MIT CSAIL defense : A few days of a robot

Speaker: Lijin Aryananda , MIT CSAIL
Relevant URL: http://people.csail.mit.edu/lijin

Abstract:

This thesis presents an implementation of a robotic head, Mertz, designed to explore incremental face recognition through natural interaction. We have seen many recent efforts in the path of integrating robots into the home for assisting with elder care, domestic chores, etc. In order to be effective in human-centric tasks, the robots must distinguish not only between people and inanimate objects, but also among different family members in the household. This thesis was driven by two specific limitations of the current technology. First, current automatic face recognition technology mostly explores the supervised solutions which are limited to a fixed training set and require cumbersome data collection and labelling procedures. Second, the lack of robustness and scalability to unstructured environments create a large gap between current research robots and commercial home products. The goal of this thesis is to advance toward a framework which would allow the robots to incrementally "get to know" each individual in an unsupervised way through daily interaction. In contrast to the target of a stand-alone and maximally optimized face recognition system, our approach is to develop an integrated robotic system as a step toward the ultimate end-to-end system capable of incremental individual recognition in a real human environment. Our main emphasis is to develop Mertz as a robotic creature with adequate overall robustness to be embedded in the dynamic and noisy human environment. Thus, we require the robot to operate for a few hours at a time and interact with a large number of passersby with minimal constraints at public locations. The robot then autonomously detects, tracks, and segments face images during these interactions and automatically generates a training set for its face recognition system. In this talk, we present the robot implementation and its unsupervised incremental face recognition framework. We describe an algorithm for clustering SIFT features extracted from a large set of face sequences automatically generated by the robot. We demonstrate the robot's capabilities and limitations in a series of experiments at a public lobby. In a final experiment, the robot interacted with a few hundred individuals in an eight day period and generated a training set of over a hundred thousand face images. We evaluate the clustering algorithm performance across a range of parameters on this automatically generated training data and also the Honda-UCSD video face database. Lastly, we present some recognition results using the self-labelled clusters.