In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend his dissertation proposal
Consensus-based Decision Making for Improving Person Re-identification Systems
Person re-identification is a problem in video-based surveillance that deals with identifying the same person in multiple different camera views or in the same camera view at a different point in time. The problem is particularly challenging since the amount of information that can be extracted from available images and videos from a surveillance camera often suffers from a plethora of difficulties. Some of these problems include changes in illumination, viewpoint and pose, possible occlusions, and differences between indoor and outdoor settings. A number of algorithms have been proposed for solving the problem of person re-identification. The most popular approach towards solving the problem is to extract features from an image of a person (also known as the "probe image") and to compute the "distance" of these features from the features extracted from a set of gallery images in order to identify which of the gallery images is the closest match to the probe image. Not only does the process of feature extraction vary widely, but the distance computation also utilizes a number of different distance metrics. These distance metrics may be pre-existing or learned from a set of annotated person images. Consequently, the distances or similarity scores which are used to determine the gallery images closest to the probe image belong to different distributions and are of widely varying range of values.
In the current study, we propose a number of algorithms for the aggregation of similarity scores obtained from different person re-identification algorithms. The aggregation of scores or distances result in a significant improvement in person over the individual algorithms themselves. The algorithms proposed are unsupervised and do not require any tuning depending on the datasets used or on the individual algorithms used for computing similarity scores. The score aggregation algorithms assist in a Consensus-based Decision Making process that boosts the re-identification rates, as compared to the algorithms used individually. Further, the proposed algorithms treat the individual person re-identification algorithms as "black-box" methods. Or in other words, no prior knowledge of how the individual algorithms work is assumed or studied. This lends a generalization to the algorithms which allow them to be used not only for person re-identification but for any matching or ranking problem in general.
Date: Thursday, April 13, 2017
Time: 1:00 PM
Place: PGH 550
Advisor: Dr. Shishir Shah
Faculty, students, and the general public are invited.