Zero-shot Learning of 3D Point Cloud Objects

Created by Ali Cheraghian from Australian National University.

Introduction

This work is based on our arXiv tech report, which appeared in MVA 2019. We proposed a novel zero-shot learning framework for 3D point clouds.

Zero-shot learning, the task of learning to recognize new classes not seen during training, has received considerable attention in the case of 2D image classification. However despite the increasing ubiquity of 3D sensors, the corresponding 3D point cloud classification problem has not been meaningfully explored and introduces new challenges. This project extends, for the first time, transductive Zero-Shot Learning (ZSL) and Generalized Zero-Shot Learning (GZSL) approaches to the domain of 3D point cloud classification.

In the "class_name" folder, the class names of seen and unseen sets from all 3D datasets are shown. Also, the "word_vector" folder contains the semantic word vectors of 3D datasets.

Train & Test codes

Coming soon ...

Feature vector

You can download the feature vectors, which are extracted from a pretrained PointNet architecure, of ModelNet, McGill, and SHEREC2015 datasets from the following link,

feature vectors of ModelNet, McGill, and SHEREC2015 datasets using PointNet

Citation

If you find our work useful in your research, please consider citing:

@article{cheraghian2019ZSL3D,
  title={Zero-shot Learning of 3D Point Cloud Objects},
  author={Ali Cheraghian, Shafin Rahman, and Lars Petersson},
  journal={2019 16th International Conference on Machine Vision Applications (MVA)},
  year={2019}
}