Tian Zhi

@Cynthia123liu Please check your data to see if there are some images with no ground-truth boxes.

@Mahmood-Hussain It's correct. You need to upload your results to CodaLab to get the mAP on test-dev. Please refer to https://cocodataset.org/#upload

@liushuai839 `label` contains the label num and will be converted to one-hot vectors.

@LMerCy It should be trained with the centerness branch.

@Wufang1997 Yes. We tried this, but it only results in minor improvements on COCO.

@Zzh-tju Please check your evaluation code.

@MsRabbits Running it on Windows should be OK. FCOS is written by python and built on PyTorch, so it is platform-independent.

@Len-Li I suggest that you can try to clean all the compiled files (or re-clone the project) and recompile the project.

@xuetingdai please try the latest code.

@1356789-rr there seems a typo. I think `DATASETS.RAIN` should be `DATASETS.TRAIN`.