prefetching in video annotations #9519
Description
Actions before raising this issue
- I searched the existing issues and did not find anything similar.
- I read/searched the docs
Is your feature request related to a problem? Please describe.
20250610-0604-27.4260854.mp4
Please see the video. every 36 frames, a buffering event occurs.
This happens in all of the following cases and/or combinations
- video/images compressed to 7% size (see quality)
- on official cloud/on local machine with direct link to files
- as video in multiple bitrates
- I frames every 1, 10 , 100 frames
- as a list of frames in a folder
Same delay in all cases above, no matter the host server, no matter image size, no matter formatting.
The above leads me to believe this is inherent to the CVAT architecture:
- there is no prefetching
- when there is a fetch event, it only fetches the next 36 frames.
- Fetching always takes the same time, leading me to believe the bottleneck is communication related in the architecture, and not data related.
For sparse-tagging (such as video), this makes CVAT unusable.
I looked everywhere for a workaround or a solution, none to be found.
Describe the solution you'd like
Any video player needs prefetching. I want to have the video play with no (or very little) buffering in normal speed. X2 can be great.
The next frames should be loaded BEFORE the person lands on them, in video-play mode, as well as in frame-by-frame mode.
When clicking D, D, D, D ... and getting to frame 35, frame 36 is still not loaded, even if very slow. this shouldn't be the case.
When going to frame 36 and preloading starts and going back to frame 35, one would expect the next chunk to load in the background. Not the case now. Loading stops and you have to wait twice.
Describe alternatives you've considered
Paying for different software. This is unusable for video tagging.