Information Technology Reference
In-Depth Information
(a) Camera 1
(b) Camera 2
(c) Camera 3
(d) Camera 4
(e) Camera 5
Fig. 5. The kick action in the IXMAS dataset from the five available views
x i ,y i }
, model parameters are adjusted max-
imizing the L 2 regularized conditional likelihood function of the model:
L ( θ )= n
Given a set of training samples
{
2
log P ( y i | x i )+ ||
θ
||
(13)
2 σ
i =1
The optimal parameters θ maximizing the conditional likelihood function are
found using Quasi-Newton gradient based methods. Both the computation of
the posterior probability on equation 11 and the auxiliary distributions that
appear on the gradient of 13 can be eciently made using belief propagation,as
proposed in [14].
5 Experiments
5.1 Experimental Setup
The proposed algorithms are going to be tested in the classification of IXMAS
dataset [21]. This dataset contains 11 actions performed by 10 different actors
at least 3 times each. The actions are recorded from 5 different viewpoints. The
algorithms are going to be tested using Leave-One-Actor-Out Cross Validation
(LOAO-CV): The algorithms are trained with all the actors unless one, used for
validation.
The system is going to be tested using the action descriptor proposed by Tran
et al. [18], combining optical flow and appearance information. It is used in the
system because it has shown a high experimental performance. The bounding
box of a human being is normalized to a square box, from which human shape
 
Search WWH ::




Custom Search