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.
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.
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)