ResNet101 code

[GuoleiSun]

Hi Peng-Tao, In your code, only vgg16 is provided. Could you also provide resnet101? Thanks

[PengtaoJiang]

Hello, the accumulation process for vgg16 is the same with resnet101, you just need to replace the vggnet with the resnet. I will consider uploading the code later.

[Ferenas]

Hello, the accumulation process for vgg16 is the same with resnet101, you just need to replace the vggnet with the resnet. I will consider uploading the code later.

Thanks for your great work ! Have you tested your model in ResNet50 or other backbones and what performance do they achieve ?

[PengtaoJiang]

I have not tried this. I could test the accumulation process for resnet50. When I obtain the performance, I will tell you. Thanks for following our work.

[Ferenas]

I have not tried this. I could test the accumulation process for resnet50. When I obtain the performance, I will tell you. Thanks for following our work.

Hi, thanks for your reply! I was wondering how you test your CAMs made by OAA, here you did not provide the test file for the mIoU of your CAM (58.6% in your paper). How do you evaluate the background score? Do you follow the common sense that treating the bg score as 1-max(foreground_score)?

[PengtaoJiang]

58.6% is not the mIoU of CAM. It is the mIoU of the segmentation network trained using the pseudo labels generated by CAM. In the gen_gt.py file, we utilize the 1-saliency as the background cues, not 1-max(foreground_score).

[Ferenas]

58.6% is not the mIoU of CAM. It is the mIoU of the segmentation network trained using the pseudo labels generated by CAM. In the gen_gt.py file, we utilize the 1-saliency as the background cues, not 1-max(foreground_score).

Thanks for your reply! However, I was more confused about it. In Tab. 1, the mIoU of segmentation masks reached 61.6% in val set, but here 58.6% (data in Tab. 2) is also the segmentation masks. Could you provide more details about it ? In WSSS, the CAM or refined CAM is firstly generated, then is treated as the pseudo masks as the semantic segmentation. Therefore, I was confused about your claim.

[PengtaoJiang]

Hi, I have written that the ablation experiments are all conducted on a single test in the paper. The final segmentation results are evaluated on a multi-scale test with CRF (58.6 -> 61.6). Moreover, the CAM and refined CAM are first used to generate pseudo segmentation labels, then the segmentation labels are used to train deeplab.

[Ferenas]

I got that！Thanks for your reply！

发自我的iPhone

在 2021年6月27日，下午12:42，Jiangpengtao @.***> 写道：

 Hi, I have written that the ablation experiments are all conducted on a single test in the paper. The final segmentation results are evaluated on a multi-scale test with CRF (58.6 -> 61.6). Moreover, the CAM and refined CAM are first used to generate pseudo segmentation labels, then the segmentation labels are used to train deeplab.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.