Scannet validation set

[OPradelle]

Hi, thanks for sharing your work.

I tried to use the ScanNet pretrained model to obtain the validation set mIoU score but the network gives poor result. I generated the HHA image from the raw depth (following https://github.com/charlesCXK/Depth2HHA-python) and use the images at a resolution of 640x480. Did i miss something there ?

Thanks for your answer

[elnino9ykl]

Hi, thank you for your interest in our work!

We have some suggestions regarding your question:

1. Please make sure that the pre-trained model is correctly loaded. For example, you can try visualizing some segmentation maps.
2. Please check whether the weights of some layers are missing after loading.
3. Please consider visualizing some HHA maps and checking whether they are similar to the maps shown in our paper.
4. If these do not work, please consider training a model using our provided code.

Best,

[OPradelle]

Hi, thanks for your answer.

I used directly the code you provide and update only on the config file (ie. data path, image size) using the configuration provided in your paper for the ScanNet dataset. Can you provide the config file used for this task to check on my side if anything went wrong ?

Looking at the HHA images, it look like the same provided in your paper, but i want to know if you first inpaint the depth to fill the empty value or if you directly use the raw depth.

Thanks.