result

[anggirl]

Hello, I would like to ask you a question. After a complete training according to the instructions in the document, the mIoU index of the result is only 0.65, which is quite different from the result mentioned in your paper. I wonder what is the reason? And I would like to ask again whether the ground-truth of the following test data set in segmentation training is pure black?

[Elliot-QxZhang]

Hi, I successfully reproduce the paper's result with mIOU of 78.91%. And the reason why groundtruth is totally black is that it is not correctly visualized in RGB format. If you want to see the foreground, you need to multiply the ground-truth mask by 255.

[ZlqKevin]

@Eric-jinkens Hello, I found in the process of reproduction that when training the segmentation model, the mIOU always hovered around 18%, and when testing, the mIOU was only 18%. Where do you think I might be ashamed? Thank you very much for your answer

[Elliot-QxZhang]

Hi! Could you please provide a little more information? I haven't encountered such a low mIOU, have you correctly moved the pseudo-label generated from the classification stage to the input folder?

[Elliot-QxZhang]

By using ln -s OEEM_resources/glas_seg segmentation/glas, you could move the official pseudo-label to the input folder. Or if you wanna train the classification model yourself, you need to use the following command to prepare the segmentation inputs:

python classification/prepare_seg_inputs.py -d 0 -ckpt res38d_best
python segmentation/tools/crop_img_and_gt.py segmentation/glas/images classification/res38d_best_train_pseudo_mask segmentation/glas

[ZlqKevin]

Thank you for taking the time to reply. I used the officially trained classification model to generate pseudo label and then trained the segmentation model. The whole process was executed according to the instructions in readme. The picture shows some data of the segmentation model training process. I wonder if it can provide you with further information. Thank you very much.

[Elliot-QxZhang]

Oh! I know why. That is a bug, I have encountered that before. The pseudo label is reversed, so the performance is so low. You need to firstly reverse the foreground and background so that you could get the correct results. You can check this by visualizing the pseudo label by multiplying it with 255.

[ZlqKevin]

@Eric-jinkens Oh, I see. Thank you very much. Without your help, I might still be confused, haha. I will try to correct this error. Thank you again for your help.