Method

CenterNet3D:An Anchor free Object Detector for Autonomous Driving [CenterNet3D]


Submitted on 16 Jul. 2020 04:34 by
Tong Xu (Sun Yat-sen University)

Running time:0.04 s
Environment:GPU @ 1.5 Ghz (Python)

Method Description:
An anchor-free CenterNet3D Network that performs 3D object
detection without anchors. Our CenterNet3D uses keypoint
estimation to find center points and directly regresses 3D
bounding boxes. Beside, our CenterNet3D is Non-Maximum
Suppression free which makes it more efficient and simpler. 
Parameters:
TBD
Latex Bibtex:
@misc{2007.07214,
Author = {Guojun Wang and Bin Tian and Yunfeng Ai and Tong Xu and Long
Chen and Dongpu Cao},
Title = {CenterNet3D:An Anchor free Object Detector for Autonomous
Driving},
Year = {2020},
Eprint = {arXiv:2007.07214},
}

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) 95.76 % 92.69 % 89.81 %
Car (Orientation) 95.71 % 92.48 % 89.54 %
Car (3D Detection) 86.20 % 77.90 % 73.03 %
Car (Bird's Eye View) 91.80 % 88.46 % 83.62 %
This table as LaTeX

2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot




eXTReMe Tracker