The KITTI Vision Benchmark Suite

Method

PointRGBNet [PointRGBNet]

Submitted on 29 Jun. 2021 10:46 by
William Xie (Tianjin University)

Running time:		0.08 s
Environment:		4 cores @ 2.5 Ghz (Python + C/C++)

Method Description:

We use 6D RGB point clouds as input,and deep
convolution network is adopted to learn features
from input. The results of our proposed network is
incredible.

Parameters:

alpha=0.25 gamma=2

Latex Bibtex:

@article{Xie Desheng:340,
author = {Xie Desheng,Xu Youchun,Lu Feng,Pan
Shiju},
title = {Real-time Detection of 3D Objects
Based on Multi-Sensor Information Fusion},
publisher = {Automotive Engineering},
year = {2022},
journal = {Automotive Engineering},
volume = {44},
number = {3},
eid = {340},
numpages = {9},
pages = {340},
keywords = {;object detection;2D image;3D point
clouds;deep neural network},
url =
{http://www.qichegongcheng.com/CN/abstract/arti
cle_1214.shtml},
doi = {10.19562/j.chinasae.qcgc.2022.03.005}
}

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).

Benchmark	Easy	Moderate	Hard
Car (Detection)	95.40 %	91.48 %	86.50 %
Car (Orientation)	95.39 %	91.33 %	86.29 %
Car (3D Detection)	83.99 %	73.49 %	68.56 %
Car (Bird's Eye View)	91.39 %	85.73 %	80.68 %
Pedestrian (Detection)	44.35 %	33.92 %	30.43 %
Pedestrian (Orientation)	43.08 %	32.57 %	29.17 %
Pedestrian (3D Detection)	34.77 %	26.40 %	24.03 %
Pedestrian (Bird's Eye View)	38.07 %	29.32 %	26.94 %
Cyclist (Detection)	79.87 %	65.98 %	59.75 %
Cyclist (Orientation)	79.64 %	65.68 %	59.48 %
Cyclist (3D Detection)	67.05 %	52.15 %	46.78 %
Cyclist (Bird's Eye View)	73.09 %	57.59 %	51.78 %