Implementation of the papers
"Exact Acceleration of Linear Object Detectors"
12th European Conference on Computer Vision, 2012.
"Deformable Part Models with Individual Part Scaling"
24th British Machine Vision Conference, 2013.
Copyright (c) 2013 Idiap Research Institute, <http://www.idiap.ch/>
Written by Charles Dubout <charles.dubout@idiap.ch>
INTRODUCTION
The train executable can be used to train a deformable part-based model on a
Pascal VOC dataset.
The test executable can be used to run a deformable part-based model either on
an image or on a Pascal VOC dataset.
The first time you run it it will be slow as the FFTW library will search for
the best plans using runtime measurements. The resulting plans will then be
saved to a file named wisdom.fftw and reused in the future.
COMMAND LINE OPTIONS
After building the train and test executables you can run them without any
argument to get a list of all the possible parameters.
-m,--model <file>
Read the input model from <file> (default "" for train, "model.txt" for test)
The models are stored in a text file format with the following grammar (an
example can be found in the file bicycle.txt)
Mixture := nbModels Model*
Model := nbParts bias Part*
Part := nbRows nbCols nbFeatures xOffset yOffset zOffset a b c d e f value*
Where nbModels is the number of mixture components (models); nbParts is the
number of parts (including the root) in the model; bias is the offset to add to
the scores of the model; nbRows, nbCols, nbFeatures are the dimensions of the
part filter; xOffset, yOffset, zOffset are the offsets of the part relative to
the root (anchor); a, b, c, d, e, f are the deformation coefficients
(ax^2 + bx + cy^2 + dy + ez^2 + fz); values are the filter coefficients, stored
in row-major order, and of size nbRows x nbCols x nbFeatures.
In the current implementation nbFeatures must be 32, the number of HOG features
(or 48 if FFLD was compiled with FFLD_HOGPYRAMID_EXTRA_FEATURES=ON).
One can use the provided Matlab script 'convertmodel4.m' to convert to this
format the models of P. Felzenszwalb, R. Girshick and D. McAllester.
Discriminatively Trained Deformable Part Models, Release 4.
http://people.cs.uchicago.edu/~pff/latent-release4/.
One can use the provided Matlab script 'convertmodel5.m' to convert to this
format the models of P. Felzenszwalb, R. Girshick and D. McAllester.
Discriminatively Trained Deformable Part Models, Release 5.
http://people.cs.uchicago.edu/~rbg/latent-release5/.
-n,--name <arg>
Name of the object to train or detect (default "person")
-r,--results <file>
Write the trained model or the detection results to <file> (default
"model.txt" for train, "" for test)
The test executable can outputs the list of all the detections into a file, in
the format of the Pascal VOC challenge (one line by detection, and for each
detection the scene id, the score and the bounding box: xmin, ymin, xmax, ymax).
-i,--images <folder>
Draw the detections to <folder> (default none)
The test executable can also output images with the detections drawn. In that
case it might be useful to set a higher detection threshold so as to only draw
detection with a high enough score.
-z,--nb-negatives <arg>
Maximum number of negative images to consider (default all)
It might be useful to run the detector only on a reduced number of negative
(background) Pascal VOC scenes in order to save time while training or
evaluating the performance of a detector.
-p,--padding <arg>
Amount of zero padding in HOG cells (default 6)
Must be greater or equal to half the greatest filter dimension.
-e,--interval <arg>
Number of levels per octave in the HOG pyramid (default 5)
-t,--threshold <arg>
Minimum detection threshold (default -10)
To set a negative threshold you need to use the option as in -t=-1
-v,--overlap <arg>
Minimum overlap in in latent positive search and non maxima suppression
(default 0.7 for train, 0.5 for test)
-x,--nb-components <arg>
Number of mixture components (without symmetry, default 3).
-l,--relabel <arg>
Maximum number of training iterations (default 8, half if the model has no
part).
-d,--datamine <arg>
Maximum number of data-mining iterations within each training iteration
(default 10).
-c,--C <arg>
SVM regularization constant (default 0.002).
-j,--J <arg>
SVM positive regularization constant boost (default 2).
-s,--seed <arg>
Random seed (default time(NULL)).
EXAMPLES
To train or evaluate a model on a Pascal VOC dataset you must pass to the
executable the corresponding image set file. It will look for the Pascal
annotations in the folder 'Annotations' two levels below ("../../Annotations/")
and for the jpeg images in the folder 'JPEGImages' two levels below
("../../JPEGImages/").
A complete example for train could be
./train --name bicycle --result bicycle.txt --nb-negatives 200 VOC2007/ImageSets/Main/bicycle_trainval.txt
./train --model bicycle.txt --name bicycle --result bicycle.txt --nb-negatives 2000 --relabel 1 VOC2007/ImageSets/Main/bicycle_trainval.txt
Which takes ~1 hour to complete on my laptop.
While a complete example for test could be
./test --model ../models/bicycle_2d.txt --images . --threshold=-0.5 VOC2007/ImageSets/Main/bicycle_test.txt
Which takes ~15 minutes to complete on my laptop.
MODELS
The "models" subfolder contains 2D and 3D models trained by this version of
FFLD. The *_init.txt files are the starting points for both kinds of DPM.
Also note that during training after every iterations the latest model is
continuously saved in the file "tmp.txt".