About the results on PASCAL Context

[junfu1115]

About the results on PASCAL Context

[junfu1115]

Q#1, we report the results on 60 classes, while EncNet achieves 51.7 mIoU on 59 classes w/o background.
Q#2, we didn't try the latter one.

[qiulesun]

Question 1:
For Q#1, the paper achieves 52.6 in mIoU without using DA, Multi-Grid or MS ?
Question 2:
For the two attention modules, do you try to empoly sigmoid op to obtain weight rather than softmax ?

[junfu1115]

Q#1, Yes
Q#2, We don't try sigmoid, in fact, we think softmax is more appropriate since it involves all pixels

[qiulesun]

Question 1:
args.lr = lrs[args.dataset.lower()] / 8 * args.batch_size in option.py means that the lr is relate to batch_size you give. Is that the lr not fixed depending on the batch_size (GPU memory)?

[qiulesun]

Question 1:
I am sorry for bothering you again. What are the base size and crop size of PContext dataset for training and multi-scale test?

[junfu1115]

On the PContext dataset, the base size is set to 608 and crop size is set to 576. Multi-scale test is adopted, you can refer to the experimental settings of PyTorch-Encoding

[HolmesShuan]

@qiulesun @junfu1115 Hi, I have mailed the authors of CCL(i.e., [6] in DANet). The result of [6] in Table 6 is actually 59 classes without background and with single scale testing. It may be unfair to compare with CCL directly :(

[qiulesun]

@junfu1115 trivial questions and suggestions #179 and Performance on Pascal Context #78 give different strategies to compute mIoU on 60 classes for PASCAL Context. They confused me deeply. How does DANet calculate the mIoU on 60 classes ?

[qiulesun]

@junfu1115
Q1: In DANet, could you provide the settings of base size and crop size for each dataset ?
Q2: CAM_Module computes the energy_new before softmax while PAM_Module not, why ?
Q3: Your repo mainly follows Pytorch-Encoding. I note that there are two different ways to compute mIoU on 60 classes for PASCAL Context (see issues from trivial questions and suggestions #179 and Performance on Pascal Context #78). How does DANet calculate the mIoU on 60 classes ?