A multi-scale 3D scene flow estimation network based on transformer
Yingxin Wei, Liwei Chen · 2024
The purpose of scene flow estimation is to capture the intricate motion patterns within point clouds across successive frames. We incorporate the channel self-attention (CSA) into the estimation of scene flow for point clouds. Specifically, the channel self-attention mechanism prioritizes channels with significant disparities to prevent the merging of similar and redundant information. Through the subtraction operation embedded in the structure, attention weights are concentrated in regions with salient characteristics and crucial information within the point cloud, thereby reducing attention toward noise points. By incorporating channel self-attention at each stage of the network, we can extract local features and capture rich contextual information. Additionally, we introduce a channel excitation module to enhance the global correlation among channels and enhance the model's representation capability by introducing additional nonlinear relationships. Experimental results verify that our proposed method is effective.