Method

Range Conditioned Dilated CNN [RCD]


Submitted on 3 Jun. 2020 09:14 by
Alex Bewley (Google Research)

Running time:0.1 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
A two-stage detector operating on the 3D range
image view for RPN and then in 3D for the final
RCNN refinement. The RPN makes use of a novel range
conditioned dilated convolutional operator to
account for the change of scale at different
ranges.
Parameters:
RCNN, KITTI only training.
Latex Bibtex:
@inproceedings{bewley2020range,
title={Range Conditioned Dilated Convolutions for
Scale Invariant 3D Object Detection},
author={Bewley, Alex and Sun, Pei and Mensink,
Thomas and Anguelov, Dragomir and Sminchisescu,
Cristian},
booktitle = {Conference on Robot Learning (CoRL)},
year={2020}
}

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) 92.52 % 88.46 % 83.73 %
Car (3D Detection) 70.54 % 60.56 % 55.58 %
Car (Bird's Eye View) 82.26 % 75.83 % 69.61 %
This table as LaTeX


2D object detection 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