search

jiwoon-ahn / irn

Weakly Supervised Learning of Instance Segmentation with Inter-pixel Relations, CVPR 2019 (Oral)
MIT License
519 stars 100 forks source link

Performance is poor after re-train a Mask RCNN #6

Closed bityangke closed 5 years ago

bityangke commented 5 years ago

Hi, I took the instance-level pseudo labels generated by running `make_ins_seg_labels.py' and kept the instance mask whose score is higher than 0. Then, I transfered these labels from *.npy to cocostyle json annotation and trained the standard Mask R-CNN with ResNet-50-FPN. However, the performance I've get is:

image

Specifically, box mAP of AP50 is 45.8, segmentation mAP of AP50 is 22.6. I noticed that the instance number in pseudo label is about 2/3 of the gt instance number for `train_aug' set. Did I miss something to reimplement the performance of Mask R-CNN with pseudo label?

Thanks a lot!

jiwoon-ahn commented 5 years ago

Hi @bityangke, It is hard to give you a straight-forward answer to your question. Coco mask annotation is composed of polygons, not pixel-wise instance labels. Hence, there are many ways of transfering the psuedo labels for mask rcnn. Please make sure you did not make any mistakes during the conversion process. Plus, the mask rcnn should be pretrained on imagenet.

bityangke commented 5 years ago

Thanks very much！

discretecoder commented 4 years ago

@bityangke did you fix this issue? do you have a code for label transformations? thanks