Fully Sparse Two-Stage Object Detection Network with Auxiliary Learning for ALS Point Clouds

Airborne Laser Scanning (ALS) systems can efficiently and directly acquire three-dimensional (3D) spatial information of scenes. However, due to the long sampling range and occlusion by tree canopies, which obscures details of ground objects, ALS point clouds often exhibit high similarity between objects and backgrounds while containing sparse point distributions for individual objects. These challenges motivate our proposal of Aux-RCNN, a fully sparse two-stage network that leverages auxiliary learning for robust 3D object detection. The network enhances feature representation capability through an auxiliary branch based on sampling-grouping, effectively suppressing background interference. Furthermore, it refines proposal predictions by self-attention-based RoI-Pooling, enabling the network to focus on sparse objects. Due to the absence of open-source ALS point cloud dataset for 3D object detection, we construct a Harbin-L2 dataset to validate our approach. Experimental results demonstrate that Aux-RCNN effectively handles background similarity and object sparsity in ALS point clouds, achieving superior performance.