How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?

Tuan Anh Tran¹, Duy Minh Ho Nguyen^2,3, Hoai-Chau Tran⁴, Michael Barz¹, Khoa D. Doan⁴, Roger Wattenhofer⁵, Vien Anh Ngo⁶, Mathias Niepert^2,3, Daniel Sonntag^1,7, Paul Swoboda^8
1German Research Centre for Artificial Intelligence (DFKI), ²Max Planck Research School for Intelligent Systems (IMPRS-IS), ³University of Stuttgart
⁴College of Engineering and Computer Science, VinUniversity, ⁵ETH Zurich, ⁶VinRobotics, Hanoi, Vietnam
⁷University of Oldenburg, ⁸Heinrich Heine University Düsseldorf

NeurIPS 2025

Code | Paper | Project Page

GitMerge3D enables merging of up to 80-95% of tokens, substantially reducing computational and memory costs while preserving model performance.

Feature PCA visualizations across different merging rates (0.15 to 0.95) demonstrating that learned representations remain distinctive or unchanged despite aggressive token reduction

A. Documentation

• Token Merging Evaluation Guide: How to evaluate models and measure GFLOPs
• Token Merging Training Guide: Fine-tuning with token merging
• Training Guide: Baseline training without token merging
• Evaluation Scripts: Ready-to-use evaluation tools

B. Project Structure

GitMerge3D/
├── Sonata/                                    # Main codebase
│   ├── configs/sonata/                       # Token merging configurations
│   ├── pointcept/
│   │   └── models/point_transformer_v3/
│   │       └── token_merging_algos.py        # Core merging algorithms
│   ├── tools/                                # Training and testing scripts
│   ├── token_merging_evaluation/             # Evaluation scripts
│   ├── TOKEN_MERGING_EVALUATION_GUIDE.md
│   ├── FINETUNING_TOKEN_MERGING_GUIDE.md
│   └── README.md                             # Base installation guide
└── README.md                                  # This file

C. Roadmap

Current Release

• Code for Sonata (Pointcept v1.6.0) - Token merging implementation with Sonata backbone
• Training scripts and configurations
• Evaluation tools and documentation

Coming Soon

• Model Checkpoints - Pre-trained weights for all retention ratios (r=0.7, 0.8, 0.9, 0.95)
• SpatialLM Integration - Code and checkpoints for SpatialLM with token merging
• PTv3 (Pointcept v1.5.1) - Code and checkpoints for Point Transformer V3 baseline

Stay tuned for updates!

D. Installation

Requirements

• Ubuntu: 18.04 and above
• CUDA: 11.3 and above
• PyTorch: 1.10.0 and above

Setup

1. Clone the repository

   git clone https://github.com/anhtuanhsgs/GitMerge3D.git
   cd GitMerge3D/Sonata
   

2. Create conda environment

   Follow the installation instructions in Sonata/README.md:

   # Option 1: Using environment.yml
   conda env create -f environment.yml --verbose
   conda activate pointcept-torch2.5.0-cu12.4
   
   # Option 2: Manual installation
   # See Sonata/README.md for detailed manual setup
   

3. Install additional dependencies

   # PTv3 dependencies
   cd libs/pointops
   python setup.py install
   cd ../..
   

4. Prepare datasets

   Follow the data preparation instructions in Sonata/README.md for:

   • ScanNet v2
   • S3DIS
   • ScanNet200

E. Quick Start

Training with Token Merging

Fine-tune a pre-trained model with token merging enabled:

cd Sonata

# ScanNet with r=0.9 (90% token retention)
sh scripts/train.sh -g 4 -d scannet \
    -c gitmerge3d-patch-0c-scannet-ft-finetune-100epochs-r0.9-s10 \
    -n gitmerge3d-patch-r0.9-scannet-ft \
    -w exp/sonata/your-pretrained-model/model/model_best.pth

Evaluation

Evaluate a trained model:

# Using evaluation script
cd Sonata
bash token_merging_evaluation/eval_scannet.sh

Or use the Python script directly:

python tools/test.py \
    --config-file configs/sonata/gitmerge3d-wpatch-0c-scannet-ft-finetune-100epochs-r0.9-s10.py \
    --options weight=exp/sonata/your-checkpoint/model/model_best.pth

Measuring GFLOPs

Measure computational efficiency across multiple retention ratios:

cd Sonata/token_merging_evaluation
python run_cal_flops_sweep.py \
    --config ../configs/sonata/gitmerge3d-wpatch-0c-scannet-ft-finetune-100epochs-r0.9-s10.py \
    --merge-rates 0.7 0.8 0.9 0.95 \
    --max-scene-count 10 \
    --gpu 0

F. Performance

Results on ScanNet semantic segmentation:

Up to 21% reduction in computational cost with minimal accuracy loss

Citation

If you find this work useful in your research, please cite:

@inproceedings{gitmerge3d2025,
    title={How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?},
    author={Tuan Anh Tran and Duy Minh Ho Nguyen and Hoai-Chau Tran and Michael Barz and Khoa D. Doan and Roger Wattenhofer and Vien Anh Ngo and Mathias Niepert and Daniel Sonntag and Paul Swoboda},
    booktitle={Advances in Neural Information Processing Systems},
    year={2025}
}

Also cite the base Pointcept and Sonata work:

@misc{pointcept2023,
    title={Pointcept: A Codebase for Point Cloud Perception Research},
    author={Pointcept Contributors},
    howpublished = {\url{https://github.com/Pointcept/Pointcept}},
    year={2023}
}

Acknowledgements

This project is built upon:

• Pointcept - Point cloud perception codebase
• Point Transformer V3 - Efficient point cloud backbone
• Sonata - Self-supervised learning for point clouds

License

This project is released under the MIT License. See LICENSE for details.