Target-aware attentional network for rare class segmentation in large-scale LiDAR point clouds

Semantic interpretation of 3D scenes poses a formidable challenge in point cloud processing, which also stands as a requisite undertaking across various fields of application involving point clouds. Although a number of point cloud segmentation methods have achieved leading performance, 3D rare class segmentation continues to be a challenge owing to the imbalanced distribution of fine-grained classes and the complexity of large scenes. In this paper, we present target-aware attentional network (TaaNet), a novel mask-constrained attention framework to address 3D semantic segmentation of imbalanced classes in large-scale point clouds. Adapting the self-attention mechanism, a hierarchical aggregation strategy is first applied to enhance the learning of point-wise features across various scales, which leverages both global and local perspectives to guarantee presence of fine-grained patterns in the case of scenes with high complexity. Subsequently, rare target masks are imposed by a contextual module on the hierarchical features. Specifically, a target-aware aggregator is proposed to boost discriminative features of rare classes, which constrains hierarchical features with learnable adaptive weights and simultaneously embeds confidence constraints of rare classes. Furthermore, a target pseudo-labeling strategy based on strong contour cues of rare classes is designed, which effectively delivers instance-level supervisory signals restricted to rare targets only. We conducted thorough experiments on four multi-platform LiDAR benchmarks, i.e., airborne, mobile and terrestrial platforms, to assess the performance of our framework. Results demonstrate that compared to other commonly used advanced segmentation methods, our method can obtain not only high segmentation accuracy but also remarkable F1-scores in rare classes. Ina submission to the official ranking page of Hessigheim 3D benchmark, our approach achieves a state-of-the-art mean F1-score of 83.84% and an outstanding overall accuracy (OA) of 90.45%. In particular, the F1-scores of rare classes namely vehicles and chimneys notably exceed the average of other published methods by a wide margin, boosting by 32.00% and 32.46%, respectively. Additionally, extensive experimental analysis on benchmarks collected from multiple platforms, Paris-Lille-3D, Semantic3D and WHU-Urban3D, validates the robustness and effectiveness of the proposed method.