Open zengwz opened 1 year ago
It seems that your model doesn't converge well. How did you add L1 loss?
According to my understanding, it should depend on "YOLOXModeSwitchHook". It will close the mosaic and mixup in the last num epochs and add L1 LOSS. the concrete implementation depends on the official implementation code of yolox.
Could you share how you processed the BDD100K dataset in mmtracking and where did you download the BDD100K dataset? Thank you very much!
Could you share how you processed the BDD100K dataset in mmtracking and where did you download the BDD100K dataset? Thank you very much!
I processed BDD100K dataset follow QDTrack(https://github.com/SysCV/qdtrack). And you can get the BDD100K dataset from its official website.
I train the ByteTrack using official config, only modifying some hyperparameters and the input dataset. And it performs well before adding the L1 Loss. After adding the L1 Loss, the test result in the 4th epoch is:
the config is: