The performance is lower than reported ?

[alex-marked]

Amazing Work! When I used the provided code and followed the method in your paper, I only got val 64.9 and test 65.2 mIoU, respectively. How can I get the same performance as reported in your paper？

[dongzhang89]

Thanks for your attention. From my experience, it looks like that you did not initialize the segmentation model with the ImageNet pre-trained weight ? Besides, using the segmentation model trained parameters in the new round for classification will also beneficial.

[alex-marked]

@ZHANGDONG-NJUST Thx for your reply.

[kkechen]

Thanks for your attention. From my experience, it looks like that you did not initialize the segmentation model with the ImageNet pre-trained weight ? Besides, using the segmentation model trained parameters in the new round for classification will also beneficial.

Thank you for your answer! I downloaded the pretrained resnet-101 model on the pytorch website according to your tips, but I didn't get the desired result. Could you tell me how to download your imagenet pretrained model you mentioned above?

[SunGuoYing]

Thanks for your attention. From my experience, it looks like that you did not initialize the segmentation model with the ImageNet pre-trained weight ? Besides, using the segmentation model trained parameters in the new round for classification will also beneficial.

Hello,I am very confused about how to use the parameters of the segmentation model training to classify in the new round

[dongzhang89]

@SunGuoYing The classification model and the segmentation model share the same backbone network. Using the segmentation weight in a new round classification is beneficial for generating CAMs.

[dongzhang89]

@hiddlephila I just used the common pre-trained model in pytorch.