DepthVoting: A Few-Shot Point Cloud Classification Model Incorporating a Projection-Based Voting Mechanism

Yunhui Zhu, Jiajing Chen, Senem Velipasalar

Abstract: Despite the significant progress in few-shot 2D image classification, few-shot 3D point cloud classification remains relatively under-explored, particularly in addressing the challenges posed by missing points in 3D point clouds. Most existing methods for few-shot 3D point cloud classification are point-based, and thus, highly sensitive to missing points. Despite recent attempts, such as ViewNet, which introduce projection-based backbones to increase robustness against missing points, the reliance on max pooling, to extract information from multiple images simultaneously, makes them prone to information loss. To address these limitations, we introduce DepthVoting, a novel projection-based approach, for few-shot 3D point cloud classification. Instead of extracting features from multiple projection images simultaneously, DepthVoting captures features from pairs of projection images (obtained from opposite view angles) separately, enhancing the extraction of more comprehensive information. These features are sent to multiple few-shot heads, which share parameters. To further refine predictions, DepthVoting incorporates a voting mechanism, allowing contribution and incorporating information from different pairs. We conduct extensive experiments on three datasets, namely ModelNet40, ModelNet40-C, and ScanObjectNN, along with cross-validation. Our proposed method consistently outperforms the state-of-the-art baselines on all datasets in terms of average accuracy with even higher margins on the challenging ScanObjectNN dataset.