This repository contains the official model and benchmarking system implementation of our paper "Pointy – A Lightweight Transformer for Point Cloud Foundation Models".
├── configs/ # Configuration files for experiments
├── scripts/ # Scripts for running experiments
├── src/
│ ├── data/ # Data processing
│ ├── datasets/ # Dataset implementations
│ ├── metrics/ # Evaluation metrics
│ ├── models/
│ │ ├── baselines/ # Benchmark models
│ │ └── pointy*.py # Pointy model
│ ├── trainers/ # Training and evaluation pipelines
│ └── *.py # Utility functions
└── requirements.txt # Dependencies-
Clone the repository
-
Install the dependencies:
pip install -r requirements.txt- Create a
.envfile in the root directory with the following paths:
DATA_PATH=/path/to/datasets/
CHECKPOINTS_PATH=/path/to/checkpoints/
RESULTS_PATH=/path/to/results/Pre-trained model weights and configurations will be released soon. Once available, you'll be able to download and place them in your CHECKPOINTS_PATH directory for zero-shot experiments.
To train the model from scratch:
python -m scripts.train_classifier --config configs/classification_pointy.yamlUpdate the checkpoint path in configs/zero_shot_evaluation_*.yaml to point to your downloaded weights.
Run evaluation:
python -m scripts.evaluate_zero_shot \
--config configs/zero_shot_evaluation_classification_modelnet40.yaml