Issue with MOT17

[HemangParmar17]

Issue with MOT17

[ocetintas]

Hi,
Unfortunately, this code base is not directly compatible with MOT17 training or evaluation so you might need to fix some bugs if you want to apply it for MOT.

I'd suggest you use the original mot_neural_solver repo for MOT because the MOTS20 dataset doesn't contain all MOT17 sequences; therefore, you won't be able to train your model on the full MOT17 dataset with this code base. In addition, MPNTrack has some implementation details that give you a boost for MOT such as frame sampling which is not used here for MOTS. I am not sure why you got a lower accuracy for MOT17 with mot_neural_solver repo but tracking scores (MOTA, IDF1) should be similar to the ones reported in the CVPR paper.

[HemangParmar17]

Thank you for the response.
As you explained, MPNTrackSeg is not compatible with MOT17. Can you please explain frame sampling or how to implement it in the mot-neural-solver? if you can give us some hints or ideas about the code it would be a great help.