PVP-SSD: Point-Voxel Fusion with Partitioned Point Cloud Sampling for Anchor-Free Single-Stage Small 3D Object Detection
Loading...
Date
2024
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Single-stage object detection from 3D point clouds in autonomous driving faces significant challenges, particularly in accurately detecting small objects. To address this issue, we propose a novel method called Point-Voxel dual-branch feature extraction with Partitioned point cloud sampling for anchor-free Single-Stage Detection of 3D objects (PVP-SSD). The network comprises two branches: a point branch and a voxel branch. In the point branch, a partitioned point cloud sampling strategy leverages axial features to divide the point cloud. Then, it assigns different sampling weights to various segments to enhance the sampling accuracy. Additionally, a local feature enhancement module explicitly calculates the correlation between key points and query points, improving the extraction of local features. In the voxel branch, we use 3D sparse convolution to extract instance structural features efficiently. The point-voxel dual-branch fusion dynamically integrates instance features extracted from both branches using a self-attention mechanism, which contains not only the category information of the detected object but also the spatial dimensions and heading angle. Consequently, PVP-SSD achieves a certain balance between preserving detailed information and maintaining structural integrity. Experimental results on the KITTI and ONCE datasets demonstrate that PVP-SSD excels in multi-category small 3D object detection.
Description
@inproceedings{10.2312:pg.20241279,
booktitle = {Pacific Graphics Conference Papers and Posters},
editor = {Chen, Renjie and Ritschel, Tobias and Whiting, Emily},
title = {{PVP-SSD: Point-Voxel Fusion with Partitioned Point Cloud Sampling for Anchor-Free Single-Stage Small 3D Object Detection}},
author = {Wu, Xinlin and Tian, Yibin and Pan, Yin and Zhang, Zhiyuan and Wu, Xuesong and Wang, Ruisheng and Zeng, Zhi},
year = {2024},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-250-9},
DOI = {10.2312/pg.20241279}
}