Method

Voting for Voting in Online Point Cloud Object Detection [la] [Vote3D]


Submitted on 20 Apr. 2015 19:01 by
Dominic Zeng Wang (University of Oxford)

Running time:0.5 s
Environment:4 cores @ 2.8 Ghz (C/C++)

Method Description:
We apply the sliding window paradigm to pure 3D data,
no information from the images is used. Efficiency is
achieved via conducting the sliding window search in the
form of voting. 3D detections are then projected into the
image for the sake of evaluation with the KITTI vision
benchmark only.
Parameters:
grid_resolution = 0.2m, num_angular_bins = 8,
overlap_threshold = 0.01 (Car) 0.1 (Cyclist) 0.5 (Pedestrian)
Latex Bibtex:
@inproceedings{Wang2015RSS,
author = {Dominic Zeng Wang and Ingmar Posner},
title = {Voting for Voting in Online Point Cloud Object
Detection},
booktitle = {Proceedings of Robotics: Science and
Systems},
year = {2015},
address = {Rome, Italy},
month = {July}
}

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) 54.38 % 45.94 % 40.48 %
Pedestrian (Detection) 42.66 % 33.04 % 30.59 %
Cyclist (Detection) 39.81 % 27.99 % 25.19 %
This table as LaTeX


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



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



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




eXTReMe Tracker