Method

VoxelJones [VoxelJones]
https://github.com/motrom/voxeljones

Submitted on 3 Nov. 2019 22:47 by
Michael Motro (University of Texas at Austin)

Running time:.18 s
Environment:1 core @ 2.5 Ghz (Python + C/C++)

Method Description:
Viola-Jones-type detector. The base feature is a binary voxel grid of lidar returns. The 3d integral image is computed with simd commands for speed. A boosted decision tree classifier is applied at every voxel step for 32 possible orientations. Nms removes overlapping detections.
Parameters:
classifier = 30 depth-3 decision trees
voxel resolution = .125m.
first 10 trees are applied to fewer anchors (.25m spacing and 16 orientations)
Latex Bibtex:
@article{motro2019vehicular,
title={Vehicular Multi-object Tracking with Persistent Detector Failures},
author={Motro, Michael and Ghosh, Joydeep},
journal={arXiv preprint arXiv:1907.11306},
year={2019}
}

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) 43.89 % 36.31 % 34.16 %
Car (Orientation) 17.83 % 15.41 % 14.13 %
Car (3D Detection) 7.39 % 6.35 % 5.80 %
Car (Bird's Eye View) 66.21 % 53.96 % 47.66 %
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