This repository contains an implementation of the PointNet model for pointcloud classification using pure Numpy for the forward pass. The goal is to provide an easy-to-use version of PointNet without requiring heavy deep learning libraries like Keras, TensorFlow, or PyTorch.
Models are first trained using Keras and then converted to Numpy. Pretrained models (both in Keras and Numpy) are available in the models/saved directory. You can use these models for inference directly.
So far, models were only trained on the ModelNet40 dataset, but plans are to extend to other popular datasets.
The predict_numpy_model_example.py file provides a minimal example for prediction.
Example of prediction
The train_keras_model.py file may be use to train the model on a custom dataset (requires Keras and likely a GPU). Trained model is "converted" using the convert_keras_model_to_numpy.py file.