Method

Aggregate Channel Features [ACF]


Submitted on 26 Sep. 2014 06:27 by
Cesar Cadena Lerma (University of Adelaide)

Running time:0.2 s
Environment:1 core @ >3.5 Ghz (Matlab + C/C++)

Method Description:
Out-of-the-box Aggregate channel features object
detector from Piotr Doll\'ar with small adaptations
to KITTI. One model is learnt for pedestrians
and 8 (orientations) models for cars. There is not
exhaustive search for the detector parameters.
Parameters:
In training default parameters, except:

opts.nNeg = 10000;
opts.nAccNeg = 50000;
opts.nWeak=[32 128 512 2048];

Pedestrian:
opts.modelDs=[50 20.5]; opts.modelDsPad=[64 32];
opts.pJitter=struct('flip',1);
opts.pBoost.pTree.maxDepth = 3;
opts.pLoad={'squarify',{3,.41}, 'lbls',
'Pedestrian'},'ilbls',{'Cyclist',
'Person_sitting', 'DontCare'}, 'hRng',[50 inf] };

Car:
opts.pBoost.pTree.fracFtrs=1/16;
opts.pLoad={ 'lbls',
{'Car_model_equalto_i'},'ilbls',{'Van', 'Truck',
'DontCare','Car_model_differentto_i'},'hRng',[25
inf] };

model_i= 1 or 5: (sideviews)
opts.modelDs=[50 150]; opts.modelDsPad=[64 176];
model_i= 2 or 4 or 6 or 8:
opts.modelDs=[50 100]; opts.modelDsPad=[63 115];
model_i= 3 or 7: (frontal and rear views)
opts.modelDs=[50 75]; opts.modelDsPad=[63 91.5];
-------------------------------------------------
------------
In testing:
detector =
acfModify(detector,'cascThr',-1,'cascCal',0);
detector = acfModify(detector,'nOctUp',1);
NMS = 0.5;
Latex Bibtex:
@article{Dollar2014PAMI,
author = {Piotr Doll\'ar and Ron Appel and Serge
Belongie and Pietro Perona},
title = {Fast Feature Pyramids for Object
Detection},
journal = {PAMI},
year = {2014},
}
@misc{PMT,
author = {Piotr Doll\'ar},
title = {{P}iotr's {I}mage and {V}ideo
{M}atlab {T}oolbox ({PMT})},
howpublished =
{\url{http://vision.ucsd.edu/~pdollar/toolbox/doc
/index.html}}
}

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) 63.05 % 54.09 % 41.81 %
Pedestrian (Detection) 48.42 % 39.12 % 35.03 %
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




eXTReMe Tracker