Open chenrxi opened 3 years ago
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 lower tracking performance. What's the use cases in your situation? Why you only want to train Siamese tracker rather than the whole network?
can we use another backbone other then like maskrcnn?
is it possible to train siamese tracker without training faster rcnn?