Method

Hierarchical And-Or Model [AOG]
http://www.stat.ucla.edu/~boli/

Submitted on 7 Aug. 2015 00:19 by
Alan Li (Beijing Institute of Technology)

Running time:3 s
Environment:4 cores @ 2.5 Ghz (Matlab)

Method Description:
Integrating context and occlusion modelling. This
submission is an updated version of the model in our
ECCV14 paper.

This model mainly focuses on car detection (as
descriped in the original paper), and we omit the
view estimation task (setting alpha angle of each
detected car to 0.5 by default).

This is an updated version of our model, at this time,
we use all the training data to learn the and-or model.
While for the one described in our eccv14 paper, for
evaluation convenience, we only used half of the
training data.
Parameters:
C = 0.006;
N = 2;
NC = 10;
SC = 16;
Latex Bibtex:
@article{Wu2016PAMI,
author = {Tianfu Wu and
Bo Li and
Song{-}Chun Zhu},
title = {Learning And-Or Models to Represent
Context and Occlusion for Car
Detection and Viewpoint Estimation},
journal = {TPAMI},
year = {2016}
}
@inproceedings{Li2014ECCV,
author = {Bo Li and Tianfu Wu and Song-Chun
Zhu},
title = {Integrating Context and Occlusion
for Car Detection by Hierarchical And-Or Model},
booktitle = {ECCV},
year = {2014}
}

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) 86.08 % 76.24 % 61.51 %
Car (Orientation) 33.28 % 29.81 % 23.91 %
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




eXTReMe Tracker