FSCIL 3D

Few-shot Class-incremental Learning for 3D Point Cloud Objects, ECCV 2022
Townim Chowdhury, Ali Cheraghian, Sameera Ramasinghe, Sahar Ahmadi, Morteza Saberi, Shafin Rahman

This paper addresses the problem of few-shot class incremental learning for the 3D domain alongside the domain gap from synthetic to real objects.

Figure: Overall Architecture

Dependencies

FSCIL 3D is implemented in PyTorch and tested with Ubuntu 20.04.2 LTS, please install PyTorch first in the official instruction. You can also install the anaconda environment using the provided environment.yml file.

Configurations

Here, we only include the configuration for the large cross-dataset experiment: ShapeNet -> CO3D. The configuration files are in configs folder.

You can create similar configuration files for other experimental setups.

Dataset

Please check the readme file in here.

K-Means

You can build your own centroids with this code.

Command

python train_pointnet_incremental_with_knn_with_w2v.py

For this ShapeNet -> CO3D experiment, we provide required centroids and model here.

Citation

@inproceedings{FSCIL3D,
  title={Few-shot Class-incremental Learning for 3D Point Cloud Objects},
  author={Chowdhury, Townim and Cheraghian, Ali and Ramasinghe, Sameera and Ahmadi, Sahar and Saberi, Morteza and Rahman, Shafin},
  booktitle={ECCV},
  year={2022}
}