DGANet: A Dilated Graph Attention-Based Network for Local Feature Extraction on 3D Point Clouds

Feature extraction on point clouds is an essential task when analyzing and processing of 3D scenes. However, there still remains a challenge to adequately exploit local fine-grained features cloud data due its irregular unordered structure in space. To alleviate this problem, Dilated Graph Attention-based Network (DGANet) with certain feature for learning ability proposed. Specifically, we first build dilated graph-like region each input establish the long-range spatial correlation towards corresponding neighbors, which allows proposed network access wider range geometric information points their dependencies. Moreover, by integrating graph attention module (DGAM) implemented novel offset–attention mechanism, promises highlight differing importance edge constructed uniquely learn discrepancy attributes between connected pairs. Finally, all learned are further aggregated, allowing most significant representation regions graph–attention pooling fully extract detailed point. The validation experiments using two challenging benchmark datasets demonstrate effectiveness powerful generation our DGANet both object classification segmentation tasks.