I think you can try SAM3

[Ying0524]

I have collected some videos and used the project for generation, but I found that the actual results were not particularly good. The YOLO model makes errors when localizing objects in certain frames, and there doesn’t seem to be a way to manually correct them—or perhaps I just haven't found the option. Recently, I noticed an open-source model called SAM3, which does an excellent job at image segmentation. However, its recognition efficiency isn’t very high and it can’t achieve real-time processing. On an RTX 4070 Ti, it takes about 0.2 seconds per image. But we could try extracting keyframes for prediction. Bro, do you think this would be feasible?