search

ThibaultGROUEIX / AtlasNet

This repository contains the source codes for the paper "AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation ". The network is able to synthesize a mesh (point cloud + connectivity) from a low-resolution point cloud, or from an image.
http://imagine.enpc.fr/~groueixt/atlasnet/
MIT License
677 stars 119 forks source link

Dataset Split used for AtlasNet #47

Closed KunalMGupta closed 4 years ago

KunalMGupta commented 4 years ago

Hi,

I am wondering which dataset split was used for training the AtlasNet model presented in the original paper. As per my understanding, the paper claims to be using the split earlier used by 3DR2N2 paper. However, I am not able to find the split file anywhere in their repository. I could only find the code here which seems to be dealing with any dataset split. However, I do not believe that this is the same as that of 3DR2N2. Additionally, the authors of Occupancy Network mention in their paper that they could only evaluate pretrained AtlasNet model for a part of the 3DR2N2 dataset. I have checked the overlap of dataset split used by MeshRCNN, Occupancy Network (both of which use 3DR2N2's data split and are consistent with each other) and the one generated from here and I believe that they do not have any significant overlap. I am therefore not sure which dataset split is used by AtlasNet. Is it possible to share a JSON file of the dataset split used?

Thanks.

ThibaultGROUEIX commented 4 years ago

Hi @KunalMGupta, The original AtlasNet implem (check the tags) used ShapeNetv2 while 3Dr2n2 used shapenetv1, which meant the splits where different. I reverted Atlasnet code to Shapenetv1 at the end of 2019 so the splits should now be identical with 3D-r2n2. Best regards, Thibault

KunalMGupta commented 4 years ago

I see. Thanks a lot for clearing the confusion.