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

comparation with Affinity #19

Closed mt-cly closed 4 years ago

mt-cly commented 4 years ago

Hi, In train_irn step, I remove the dispalce loss part and remains only boundary loss. I notice boundary loss is similar to the AffinityNet which you published in CVPR18 even the detail has some differents. But the semantic mIoU only 37+% which is even worse than CAM result(50%)，comared to Affinity result(59%)。 So I confuse the reason for such gap in same idea, similar loss. Have you some suggests? THX

jiwoon-ahn commented 4 years ago

The displacement loss has no effect on the process of generating semantic segmentation masks. Perhaps, you have mistaken mAP with mIOU?

mt-cly commented 4 years ago

The displacement loss has no effect on the process of generating semantic segmentation masks. Perhaps, you have mistaken mAP with mIOU?

I am sorry, I found the reason of that cause some wrong operations, and the isolated boundary loss would help semantic mIoU achive about 67% as well. thanks for reply.