Point Transformer

This repository contains my code and report for the NPM3D Project (MVA 2025/2026) by, building upon the unofficial Point Transformer implementation (the code for the shape classification part of the model was never added).

• My forked implementation: https://github.com/TopAgrume/point-transformer
• Base unofficial implementation: https://github.com/POSTECH-CVLab/point-transformer

Project summary

This project proposes a study of the Point Transformer, a transformer-based architecture for 3D point cloud understanding.

• My primary goal is to reproduce the results of the original paper on the ModelNet40 shape classification task. This implementation achieves an overall accuracy of 92.7%, which is close to the 93.7% reported by the authors.
• Beyond reproduction, the project includes a series of ablation studies and architectural modifications to better understand the behavior of the model.

Experimental results

Shape classification on ModelNet40:

Model	Overall accuracy (OA)	Mean class accuracy (mAcc)
Current implementation	92.67%	90.25%
Original paper	93.7%	90.6%

My codebase additions and modifications

• Created model/pointtransformer/pointtransformer_cls.py (massively adapted from the existing segmentation model pointtransformer_cls.py).
• Created util/modelnet40.py to handle loading data specifically for the ModelNet40 dataset.
• Created util/profiler.py in order to measure the computational latency of each component of my implementation during the evaluation.
• Modified tool/train.py and tool/test.py to support classification tasks and configuration file selections.

Dependencies

• Ubuntu: 18.04 or higher
• PyTorch: 1.9.0
• CUDA: 11.1
• Hardware: GPUs required to reproduce Point Transformer
• To create conda environment, command as follows:

  bash env_setup.sh pt
  

Dataset preparation

Download the pre-processed modelnet40_normal_resampled dataset from Kaggle: ModelNet Normal Resampled

Unzip the downloaded file and save it in dataset/modelnet40_normal_resampled.

Note: Loading the dataset takes a long time during the first session because it caches the arrays directly into shared memory for fast access.

Shape classification on ModelNet40:

• Train a model (6 mn of caching): sh tool/train.sh modelnet40 pointtransformer_cls
• Modify hyperparameters: Edit the configuration file located at config/modelnet40/modelnet40_pointtransformer_cls.yaml.
• Test the best model (2 mn of caching): sh tool/test.sh modelnet40 pointtransformer_cls

Semantic segmentation on S3DIS:
Take a look at the base unofficial implementation repository.

Checkpoints and training logs

All experiments, model checkpoints, and training logs can be viewed and downloaded from the following link: Download Checkpoints & Logs (Google Drive)

Once downloaded and unzipped, you can easily visualize the training curves and metrics using TensorBoard. Navigate to the extracted folder in your terminal and run:

tensorboard --logdir=.

The following window will appear:

Reproducing figures

To reproduce the plots and graphs I used in the project report, run the following scripts:

• figures/treemap.py
• figures/pareto.py
• figures/graphs_generation_for_report.py

References

@inproceedings{zhao2021point,
  title={Point transformer},
  author={Zhao, Hengshuang and Jiang, Li and Jia, Jiaya and Torr, Philip HS and Koltun, Vladlen},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={16259--16268},
  year={2021}
}