Low Accuracy on ADE20K val

[pjohh]

Hello,

I followed the instructions and ran eval_all.m on the ADE20K Validation Dataset (the one from the Challenge). Unfortunately the results are very bad: Mean IoU over 150 classes: 0.0007 Pixel-wise Accuracy: 2.15% (see also attached file) PSPNet_ADE20K_val_results.txt

Tested on 2 different machines. First one on CPU; second one on GPU (Titan X) with Cuda 8.0 and without cuDNN.

The segmentation images don't look that bad but are a lot worse than the ones presented in the paper...

Any hints? How can I check whether the pretrained caffemodel is loaded successfully or not?

Thanks in advance

Best regards, Johannes

[hszhao]

Hi, when training, the gt labels [0-150] are mapped into 0-149 and 255 for void labels are ignored.

[hellochick]

Hey, @pjohh Did you solve this problem yet? I encounter same problem as you: Mean IoU over 150 classes: 0.0007 Pixel-wise Accuracy: 2.15% Thanks.

[shahabty]

Hello, I have the same problem: Mean IoU over 150 classes: 0.0006 Pixel-wise Accuracy: 1.96%

[qilinli]

in eval_acc.m, line 38, comment '' imAnno = imAnno + 1; " and try

[ThienAnh]

@shahabty How to training with more than class. Currently, i only train susscess for 21 classes