zhanghang1989
Supplementary material of CVPR2019
- Could you deliver your supplementary material to my email ([email protected])?
- In the subsection Ablation study of CFNet, you mention that global pooling yields 0.9% improvement. However, in Table 2 , I don't kown which two results can be seen to obtain 0.9% improvement.
- How do I get the legend of ADE20K like Fig. 2 ?
Hi, I also have a question about CFNet. In the paper, the description of Table 4 about the testing result of Pascal Context mentions that "mIoU on 60 classes w/ background". I wonder if the number of output channels of CFNet is 60 when both training and evaluating on pcontext. In other words, edit NUM_CLASS to 60(https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/encoding/datasets/pcontext.py#L19), like argued in #78 . Thanks!
@txfs1926 May be another answer can help you. See #179 . Question1 For obtaining mIoU with backgroud , computing mIoU with backgroud, i,e,. 60 classes for PContext, fisrt need to get mIoU without backgroud (mIoU_59) as you did in repo, then mIoU_60 is directly equal to mIoU_59 * 59 / 60 and will be slightly weaker than mIoU_59. Do I understand correctly?
Answer1 yes, we just simply assume the pixels from background category are all misclassied.
@zhanghang1989 I have noticed that GCU (NeurIPS2018, paper link: https://papers.nips.cc/paper/8135-beyond-grids-learning-graph-representations-for-visual-recognition.pdf) creates visualization of the assignment matrix in their method. It is meaningful to illustrate the proposed method. Do you consider visualizations of the learned assignment weight matrix in EncNet to further straightforwardly detail your method?
@zhanghang1989 I have noticed that GCU (NeurIPS2018, paper link: https://papers.nips.cc/paper/8135-beyond-grids-learning-graph-representations-for-visual-recognition.pdf) creates visualization of the assignment matrix in their method. It is meaningful to illustrate the proposed method. Do you consider visualizations of the learned assignment weight matrix in EncNet to further straightforwardly detail your method?
Thanks @qiulesun for the suggestion. Actually I am not active in research work now.
@zhanghang1989 GC = tmp.sum(0).sum(0) in https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/encoding/lib/cpu/encoding_cpu.cpp#L44 should be GC = -tmp.sum(0).sum(0) ?
@zhanghang1989 GC = tmp.sum(0).sum(0) in https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/encoding/lib/cpu/encoding_cpu.cpp#L44 should be GC = -tmp.sum(0).sum(0) ?
The weights can learn the negative value automatically if necessary.
@zhanghang1989 When args.test_val=Ture （https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/experiments/segmentation/test.py#L103), dst (https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/experiments/segmentation/test.py#L103) is the image name rather than the mask. Why is the image name obtained at this time?