Using Dynamic Strain Pattern from Video to Recognize Faces

A new face recognition method that utilizes dynamic information extracted from video is developed. The method is designed based on the hypothesis that the strain pattern exhibited during facial expression provides a unique "fingerprint" for recognition. First, a dense motion field is obtained with an optical flow algorithm. Strain pattern is then derived from the motion field by a finite difference method. In experiments with 30 subjects, PCA results indicate that strain pattern is a very useful biometric, especially when dealing with extreme conditions such as shadow light and face camouflage, given which conventional face recognition methods usually fail.
Compared to our previous work, we made the following improvements: (1) Motion computation is automated with an optical flow algorithm that replaces the manual feature matching approach; (2) Strain computation is simplified with a more efficient finite difference method; (3) A sequence of strain patterns, rather than a snap shot of strain, is gathered which, enables us to study its invariance property.

Performance Evaluation of Object Detection and Tracking Algorithms

The need for empirical evaluation metrics and algorithms is well acknowledged in the field of computer vision. The process leads to precise insights to understanding current technological capabilities and also helps in measuring progress. Hence designing good and meaningful performance measures is very critical.
The goal of this project is to evaluate the performance of Video Analysis and Content Extraction (VACE) algorithms that detect and track face, text, hand, person, and vehicle objects across the following four domains: Meetings, Broadcast News, Surveillance, and UAV.
For each task, we have developed a suite of diagnostic measures that capture different aspects of an algorithm’s performance and a comprehensive measure which captures many aspects of the task in a single score. The diagnostic measures are useful to the researchers in their failure analysis while the comprehensive measures will be used to provide a summative analysis of overall system performance.
Depending on the annotation approach for a specific task-domain pair, we classify the evaluations into two categories:

1. Area based metrics for object bounding annotations
1. Sequence Frame Detection Accuracy (SFDA)
2. Average Tracking Accuracy (ATA)
2. Distance based metrics for point annotations
1. Sequence Frame Detection Accuracy – Distance-based (SFDA-D)
2. Average Tracking Accuracy – Distance-based (ATA-D)