Dense Point Diffusion for 3D Object Detection

The backbone network adopted in state-of-the-art 3D object detectors lacks a good balance between high point resolution and large receptive field, both of which are desirable for object detection on point clouds. This work proposes Dense Point Diffusion module, a novel backbone network that solves t...

Full description

Saved in:
Bibliographic Details
Main Authors: Liu, Xu, Cao, Jiayan, Bi, Qianqian, Wang, Jian, Shi, Boxin, Wei, Yichen
Format: Conference Proceeding
Language:English
Subjects:
Convolution
Feature extraction
Object detection
Quantization (signal)
Task analysis
Three-dimensional displays
Two dimensional displays
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The backbone network adopted in state-of-the-art 3D object detectors lacks a good balance between high point resolution and large receptive field, both of which are desirable for object detection on point clouds. This work proposes Dense Point Diffusion module, a novel backbone network that solves these issues. It adopts dilated point convolution as a building block to enlarge the receptive field and retain the point resolution at the same time. Further, a number of such layers are densely connected, giving rise to large receptive field and multi-scale feature fusion, which are effective for object detection task. Comprehensive experiments verify the efficacy of our approach. The source code 1 has been released to facilitate the reproduction of the results. 1 https://github.com/AsahiLiu/PointDetectron
ISSN:2475-7888
DOI:10.1109/3DV50981.2020.00086