halbielee/eps

Official PyTorch implementation of "Railroad is not a Train: Saliency as Pseudo-pixel Supervision for Weakly Supervised Semantic Segmentation", CVPR2021

eps's Issues

What is the novelty of your paper?

I can not understand why this paper been accepted? Adopting saliency map whether in supervised or unsupervised manner as additional supervision is a cheat.

COCO training codes too.

COCO segmentation code

Hi authors,
deeplab-pytorch only support code for VOC and COCO-Stuff, how can I use this repo to run segmentation on MS-COCO 2014?

Where is the code？😅

Regarding portnig the code to google colab

Are there any plans for writnig a google colab notebook that allows us to see how the inference is done,similar to facebook's detr google colab

About the color map of coco 2014

Thanks for your wonderful work on WSSS.

Could you provide the color map (palette) in your work (on MS coco 2014) for segmentation visualization?

The coco performance 37.15 is just the performance of the pseudo mask?

Train EPS on other dataset.

I want to know about how to train EPS on coco17.
I think I should get the cls_labels.npy in the file metadata/coco.
Can you tell me how to transfer the official json to the .npy?

hyper-parameters about CRF?

Hi! Thanks for you code! This repo doesn't provided detailed hyper-parameters (i.e, crf_t, alpha) when applying dCRF to the pseudo CAM label, can you provide them?

DeepLab setting

Fellow your step, I get the mIOU about 68% which is much lower than you put on the git(71%).I check out all the steps and can't find any change that I possibly make that might influence the perfermance. Would you tell me how to achieve the ~71% mIOU on VOC . Maybe I miss out some key training strategies.

COCO training codes

Hi~ Could you provide COCO training codes, to help us reproduce the results on COCO?

Thanks!

Deeplab-v2 pretrain model

I want to know which pretrain model you used when you train your deeplab-v2, imagenet or ms-coco?
Thanks

COCO checkpoint

When I use ‘coco_cls.pth’ and ‘resnet38_cls’ to generate CAM, the eval result is very low(only 1.3%). But when I use ‘coco_eps.pth’ and ‘resnet38_eps’, the eval result is correct. Do you have any suggestions for this problem？

Performance of VGG16 based Deeplab-v1

Hello, @halbielee ,
I use the pseudo labels to train the VGG16 based Deeplab-v1 used in OAA. However, I only get 65.1% mIOU. Could you please tell me where is the difference between my model and yours?