AGS-SSD: Attention-Guided Sampling for 3D Single-Stage Detector

Abstract

3D object detection based on LiDAR point cloud has always been challenging. Existing point cloud downsampling approaches often use heuristic algorithms such as farthest point sampling (FPS) to extract the features from a massive raw point cloud. However, FPS has disadvantages such as low operating efficiency and inability to sample key areas. This paper presents an attention-guided downsampling method for point-cloud-based 3D object detection, named AGS-SSD. The method contains two modules: PEA (point external attention) and A-FPS (attention-guided FPS). PEA explores the correlation between the data and uses the external attention mechanism to extract the semantic features in the set abstraction stage. The semantic information, including the relationship between the samples, is sent to the candidate point generation module as context points. A-FPS weighs the point cloud according to the generated attention map and samples the foreground points with rich semantic information as candidate points. The experimental results show that our method achieves significant improvements with novel architectures against the baseline and runs at 24 frames per second for inference.