Mehran Kafai, Bir Bhanu, Le Analumni.cs.ucr.edu/~mkafai/papers/Poster_ICPR2012.pdf · Mehran Kafai,...

Mehran Kafai, Bir Bhanu, Le An Center for Research in Intelligent Systems, University of California, Riverside, USA

1. Introduction Goal:

Robustly estimate the pose of a given face by

classifying the face to a predetermined set of

poses.

Sample images for pose estimation

Challenge: Limited discriminative power of commonly

used classifiers such as LDA, SVM, or NN

results in low classification accuracy.

Solution: • We present the Cluster-Classification

Bayesian Network (CCBN), a graphical

model specifically designed for

classification after clustering.

• A pose layout is defined where similar

poses are assigned to the same group.

• The discriminative power increases within

the same group when similar yet different

poses are present.

• The CCBN is trained on multi-pose face

image databases.

• 2200 images from FEI database, 200 individuals, 11 poses from

profile left to profile right.

• 4200 images from CAS-PEAL database, 200 individuals, 21 poses

from 9 cameras spaced in a horizontal semicircular shelf.

2. Technical Approach

Sample images from FEI database with pose layout overlay

Sample images from CAS-PEAL database with pose layout overlay

Accuracy comparison on CAS-PEAL

3. Results

Accuracy comparison on FEI

• Define pose layout such that similar poses are

located in neighboring locations (1-11 are pose

IDs from FEI database).

• Layout can be one, two, or three dimensional.

• Layout is partitioned into groups. Each group

holds similar poses. Each pose belongs to at least

one group.

• Partitioning may be performed using systematic

or heuristic methods.

Sample pose layout

• We introduced a novel pose estimation method using the

Cluster-Classification Bayesian Network (CCBN).

• By clustering similar poses into the same block, the trained

classifier is more discriminative in these similar poses.

• Experimental results show that the CCBN has superior

performance compared to the NN, LDA, and SVM classifiers.

• CCBN with HOG as the feature descriptor achieves the

highest performance.

Goal is to compute max 𝑐𝑐𝑐𝑐𝑐𝑃(𝐶|𝐹)

Definition of CCBN nodes: • C : Class node

• Holds probability distribution over all poses • Discrete node • Size equal to number of poses

• F : Feature node • Corresponds to feature vector representing the data • Discrete or continuous based on data • Size equal to dimensionality of data

• 𝑩𝟏, … ,𝑩𝒎 ∶ group nodes • 𝐵𝑖 determines membership probability of data to group 𝑖 vs. all other groups.

and

where

Probability of a given data 𝑓 being from class 𝑐𝑘 is formulated as:

which represents the pose with the highest probability. Joint probability distribution with class node 𝐶, feature node 𝐹, and group nodes 𝐵1,𝐵2, … ,𝐵𝑚 is defined as:

• Each group is represented by a node B in the middle layer of the

corresponding CCBN.

4. Conclusions

• CCBN has greater accuracy than the other three

classifiers for 9 out of the total 11 poses on FEI database.

• The average accuracy for CCBN is 3.48% more than

SVM, 5.81% more than LDA, and 11.21% more than NN.

• Images resized to 32x32, each image represented by 240-

dimensional HOG feature vector.

• 10-fold cross validation, 150 individuals are used for training

and 50 individuals for testing.

ROC plot for performance on FEI