Method

RTS3D: Real-time Stereo 3D Detection from Feature-Consistency Embedding Space for Autonomous Driving [RTS3D]


Submitted on 31 Aug. 2020 13:10 by
peixuan li (University of Chinese Academy of Sciences)

Running time:0.03 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
Experiments on KITTI benchmark show that RTS3D is
the first true real-time system (FPS>24) for stereo
image 3D detection while achieves 10% improvement in
average precision comparing with precious state-of-
the-art method.
Parameters:
none
Latex Bibtex:

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) 80.48 % 73.08 % 64.02 %
Car (Orientation) 80.36 % 72.74 % 63.65 %
Car (3D Detection) 58.51 % 37.38 % 31.12 %
Car (Bird's Eye View) 72.17 % 51.79 % 43.19 %
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