search

chenqi1126 / SIPE

[CVPR 2022] Self-supervised Image-specific Prototype Exploration for Weakly Supervised Semantic Segmentation
MIT License
73 stars 10 forks source link

The evaluation of COCO14 segmentation masks #13

Open mt-cly opened 1 year ago

mt-cly commented 1 year ago

Hi, thanks for your great work. Your SIPE achieves considerable performance in COCO14 val set. I wonder how to obtain the image GT for COCO14 so that I can calculate miou as done in https://github.com/chenqi1126/SIPE/blob/main/eval_cam.py#L52 . As the COCO official only provides the .json file, can you please tell me how to translate the .json to .png? Thanks.

chenqi1126 commented 1 year ago

Hi @mt-cly, Thanks for your attention. You can use 'pycocotools' to convert annotation to png. An off-the-shelf code: https://github.com/zhaozhengChen/ReCAM/blob/main/mscoco/annToMask.py.

mt-cly commented 1 year ago

Thanks for your reply.

mt-cly commented 1 year ago

I meet another question, I directly use your provide deeplab_r38_coco.pth on the modified code of seamv1, but I can only achieve around 36% mIoU, (+CRF ~37% mIoU), still has a large gap with reported 43.6%, I have no idea about that, have you any suggestions? Or can you provide the evaluation code for COCO14 val set? Thanks.