Test the performance of behavior recognition using the cuboid representation.
Given n sets of data, each containing multiple data instances, we train on 1 set at a
time, and then test on each of the remaining sets. Thus there are (n x n) separate
training/testing scenarios. [Note: to get performance on set i given training on i we
use cross validation WITHIN the set]. Note that this is not cross validation where
training occurs on all but (n-1) of the sets and testing on the remaining one, giving a
total of (n) training/testing scenarios.
Clustering is performed (using recog_cluster) on cuboids from the single training set.
Once the clustering is obtained, each cuboid in all the clips in all the sets is
assigned a type and each clip is converted to a histogram of cuboid types (using
recog_clipsdesc). Afterwards standard classification techniques are used to train/test.
Parameters for clustering and classification can be specified inside this file.
INPUTS
DATASETS - array of structs, should have the fields:
.IDX - length N vector of clip types
.desc - length N cell vector of cuboid descriptors
.ncilps - N: number of clips
k - number of clusters
nreps - number of repetitions
OUTPUTS
ER - error matricies [nsets x nsets] - averaged over nreps
CMS - confusion matricies [nclass x nclass x nsets x nsets] - averaged over nreps
See also RECOGNITION_DEMO, RECOG_TEST_NFOLD, NFOLDXVAL, RECOG_CLUSTER, RECOG_CLIPSDESC