Skip to content

NPM3D: 3D Modeling & Point Clouds (MVA 2021) - Project on "Point Transformer" by H. Zhao et al.

Notifications You must be signed in to change notification settings

pierrefdz/point-transformer

Repository files navigation

Paper | Colab

Unofficial Implementation of Point Transformer for classification on ModelNet40

Point Transformer is a paper by Hengshuang Zhao et al. (2020). The project was done in the scope of the course NPM3D of Master MVA 2021. It consisted in understanding the paper and testing one part of the paper, here the classification on ModelNet40.

I also provide a notebook where all the pipeline is already implemented.

Dataset

The data can be downloaded here. It is a pre-processed version of the ModelNet40 CAD models that was created for PointNet++ (see [PointNet++]{http://stanford.edu/~rqi/pointnet2/}). It must then be unzipped and saved in data/modelnet40_normal_resampled.

Loading the dataset takes a long time the first time, since it pre-computes the samplings and saves it in the files modelnet40_test_1024pts_fps and modelnet40_train_1024pts_fps. Once pre-processed it allows the computations to be much faster for the rest of the pipeline.

Run

To run a training procedure and test a model you should run train.py (and change parameters if you want to). In the current configuration, you should obtain around 91.5% accuracy on the test split.

Each .py file can be unit-tested (by directly running them)

About

NPM3D: 3D Modeling & Point Clouds (MVA 2021) - Project on "Point Transformer" by H. Zhao et al.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages