Point Cloud Classification via Learnable Memory Bank

Lisa Liu, William Y. Wang, Pingping Cai

Abstract: Point cloud analysis has gained significant importance across various fields, in which a pivotal and formidable aspect is the classification of point clouds. To solve this problem, most existing deep learning methods follow the two-stage framework, wherein an encoder first extracts shape features and then an MLP-based classification head categorizes objects. To improve the classification performance, prior efforts have concentrated on refining the shape feature extraction through intricate encoder designs. However, the design of classification heads has remained unexplored. In this paper, we focus on the second stage and introduce a novel classification head integrated with a Learnable Memory Bank (LMB), tailored for the classification of 3D point-cloud objects. The LMB aims to learn representative category feature vectors from training objects. Subsequently, a similarity-based feature-matching mechanism is employed to generate the predicted class logits. The proposed LMB classification head can be seamlessly integrated into existing feature extraction backbones. Empirical evaluations prove the efficacy of the proposed LMB classification head, showcasing performance on par with state-of-the-art methods.