Method

3D Vehicle Detection via Geometry Constrained Keypoints [3D-GCK]


Submitted on 13 Jan. 2020 17:39 by
Nils Gählert (Mercedes-Benz, R&D)

Running time:24 ms
Environment:Tesla V100

Method Description:
3D-GCK is an extension to 2D object detection
frameworks that allows efficient 3D object
detection from monocular RGB images in realtime.
Parameters:
TBD
Latex Bibtex:
@inproceedings{gahlert2020single,
title={Single-Shot 3D Detection of Vehicles
from Monocular RGB Images via Geometrically
Constrained Keypoints in Real-Time},
author={Gählert, Nils and Wan, Jun-Jun
and Jourdan, Nicolas and Finkbeiner, Jan and
Franke, Uwe and Denzler, Joachim},
booktitle={2020 IEEE Intelligent Vehicles
Symposium (IV)},
year={2020},
organization={IEEE}
}

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) 89.55 % 80.19 % 68.08 %
Car (Orientation) 88.59 % 78.44 % 66.28 %
Car (3D Detection) 3.27 % 2.52 % 2.11 %
Car (Bird's Eye View) 5.79 % 4.57 % 3.64 %
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