment.

I'd like to concat the video segments and form a single video file without duplicate parts. I can do it programmatically using ffmpeg libraries to find the cut points by minimizing error between a frame of the current segment and the last frame of the previous segment.

Is there a way to achieve this with an ffmpeg command?

Answer 1

I wrote the following script to solve my problem. It uses framemd5 checksums to find the inpoints in time and then uses the concat demuxer (thanks @Gyan for the tip!).

# Usage: ffmpeg $(python3 script.py input1.mp4 input2.mp4) -y -an output.mp4

import os
import sys
import subprocess

ffmpeg_concat_echo, md5_last_frame = '', None

for video_path in map(os.path.abspath, sys.argv[1:]):
        framemd5 = subprocess.check_output(['ffmpeg', '-nostats', '-hide_banner', '-i', video_path, '-an', '-f', 'framemd5', '-c', 'copy', '-'], stderr = open(os.devnull, 'w'))
        tb_num, tb_den = 1, 1
        inpoint = 0.0
        for line in framemd5.split(b'\n'):
                if line.startswith(b'#tb'):
                        tb_num, tb_den = list(map(int, line.split()[-1].split(b'/')))
                elif line and not line.startswith(b'#'):
                        splitted = line.split(b',')
                        md5 = splitted[-1].strip()
                        pts_time = float(splitted[2]) * tb_num / tb_den
                        if md5_last_frame is not None and md5 == md5_last_frame:
                                inpoint = pts_time
        md5_last_frame = md5
        ffmpeg_concat_echo += 'echo "file {}"; echo "inpoint {}";'.format(video_path, inpoint);

print('-nostats -hide_banner -avoid_negative_ts make_zero -fflags +genpts -f concat -safe 0 -i <({}) -c copy'.format(ffmpeg_concat_echo))

Below is another version (source, archived) which supports exact PTS inpoints using concat video filter; computing inpoints using explicit frame differences.

# Usage: ffmpeg $(( echo input1.mp4; echo input2.mp4 ) | python3 concat_overlapping_video_framemd5_demuxer.py concat_file.txt) -y -an output.mp4

import os
import sys
import subprocess

ffmpeg_concat_file, md5_last_frame = open(sys.argv[1], 'w'), None
for video_path in map(os.path.abspath, list(map(str.strip, sys.stdin))):
        framemd5 = subprocess.check_output(['ffmpeg', '-nostats', '-hide_banner', '-i', video_path, '-an', '-f', 'framemd5', '-c', 'copy', '-'], stderr = open(os.devnull, 'w'))
        inpoint, tb_num, tb_den = 0, 1, 1
        for line in framemd5.split(b'\n'):
                if line.startswith(b'#tb'):
                        tb_num, tb_den = list(map(int, line.split()[-1].split(b'/')))
                elif not line.startswith(b'#') and len(line) > 0:
                        splitted = line.split(b',')
                        md5, pts_time = splitted[-1].strip(), float(splitted[2]) * tb_num / tb_den
                        if md5 == md5_last_frame:
                                inpoint = pts_time
        md5_last_frame = md5
        ffmpeg_concat_file.write('file {}\ninpoint {}\n'.format(video_path, inpoint))

print('-nostats -hide_banner -avoid_negative_ts make_zero -fflags +genpts -f concat -safe 0 -i {} -c copy'.format(sys.argv[1]))