Instance-aware sampling and voxel-transformer encoding for single-stage 3D object detection

Abstract

In point cloud 3D object detection tasks, single-stage detectors offer fast inference but are less accurate than two-stage detectors. We point out two main problems: first, traditional methods deal with the whole point cloud, making them vulnerable to background noise interference; second, existing methods exhibit insufficient single-channel feature encoding capability. Therefore, this paper proposes Instance-Aware Sampling and Voxel-Transformer Encoding for Single-Stage 3D Object Detection (IAVT-SSD). Specifically, we design an Instance-Aware Weighted Sampling Strategy to filter out ground reflection points, enhancing the model's focus on the foreground points. Meanwhile, we introduce a Voxel-Transformer Dual-Channel Feature Encoding Module to capture more comprehensive features through two independent channels, efficiently fusing non-empty voxels and remote context information. In addition, a Collaborative Enhancement Branch is designed to predict the complete structure of the object. Experiments show that IAVT-SSD achieves a good balance of accuracy and speed, with an inference speed of 42 FPS (frames per second) and a mAP (mean average precision) of 81.70% on the KITTI dataset, and a mAP of 66.96% on the ONCE dataset, validating its effectiveness and superiority.