PeRFception ScanNet Segmentation ground truth

First off, thank you for your great contribution.

Secondly, I have been exploring the PeRFception-ScanNet dataset, and I am trying to find the ground truth labels, but I am failing to do so. If they are included in the folders, is there a location where we can find them? If not, are there plans for you to release the ground truth files?

Additionally, since there is no documentation for the PeRFception-ScanNet, can you elaborate on the different numpy arrays (thick.npy and trans_info.npz)?

Thanks in advance for your help.

Oct 31 '22 08:10 kghandour

Sorry for being late. Our team members were busy for preparing the CVPR submission. @junha-l Could you answer this question?

Nov 16 '22 07:11 jeongyw12382

Hi @kghandour, sorry for the late response.

As you mentioned, the ground truth semantic labels are currently missing in the PeRFception-ScanNet repository. We'll upload it as soon as possible and let you know when it is done.

For the last question, trans_info.npz contains some useful metadata (e.g. frame ids that are used in training, the transformation matrix that aligns voxel grid of Plenoxels and the original ScanNet point cloud), and thick.npy is for the point cloud data that we used to initialize the voxel grid as stated in our paper.

Nov 16 '22 15:11 junha-l

Thank you for the great work. As a follow up question related to this thread. What is the difference between the "init.npy" and "thick.npy" files?

Nov 19 '22 18:11 m13ammed