ICNet Problem

[erichhhhho]

In the low-res branch, the input of the low-res feature map(x_sub4) in your code is from the medium and downsample by 0.2.

But from the original paper, it should be downsampled from the raw input by 0.25, is it?

[lxtGH]

No, the original code is a litttle different from the figure in the paper. See this issue hszhao/ICNet#7.

[erichhhhho]

@lxtGH Yes. I also noticed that the original code use element-wise summation instead of the concatenation. Would you mind sharing your trained model performance?

[lxtGH]

I will report the results and trained model as soon as possible.

[lxtGH]

Please use the lastest code for reproducing ICnet results.

[erichhhhho]

@lxtGH Thank you.
I think the ICNet loss in your code is inaccurate.

When you calculate the loss, according to the paper: the ground truth should be downsampled to the prediction size in 1/4, 1/8, 1/16 branch, instead of upsampling the sub prediction. There is a difference. And I was wondering the mIoU in your trained model is reported in the test set or val set?

[lxtGH]

You can have a try by not upsampling. and I report the results on val set