Question about the training of MoMA-M3T

[ChenN-Scott]

Hello! thank for your dedication and hard work on this open-source project!

I am interested in the training part of your work, including the design of loss and the mask for invalid loss backward(such as invalid but having the highest score match result should not be regarded as a tracklet to match subsequent detections). However, I do not find relevant explanation in your code or paper. I have several questions below:

1. In training, if a tracklet match a detection by searching the highest score but they do not match in label, will this unmatched detection be regarded as a part of this tracklet and be encoded to participate in the later calculation?
2. How do you handle the situation that some tracklets lacking matched detection in some timesteps but need to be encoded by temporal encoder?

Looking forward to your reply!
Best wishes!