Semantic-aware object detection for 3D point cloud

Abstract

In this paper, we propose a semantic-aware network (SA-Net) to improve the performance of 3D point cloud object detection, which embeds a backward attention module and a semantic attention module. The backward attention module utilizes high-level semantic features from the encoder via fusing multi-level encoder features hierarchically. In this stage, high-level features are transformed into an attention map to modulate low-level features backward. Meanwhile, semantic attention module obtains a semantic segmentation map of a given point cloud scene through supervised learning. This can be transformed into a semantic attention map and embedded into the detection head for better detection. Equipped with these modules, SA-Net can greatly improve the performance of object detection. Extensive experiments on KITTI demonstrate that the proposed method can achieve competitive results against the state-of-the-art methods.