This CSAFE Center Wide webinar was presented on April 27, 2018 by Dr. Charless Fowlkes, CSAFE researcher and associate professor of computer science at University of California, Irvine.
We investigate the problem of automatically determining shoe outsole class characteristics from crime scene impression evidence using computer vision and machine learning techniques. This problem can be formulated as an image retrieval task: given a photo of crime scene evidence, return a ranked list of matching candidates from a database of reference prints. I will describe our approach to automatically extracting tread pattern features using convolutional neural nets and discuss how these features can be robustly compared across images using normalized correlation measures. This framework can be tuned automatically from training data and currently produces state-of-the-art matching performance on benchmark evaluations. Finally, I will discuss some of the challenges in assembling and maintaining a comprehensive database of reference tread patterns.