DeepLab Code
DQDH opened this issue · 18 comments
Hi, @jiwoon-ahn, I get the same mIoU of cams and AffinityNet, but the result of Deeplab segmentation network only get 55% mIoU. Did you change something in DeepLab-CRF-LargeFOV? Can you give me your deeplab code if it be convenient to you.
Hi, hardBird123,I don't know how to get mIou.Can you give me your mIou code if it be convenient to you.Thank you!!!
@zhangxiufei you can open this link https://bitbucket.org/deeplab/deeplab-public/src , the code in path matlab/my_script/EvalSegResults.m is for mIoU. You should change the Image path with you own path.
@hardBird123,
Sorry for the late reply. I don't think the network is the problem. The performance of DeepLab-CRF-LargeFOV heavily depends on CRF parameters. I did use the exact settings the author of DeepLab revealed. (http://liangchiehchen.com/projects/DeepLab-LargeFOV.html) However, I am not sure this works for all cases.
Plus, you need to refine your segmentation labels with dense CRF as our paper describes.
Hope this may help your project. I may not be able to respond for a few days since I'm still so close to the deadline.
@jiwoon-ahn .Thanks. I will have a try with the CRF parameters. And the last problem is the plus tip which you mean refine the rw label with dense CRF? Is the code for dense CRF in DeepLab or your given code? I'm not very clear in the paper? Wish you reply the last question.
@hardBird123 ,Very thank you!!!
@hardBird123 I have one question, when you get the same IOU of cams and AffinityNet, you just use the provided weights by author or you train your own weights? when I download the weights of authors and use them directly, I found the mIoU cannot reach the same value.
@zbf1991 @jiwoon-ahn find i get the same mIoU on val dataset with author, but can't reach the same mIoU on train set that both in provided weights by author and my own weights. Have you solved this problem?
@hardBird123 Not yet. But I would not try cause it is enough for my research
hello, I'm a student who running this code. And there is a running error. Can you give me some tips about this issue.
RuntimeError: size mismatch, m1: [1 x 20], m2: [1 x 20]
@hardBird123 could you share your training setttings for ResNet38 AffinityNet? I train the ResNet38_aff network by myself with lr=0.01 and other default params but only achieve 59.077% mIoU on val set. Thanks a lot!
@zbf1991 @jiwoon-ahn find i get the same mIoU on val dataset with author, but can't reach the same mIoU on train set that both in provided weights by author and my own weights. Have you solved this problem?
@hardBird123 You says achieve same mIoU on val dataset with author, by VGG16?
@YudeWang I get the same mIoU on CAM result and rw result as the author by vgg16,but can't get the same mIoU on the final segmentation result by vgg16.
@hardBird123 May I ask how many GPUs you used for CAM and AffinityNet training?
@hardBird123 Thanks, I will try.
@DQDH Can I ask which segmentation network you use? deeplab_large_fov or resnet38? I could not found the code of the resnet38-deeplab.
Hi, @jiwoon-ahn, I get the same mIoU of cams and AffinityNet, but the result of Deeplab segmentation network only get 55% mIoU. Did you change something in DeepLab-CRF-LargeFOV? Can you give me your deeplab code if it be convenient to you.
Did you follow the readme to retrain vgg16 and get the same mIoU as author in CAM and rw result? I retrained by myself, but I failed to get cam result as cam pred mask png file were all None. Which vgg16 weight did you use to initialize your training? I did not install caffe so I cannot initialize model with pretrained [vgg16_20M.caffemodel] weight, so I used another one vgg16_20M_custom.pth., could you share me your training details? Thanks