Giters

clovaai / BESTIE

[CVPR 2022] Beyond Semantic to Instance Segmentation: Weakly-Supervised Instance Segmentation via Semantic Knowledge Transfer and Self-Refinement

Repository from Github https://github.comclovaai/BESTIERepository from Github https://github.comclovaai/BESTIE

About maskrcnn？

ZechengLi19 opened this issue · comments

commented

Thanks for your perfect work. I want to reproduce your work. However, when I apply mask rcnn on it, I found the result is not match the result you provided. On VOC2012, map50 --> 48.3, On COCO2017, map50 --> 22.2. Can you tell me, how to reproduce the result on mask rcnn? By the way, I use mask_rcnn_r50_fpn_1x on mmdetection as my mask rcnn config.

commented

Hi, I also meet the same problem now, have you solved this?

commented

Hi, I also meet the same problem now, have you solved this?

Sorry, I just give up.

commented

Hi！I want to reproduce this work, too. Can you tell me how to get BESTIE's result for maskrcnn refinement, even though the map50 is 48.3.

commented

Hi！I want to reproduce this work, too. Can you tell me how to get BESTIE's result for maskrcnn refinement, even though the map50 is 48.3.

Two tips.

(1) alpha of PAM, you need to visualize PAM to select a better alpha for coco dataset.
(2) WSSS maps with CRF is better.

commented

Thanks！But I have another problem. How can I obtain the instance segmentation results on the training set using my trained model?

commented

Thanks！But I have another problem. How can I obtain the instance segmentation results on the training set using my trained model?

This repository does not seem to provide this function. You need to implement this function yourself.

commented

Thank you！Could you briefly explain how to use the trained bestie model to obtain predicted masks on the training set? And I have another question. Are the training and validation images used for training Mask R-CNN the same as those used for training and validating bestie?