recognition_demo.m

cuboid democuboid democuboid democuboid demo

% Describes all steps of behavior recognition; example for facial expressions.
%
% The file describes the following:
%   1) data format [3 choices]
%   2) feature detection & description [2 choices]
%   3) behavior classification 
%
% Data and additional information can be obtained at:
%   http://vision.ucsd.edu/~pdollar/research/research.html
%
% ----------------------------------------------------------------------------------------
% DATA FORMAT
%
% First the directory of the data must be specified.  
% Alter datadir.m to point to the location of the data. 
%
% Behavior data can be represented in 1 of 3 ways.  In each case data is divided into a
% number of 'sets', each set contains data that should be treated together (for training
% and testing, see generating results for more info).  In each case, a set contains a
% number of behavior clips. The three methods for representing a set are:
%       [set_ind] the 2 digit set index, '00','01',...
%       [cliptype] represents behavior type, such as 'grooming' or 'smiling'  
%       [instance] represents the 3 digit instance number - '000','001',...
%   1) .avi files in a datadir/set[set_ind]/, named '[cliptype][instance].avi'
%   2) .mat files in a datadir/set[set_ind]/, named 'clip_[cliptype][instance].mat'
%      When loaded the file contains two matlab variables: 'I' and 'clipname'
%   3) a single DATASETS struct, kept in memory as DATASETS, described below.  
%
% If all the sets can be stored in memory at once, the third option can be used.  DATASETS
% is an array of nsets elements, where each element is a struct representing a set of
% data.  Each DATASETS should initally have the following two fields: IS and IDX.  IS is
% either a ...xN 4D array of N clips or an N element cell array of clips (3D arrays).  IDX
% should be a length N uint8 vector of clip types.  Note that as processing proceeds the
% contents of DATASETS will change.
%
% One can convert between the formats using the conv_* functions. To go from DATASETS
% format to .avi format, go through the .mat format.
%
% ----------------------------------------------------------------------------------------
% FEATURE DETECTION & DESCRIPTION
%
% These functions are used to detect cuboids and than apply descriptors to them, creating
% data that can than be used in various training / testing scenarios described next.  The
% two function featuresSM and featuresLG both return the same output.  The difference is
% that featuresSM works with data fully in memory (the DATASETS format), while featuresLG
% writes things back and forth to the hard disk (using the .mat format).  If the files are
% originally in .avi format, they need to be converted to either .mat format or the
% DATASETS format.
%
% ----------------------------------------------------------------------------------------
% BEHAVIOR CLASSIFICATION
%   See recog_test or recog_test_nfold
%
% ----------------------------------------------------------------------------------------
% EXAMPLE below 

% set directory in datadir.m
nsets = 4; cliptypes = { 'anger','disgust','fear','joy','sadness','surprise' };
par_stfeatures = {2, 3, 1, 2e-4, [], 1.85, 1, 1, 0};
cubdesc = imagedesc_generate( 1, 'GRAD', -1 ); kpca = 100;
for i=0:(nsets-1) conv_movies2clips( i ); end;
if( 0 ) 
    % Using DATASETS format
    DATASETS = conv_clips2datasets( nsets, cliptypes );
    save( [datadir() '/DATASETS.mat'], 'DATASETS' );
    [DATASETS,cubdesc,cuboids] = featuresSM( DATASETS, par_stfeatures, cubdesc, kpca );
    savefields = {'DATASETS', 'cubdesc', 'cuboids', 'cliptypes'};
    save( [datadir() '/DATASETSprSM.mat'], savefields{:} );
else
    %%% Using .mat format
    [DATASETS,cubdesc,cuboids] = featuresLG( nsets, cliptypes, ...
                                     par_stfeatures, 20, cubdesc, kpca );
    savefields = {'DATASETS', 'cubdesc', 'cuboids', 'cliptypes'};
    save( [datadir() '/DATASETSprLG.mat'], savefields{:} );
end
[ER,CMS] = recog_test( DATASETS, 50, 1 );
confmatrix_show( CMS(:,:,1,2), cliptypes );