Progress Report

All emotional data sets have been combined into the 7 categories for importing into Matlab. The next step is to create the 7 gaussian functions. My goal in the morning is to start working on creating the gaussian functions as well as a classification function to classify new faces by comparing results of the gaussian functions.

Progress Report

Thus far I have found and formatted three of the emotional data sets to be passed to the various gaussian discriminant functions. It's been awhile since I've had to do anything with the data so it took some time to locate the training data set from the first experiment.

I will add another entry today updating what I was able to get accomplished this morning. My goal is to finish the data formatting script that will combine the training data into 7 emotional data sets. Using these data sets I will create 7 gaussian discriminant functions to classify each new face into one of these 7 categories.

Goals for today:
1. Finish the training data formatting script
2. Format each parameter training dataset into a combined file for each emotional category

Understanding the gaussian distribution (First case)

This morning I started working on recording the steps necessary to use a gaussian distribution. Most of what was
accomplished today was simply reading and understanding how to implement the discriminant functions.

I read the Duda, Stork and Hart book: "Pattern Classification 2nd edition" (section 2.6). This section describes the
three cases for normal distribution. After reading through the first case I created a document outlining what was needed
to create the function. I'll use this document to work from in creating the discriminant functions in Matlab.

The idea is to create a discriminant function for each emotion category. Each face vector will be passed to each of the
7 emotion discriminant functions. Each function will return a value. This value will tell us how well that vector fits into that particular
category. The one with the best fit will return the highest value of the 7 categories.

Errors can be calculated based on the number of incorrect classifications.

Tomorrow I hope to move to the next step and start coding the first case discriminant functions using Matlab.

Emotion Recognition using AAM

This is the first entry in a new blog site established to record research efforts in the study of emotion recognition using computer vision.
This research is part of an ongoing research effort in using active appearance models to recognize
emotion from facial expressions.

In April, Dr. Eric Patterson and I presented at an IASTED HCI conference in Austria on the topic.
Our first classification scheme involved a simple euclidean distance measure from the unknown expression to mean
of each emotion cluster. The results were promising, but warranted further investigation.

This blog will pick up where we left off by exploring gaussian distributions of the data set, and possibly
SVM and HMM for image sequences in video.