Bing Shuai

Technically, you can freeze the backbone, RPN as well as the detection branch so that you only train Siamese tracker alone. But as our experiments show, this will lead to...

The loss for tracker_center usually does not drop much, the values you provide looks normal to me.

Yes, these three thresholds (here: https://github.com/amazon-research/siam-mot/blob/main/siammot/configs/defaults.py#L54-L56) are probably what you need to tune.

My initial guess is that the gap may comes from the fact you use "max_size_train == 1200". I haven't tested it with model training on 4 GPUs, and I don't...

You can follow the command listed here: https://github.com/amazon-research/siam-mot#dataset-evaluation-and-training. You only need to use your own train_dir and model_suffix. Training the whole network on 8 GPUs (V100) with the default configuration...

Yes, in the demo, we provide the models that can track person as well as other multiple vehicle instances.

Not sure whether you've loaded the ground truth correctly, as all predictions are considered as "False Positive (FP)", and there is no False Negative (FN) and ID switches. This means...

Not sure wether I understood your question correctly. If you are talking about doing spatial matching within the motion model, you should be able to replace the one in the...

Hi, yes, you can do something similar, and IMM should be pretty straightforward to implement.

This refinement is simply to refine the location of the target objects from motion model. We can think that the motion model is not very good regressing its exact location.