<scp>ADAR‐Seg</scp>: A Lightweight Network With Adaptive Dependency and Receptive Fields for Point Cloud Segmentation in Dynamic

ABSTRACT 3D point cloud segmentation aims to assign semantic labels to each point, which supports high‐precision environment perception and safe decision‐making for autonomous driving systems. However, on one hand, the inherent sparsity of point cloud data hinders existing methods from effectively capturing fine‐grained details of small or fast‐moving objects in dynamic driving scenes. On the other hand, current Transformer‐based methods face pressing challenges in balancing model efficiency and segmentation accuracy when handling large‐scale point cloud inputs. To address these issues, this paper proposes a lightweight unified network with adaptive dependencies and receptive fields for point cloud segmentation, named ADAR‐Seg. First, to tackle the problem of nonuniform sampling in point cloud data, we design an adaptive sparse convolution‐based superpoint feature extraction (SFE) module. It adaptively captures object superpoints with significant differences in geometric structures via a learnable dynamic receptive field adjustment mechanism. Second, a centroid‐aware flexible pooling strategy is proposed, which extracts cluster centroids as proxies for long‐range dependencies by modeling superpoint correlations to reduce the computational overhead of the Transformer at redundant points. Experiments demonstrate that ADAR‐Seg improves 4% in mIoU and 5.1% in PQ, which validates its effectiveness in complex dynamic driving environments. The code is available at https://github.com/duo‐d/ADAR‐Seg .