Xingyi Zhou

Hi, Sorry for the delayed response. Can you specify your command?

Thank you for your interest. I don't have such a plan. But you can just copy-paste the code under projects/CenterNet2 to your detectron2 repo and use it as is.

Can you show your pytorch/ cuda/ cudnn/ detectron2 version? You can find them on the very beginning of the output log. The output log should also show the runtime breakdown...

It's likely your terminal and IDE have different environments. Can you try to install a new conda environment and use the same one for both?

Hi, Thank you for your interest and sorry for my delayed response. This means the training diverged. Common ways to avoid this issue are to increase the warmup iterations (SOLVER.WARMUP_ITERS)...

Hi, Great question! This is mostly designed for LVIS, where the objects are not fully annotated, and the high FPs are sometimes true objects. It makes an ~0.5 mAP difference...

Hi, That's interesting. Can you try to manually load the annotation file and check the number of images? Can you check your config file only contains that single dataset in...

Hi, Thanks a lot for your effort in reproducing our work on mmdetection and sorry for my very delayed response (I didn't check Github issues regularly. For detailed or urgent...

Hi, Thank you for your interest and sorry for my delayed response. The motivation to init bias to 8 is to provide initialization for the IoU-based setting (when MODEL.CENTERNET.MORE_POS is...

Thank you for your interest. I haven't tried to train on VOC with this code. I guess you can follow the instructions in the original detectron2. I.e., place your data...