search

krumo / Domain-Adaptive-Faster-RCNN-PyTorch

Domain Adaptive Faster R-CNN in PyTorch
MIT License
304 stars 68 forks source link

Serious bug: adding target domain annotation boxes into rpn proposals #10

Closed Liyw979 closed 4 years ago

Liyw979 commented 4 years ago

https://github.com/krumo/Domain-Adaptive-Faster-RCNN-PyTorch/blob/276fcf8ae41be483db94682722c8093703750fdd/maskrcnn_benchmark/engine/trainer.py#L137-L143

https://github.com/krumo/Domain-Adaptive-Faster-RCNN-PyTorch/blob/276fcf8ae41be483db94682722c8093703750fdd/maskrcnn_benchmark/modeling/rpn/inference.py#L147-L148

https://github.com/krumo/Domain-Adaptive-Faster-RCNN-PyTorch/blob/276fcf8ae41be483db94682722c8093703750fdd/maskrcnn_benchmark/modeling/rpn/inference.py#L51-L72

krumo commented 4 years ago

Hi @liyiwei979621500 Thanks for pointing it out. Actually during training you should generate target domain json with pseudo annotations like using this script. According to my experience, there is no difference on performance between using ground-truth annotations and pseudo annotations because of the random sampling strategy. But you are right, the gt_box should be filtered out if you plan to use the ground-truth annotations. This bug has been fixed in the recent commit.