Cross-Modal Feature Learning for Point Cloud Classification

李婉桦, Chun-Yang Zhang

Abstract: Traditional 3D shape classification methods face challenges due to the complexity and variability of point cloud data. To address this issue, we propose CFL framework that integrates proxy weights from two modalities through an average fusion approach and adopts a proxy-based contrastive learning strategy to enhance feature representation. By using the average fusion method. we can effectively capture both texture features and geometric features via integrating complementary information from different modalities. Furthermore, the proxy-based contrastive learning method is designed to acquire representations by learning a unified space. Experimental results demonstrate that our CFL method significantly improves classification performance on the ModelNet10 dataset. Meanwhile, we conduct ablation studies on ModelNet10 dataset to confirm the pivotal role of the average fusion method and the proxy-based contrastive learning method, highlighting the potential of cross-modal feature learning method in advancing 3D shape classification tasks.

Submission Number: 167