DFAF3D:A dual-feature-aware anchor-free single-stage 3D detector for point clouds

Abstract

Currently, anchor-free single-stage 3D object detection methods based on point clouds have attracted extensive attention due to their high efficiency. It is crucial to enhance the ability of the center features to represent the object for such methods. In this paper, we propose a dual spatial-context feature extraction (SCFE) module to extract both spatial and context features of point clouds in parallel for 3D object detection, in which, we design a deformable offset self-attention (DOSA) structure in the context feature extraction branch to learn the relative structure information of point clouds. Correspondingly, we design a coordinate attentional feature fusion (CAFF) module, which adaptively fuses spatial and context features of different resolutions while preserving coordinate information thus making the features of center point more representative. Experiments on KITTI demonstrate that the proposed method outperforms all previous anchor-free methods in general and has comparable performance to anchor-based methods in comprehensive performance and it achieves good trade-offs in terms of accuracy, speed and parameter size.