MIT Department of Electrical Engineering & Computer Science

Novel Machine Learning Approaches for Computer Vision

Shumeet Baluja
Carnegie Mellon University

Abstract

A crucial building block for automatic video indexing is object detection. We have developed a neural-network based system for the detection of frontal, upright, faces in cluttered scenes. Using novel techniques for network sensitivity analysis, I will demonstrate the types of features the networks exploit to localize the face. The detection algorithms compare favorably, both in terms of accuracy and speed, with other state-of-the-art face detection systems.

The above project addresses the need for focusing attention by examining all regions of the input serially. However, in complex dynamic environments, such as driving, it is crucial to quickly and accurately discriminate between relevant and irrelevant features. The second approach explores expectation-based selective attention. Here, the network's ability to predict the future (or more precisely, its inability to predict the noise) is used to focus its attention in domains with temporal structure. These selective attention algorithms have been successfully used in systems for autonomous navigation, visual hand-tracking, and the detection of errors in the plasma-etch step of semiconductor wafer fabrication.

The need for focusing attention in vision-based systems is paralleled by the need for focusing computational resources in search algorithms. Time permitting, I will present an efficient abstraction of the standard genetic algorithm (GA), termed Population-Based Incremental Learning (PBIL). PBIL explicitly maintains the statistics contained in a GA's population, without the use of "genetic" operators. PBIL is simpler than a GA, both theoretically and computationally. Empirical results show that PBIL is faster and more effective than standard GAs on a large set of benchmark problems. PBIL has been used for discrete point data selection for object localization - for use in computer aided surgery, the design of low-level vision controllers for autonomous navigation, and the design of high-level reactive controllers for robot vehicles.

HOST: Prof. Eric Grimson