ShapeNet-55 - Githubissues

Julie-tang00 / Point-BERT

[CVPR 2022] Pre-Training 3D Point Cloud Transformers with Masked Point Modeling

MIT License

541 stars 65 forks source link

ShapeNet-55 #32

Closed Phoebe-ovo closed 2 years ago

Phoebe-ovo commented 2 years ago

Why not use the original ShapeNetCore dataset directly? And how did you generate ShapeNet-55 from ShapeNetCore?

yuxumin commented 2 years ago

Hi, thanks for your interest in our paper. The complete point cloud in ShapeNet-55 is the same as in ShapeNetCore. So, in fact, we use the original ShapeNetCore for pre-training. (ShapeNet-55 is the dataset used for point cloud completion. But the data for both datasets is the same. We claim that we used ShapeNetCore in our paper. We use it in our repo only because our repo is based on PoinTr, which proposes ShapeNet-55)

Phoebe-ovo commented 2 years ago

Thank you for your quick reply~ What's the code for generating ShapeNet-55?

yuxumin commented 2 years ago

In fact, they are the same in terms of the data.

Phoebe-ovo commented 2 years ago

Given the original 3D model in the shapenet dataset(.obj files), I want to know how did you get the groundtruth point cloud?

yuxumin commented 2 years ago

We use the data from ShapeNetCore in the training of dVAE and MPM tasks for Transformers.

In dVAE: The model (dVAE) is required to reconstruct the input point clouds. In that case, the groundtruth point cloud is exactly itself.
In MPM tasks: We randomly mask some local patches in point cloud from ShapeNetCore and require the Transformers to predict the missing local patches by discrete tokens. The groundtruth discrete tokens are generated by infering the missing local patches with the Tokenizer in the pretrained dVAE.

yuxumin commented 2 years ago

If I misunderstood the question you want to ask, plz feel free to continue the question.

Phoebe-ovo commented 2 years ago

ShapeNetCore seems don't provide point cloud directly, it provides the mesh and normal vector. I want to ask how to generate the point cloud from the mesh and normal vector?

yuxumin commented 2 years ago

Sorry for misunderstanding your question. The point clouds are obtained by random-sampling or other sampling methods. An example code is provided in AtlasNet/OCNN

Phoebe-ovo commented 2 years ago

Thank you for your patient reply. I will try it by myself.

yuxumin commented 2 years ago

Close it since no response. Feel free to re-open it if problems still exist

yuxumin commented 2 years ago

@Phoebe-ovo , Hi, I find the code for obj2pc code and update it here (https://github.com/yuxumin/PoinTr/blob/master/datasets/shapenet2pc.py). Hope it can help.