Thank your brilliant Idea! I've trained a deeplabv3+ by your official pseudo labels and your network. However, I evaluate the mIOU of val_cls.txt that only achieve about 67.1 without dCRF. All of the hyperparameter is arranged to such as yours. I wanna know why my mIOU is lower than yours. Thank your for your reply. Best wishes with you.

Hi, thanks for your great work! I was wondering if DRS could be adapted in RESNET50. Have you ever tried it on ResNet? I have tried DRS in some layers buy found limited improvement. Do you have any suggestion?

Greetings,

Thank you for your code and paper, they give us a novel way to process the relationship with most descriptive part and the other object part. When I read your code, I find some question, would you answer them for me? I find that the pretrained link in your deeplabv2 table turn into the download for a checkpoint file with name "final checkpoint.pth", is that truly the pretrained model for deeplabv2 or just the final checkpoint of your deeplabv2 training? Another reason let me think like this is that I find the pretrained path in your config file is "data/models/coco/deeplabv1_resnet101/caffemodel/deeplabv1_resnet101-coco.pth". Is it the same checkpoint file with the "checkpoint.pth"?

Hi, Thank you for your code and paper. I've already tried your code, but I cannot achieve the same performance as the paper. Would you please help me figure out where the problem is? I retrained DeepLabV2 with your shared refined pseudo segmentation labels. But the performance on val only achieves 64.8 and 65.7(after crf). All hyper-parameters follow your provided code. Could you please point out the possible mistake? Thank you!

Amazing Work! When I used the provided code and followed the method in your paper, I only got val 69.45 with learnable DRS. How can I get the same performance as reported in your paper？

Thank you for a great job! "In contrast, DeepLab-V3 does not require the COCO-pretrained weight due to the recent large memory GPUs and Synchronized BatchNorm" confused me. Is deeplab-V3+ is trained from scratch without any pre-trained weight? Then I have a question about the link. The link of Deeplab-V3+ with ResNet-101 is pre-trained weight? or it's the model to test and get the results, 71%.

Model | mIoU | mIoU + CRF | pretrained -- | -- | -- | -- DeepLab-V2 with ResNet-101 | 69.4% | 70.4% | [link] DeepLab-V3+ with ResNet-101 | 70.4% | 71.0% | [link]

Hi qjadud,

I've got all the other data in the data structure shown in README file but I can not find the dir name 'refined_pseudo_segmentation_labels/'. Would you mind help me with it?

As the pretrained network you provide for deeplab v2, https://drive.google.com/file/d/1Wj8Maj9KGQgwtDfvIp8FChsdAIgDvliT/view, the name of it is deeplabv1_resnet101-coco.pth, is it correct?

Hey, thanks for your work. When I see your result in Pseudo Segmentation Labels, I found there are some white regions, which I guess is generated because of the saliency. Since these regions belong to no categories, what did you do for it when you train a fully supervised segmentation network, like Deeplab2 or 3? Does these regions regarded as background or categories nearby?

Hey, thank you so much for your interesting work. When I run with train_cls.py, there is a error occurs in validate() function. In the last of the function, mIoU is made between pred_map and gt_map. The pred map has the size of HW1, and each pixel value in the map shows the prediction labels. However, the gt_map is HW3, which is the map read directly from the 'SegmentationAug' dir, it is a RGB image. Could you please tell the convert method from the RGB into the Seg label? Thank you so much for your time.

Patching CVE-2007-4559

Hi, we are security researchers from the Advanced Research Center at Trellix. We have began a campaign to patch a widespread bug named CVE-2007-4559. CVE-2007-4559 is a 15 year old bug in the Python tarfile package. By using extract() or extractall() on a tarfile object without sanitizing input, a maliciously crafted .tar file could perform a directory path traversal attack. We found at least one unsantized extractall() in your codebase and are providing a patch for you via pull request. The patch essentially checks to see if all tarfile members will be extracted safely and throws an exception otherwise. We encourage you to use this patch or your own solution to secure against CVE-2007-4559. Further technical information about the vulnerability can be found in this blog.

If you have further questions you may contact us through this projects lead researcher Kasimir Schulz.

Hey, maybe this is question is not strongly related to this proj, but I will appreciate a lot if you could provide any advice for me. The question is, do you have any idea about the miou performance for a classification network based on vgg (13 layers as backbone and with other 3 extra_conv). When I try to reproduce your code ( The first step, train_cls.py), I remove the saliency and computing the background by a given threshold. I find that the miou drops. This is acceptable since saliency provides much useful information. But when I try to remove other module in the vgg.py, to make it as a classification network, the best CAM I could get is among 42. You could find that in PSA, same backbone network(vgg16) could reach 48+. At first, I wonder it is caused by the different parameters in the last three conv layers, but when I keep them as the same, I still can't reach 48. Do you have any idea about the limitation? If so, I will appreciate a lot for your sincere help!

Thank you for providing such a convenient code, but I have some questions about it. On the line 75 of eval.py, you directly use the image after crop for performance verification. Is this accurate? Another question is why not apply ms+flip to inference?

Greetings！

I read your paper again and again recently, and I really learned a lot from your work. For the retrain part, I met some problems, would you like to help to answer?

At first, I noticed that your results in Table.4 were devided into 3 parts, the subtitle of first part is Segmentation Network : DeepLab-Large-FOV (VGG-16), the 2st part is Segmentation Network : DeepLab-Large-FOV (ResNet-101), the 3rd part is Segmentation Network : DeepLab-ASPP (ResNet-101).

My first question is that your deeplab retrained script(/DRS/DeepLab-V2-PyTorch/train.sh) indicate which part in Table.4 we mentioned before? What about other two parts' code? I could not find a VGG-Backboned model in DeepLab-V2-PyTorch folder.