Enhancing Few-Shot 3D Point Cloud Classification with Soft Interaction Module and Self-Attention Residual Feedforward

Environment Requirements

Before you get started, ensure you have the following dependencies installed:

• Python3, PyTorch, json, h5py

Getting Started

Dataset Download and Split for ModelNet40

1. Navigate to the ../dataset/ directory. (Please note that the 'dataset' folder should be at the same level as the project folder. If it doesn't exist, create the directory.)
2. Download the dataset by clicking on this ModelNet40 link, and unzip it into the ../dataset/modelnet40_normal_resampled folder.
3. Dataset Split: You can split the dataset manually or use our predefined split by copying the './data/base.json' and './data/novel.json' files to ../dataset/modelnet40.

Data Preprocessing (ModelNet40, ShapeNetCore, ScanObjectNN, etc.)

Set all the file paths according to config.py and refer to the DataProcess folder to organize and build JSON files for the datasets.

Train Your Model

To start training your model, run the following command:

python ./train.py --model [BACKBONENAME] --method [METHODNAME] [--OPTIONARG]

For example, you can run python ./train.py --model pointnet --method protonet --weightdecay 0.01. If you wish to implement different methods with varying hyperparameters, please refer to 'io_utils.py'.

Save Features

After training, saving features first is advisable, as this can significantly speed up repeated experiments during testing. For instance, you can run the following command:

python ./save_features.py --model pointnet --method protonet

If you would like more details, you can consult 'io_utils.py'. You can locate your stored feature files in the './features' directory.

Test Your Model

To test your trained model, run a command like this:

python ./test.py --model pointnet --method protonet

For further details, refer to 'io_utils.py'.

Reminder: You can run the test command directly without executing 'train.py' and 'save_features.py' since we have provided a pre-trained model for PointNet+ProtoNet. You can check your experiment results in ./record/results.txt.

Download Pre-trained Weights/Checkpoints

Download Features and Checkpoints

References

This code is adapted from: https://github.com/PeiZhou26/Few-Shot-3D-Point-Cloud-Classification.

We thank the author for providing code and guidance.