Speed up the video splitting when prepare the training data

[gulucaptain]

The original speed of video splitting in /easyanimate/video_caption/video_splitting.py is very slow, due to the 'ffmpeg' operation.

Improvement:
change the FFmpeg code in Line 17-32 in </easyanimate/video_caption/video_splitting.py> to:

return [
'ffmpeg',
'-ss', str(start_time.time()),
'-i', video_path,
'-t', str(video_duration),
'-c:v', 'copy',
'-c:a', 'copy',
output_path
]

This will speed up the splitting process.

[hkunzhe]

@gulucaptain Directly copy video/audio streams without re-encoding would indeed be faster, but this would sacrifice the precision, resulting in incorrect start /end timestamps or scene transitions in the video clips. Thus, I do not recommend doing this.

[gulucaptain]

@gulucaptain Directly copy video/audio streams without re-encoding would indeed be faster, but this would sacrifice the precision, resulting in incorrect start /end timestamps or scene transitions in the video clips. Thus, I do not recommend doing this.

Thanks, in my experiments, I don't observe obvious error start/time timestamps in my splitting. And I also did not check the accurate index of frames after the splitting. However, the running speed difference between the two functions is huge. After modification, I can process 1,000 videos only with 9.x seconds. So, if there is actually a problem with precision, we should consider the trade-off between the precision and the running time.