Closed fangjin-cool closed 3 years ago
The orpn_r50_fpn_1x_dota10.py only contains "orpn" which generates the oriented proposals for the second stage. It only classifies the samples into the foregrounds and backgrounds.
The entire model of oriented rcnn is faster_rcnn_orpn_r50_fpn_1x_dota10.py.
Got it, thank you very much!
Thanks for your great work!
Is the code ready for Oriented R-CNN? When I use the model config file "obb/orpn_r50_fpn_1x_dota10.py", It's OK for training. But when I want to do the inference and submit it to the testing server (DOTA-v1.0), I found the results are very bad, the mAP is only 0.026. The visualization result is shown below if I set the "args.show" and "args.show_dir" in test.py.
I find that for each patch, there are 2000 detection results even after nms operation, and many of them are with very high scores. Do I miss something or is the code still under development for "obb/orpn_r50_fpn_1x_dota10.py" setting?
I also found that the output anchors of FPN have 5 levels and they are using independent nms, and the horizontal instead of the rotation bbox are sending to the nms function, why is that?
Thanks very much and looking forward to your reply.