search

wangbo-zhao / 2021CVPR-WSVSOD

The code for CVPR2021 Weakly Supervised Video Salient Object Detection
43 stars 4 forks source link

reproduce your work #4

Closed muchengxue0911 closed 2 years ago

muchengxue0911 commented 2 years ago

Thanks for your great work.However, I got trouble reproducing your work.

  1. the images are jpg format while the gt, mask, etc. are png format. But the load of the dataset didn't transform it; thus I got a file not found error. After I fixed it myself, I found the pretrained process use the optical flow, which I think it's not reasonable.So I wonder if you uploaded the wrong version of your code. 2 The fine-tune process needs 8 file directory, I wonder how can I get fixation_maps, fixation_maps_smoothed, flow and grey? Could you please provide the files, since I didn't find them in dataset.rar in the link you provided of BaiduNetDisk so that we can reproduce your wonderful work more easily?

Thanks.

wangbo-zhao commented 2 years ago

Thanks for your interest.

  1. In fact, we adopt optical flow maps generated from static images. This process is described in " Implementation details " in our paper.
  2. fixation_maps and fixation_maps_smoothed are directly downloaded from the datasets that we talked about in our paper. You can use any optical flow estimation method to generate the flow. The gery can be generated from original RGB frames.

muchengxue0911 @.***> 于2022年7月21日周四 16:28写道：

Thanks for your great work.However, I got trouble reproducing your work.

  1. the images are jpg format while the gt, mask, etc. are png format. But the load of the dataset didn't transform it; thus I got a file not found error. After I fixed it myself, I found the pretrained process use the optical flow, which I think it's not reasonable.So I wonder if you uploaded the wrong version of your code. 2 The fine-tune process needs 8 file directory, I wonder how can I get fixation_maps, fixation_maps_smoothed, flow and grey? Could you please provide the files, since I didn't find them in dataset.rar in the link you provided of BaiduNetDisk so that we can reproduce your wonderful work more easily?

Thanks.

— Reply to this email directly, view it on GitHub https://github.com/wangbo-zhao/2021CVPR-WSVSOD/issues/4, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANR3QJRJJ444UQNXIHRHQZTVVEC2VANCNFSM54GXKLKA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

muchengxue0911 commented 2 years ago

Thanks for your patient answer, I'll try it.

muchengxue0911 commented 2 years ago

After generating the optial flow, another error appeared: too many values to unpack (expected 5) in line 272 in your model.py file, which is because you want to unpack 5 tensors while the rgb_feature is a 4 dimentions tensor. Thus, once again, can you recheck your code and uploaded a runnable version of your great work? Thanks!

muchengxue0911 commented 2 years ago

And what's more, I want to finetune the model directly using your pretrained model, however, I have difficulties in the fiation_maps and fixation_maps_smoothed files. I noticed that you mentioned reference 37 provided the fixation map of DAVIS, but the link in his github is invalid. Moreover, where can I find the fixation_map and smoothed files for DAVSOD? Can you kindly upload them to your BaiduNetDisk? Thanks a lot.

wangbo-zhao commented 2 years ago

Sorry for confusing you. I remember that I do not use fiation_maps and fixation_maps_smoothed files during training and finetune in fact. Please try commenting codes related to fixation_map and fixation_map_smoothed in the dataset. I will fix this bug later. image

muchengxue0911 commented 2 years ago

OK，Thanks a lot. I will have a try ~

wangbo-zhao commented 2 years ago

I would like to know whether you solve the problem "too many values to unpack (expected 5) in line 272 in your model.py file, which is because you want to unpack 5 tensors while the rgb_feature is a 4 dimentions tensor". Since I do not have the environment to run this code now, please tell me the bug if you find it.

muchengxue0911 commented 2 years ago

No， I'm just trying the finetune period now. I'll let you know when it's done.

muchengxue0911 commented 2 years ago

Sorry to bother you again. Could you please upload your optical flow images? I use a network to generate the optical flow, but the results are lower than your paper.

wangbo-zhao commented 2 years ago

Sorry. I do not have these optical flow images. Because I have graduated from NWPU. What network do you use to generate the optical flow?

muchengxue0911 @.***> 于2022年10月21日周五 16:35写道：

Sorry to bother you again. Could you please upload your optical flow images? I use a network to generate the optical flow, but the results are lower than your paper.

— Reply to this email directly, view it on GitHub https://github.com/wangbo-zhao/2021CVPR-WSVSOD/issues/4#issuecomment-1286642915, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANR3QJSZON6TYQY2XXVGV2DWEJIUJANCNFSM54GXKLKA . You are receiving this because you modified the open/close state.Message ID: @.***>

muchengxue0911 commented 2 years ago

I use a pytorch version of pwcnet.

wangbo-zhao commented 2 years ago

Is the difference between your results and our results large?

muchengxue0911 @.***> 于2022年10月21日周五 18:39写道：

I use a pytorch version of pwcnet.

— Reply to this email directly, view it on GitHub https://github.com/wangbo-zhao/2021CVPR-WSVSOD/issues/4#issuecomment-1286777110, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANR3QJT72WO2UHGT6U5FRR3WEJXGHANCNFSM54GXKLKA . You are receiving this because you modified the open/close state.Message ID: @.***>

muchengxue0911 commented 2 years ago

Yes，S and F is about 3% lower in DAVIS and ViSal, 1% at other datasets