A Lightweight Structure Based on Feature Fusion for Point Cloud Analysis

Abstract

Point cloud analysis is challenging due to the irregularity and sparsity, making it difficult to capture the underlying geometric characteristics. This paper proposes a lightweight structure that effectively aggregates the local patterns and the spatial layout of them extracted from the irregular and sparse point cloud. Unlike previous works that seek sophisticated feature extraction methods, the key to this structure is simultaneously exploring the local features and their distribution features in a concise manner. Specifically, the two features, which correspond to two different scales, are extracted independently and then fused. In this way, an abstract shape-level feature that contains much shape awareness and robustness is obtained. Moreover, since the structure exhibits rapid convergence with conventional learning schedule, we upgrade it by introducing the snapshot ensemble and creatively design a more flexible and effective cyclic annealing learning schedule. We evaluate the structure on challenging benchmarks, and experiment results prove that our model achieves on-par or better performance than previous state-of-the-art (SOTA) methods, although with the simple shared multi-layer perceptrons (MLPs) as feature extractors.