Large-scale Tread Pattern Matching for Forensic Analysis

Abstract

We implement similarity measures for tread patterns, Multi-Channel Normalized Cross-Correlation (MCNCC)[1] and MCNCC with compact bilinear pooling[2], into an online interactive query site and investigate ways to improve unsupervised MCNCC by (a) extending an existing dataset with large-scale web-crawled images from online shopping sites and (b) introducing supervision signals with trainable transformations such as soft attentions and Spatial Transformer Networks (STN). Empirical results on the extended dataset show significant improvement of precision-recall characteristics of MCNCC + STN over MCNCC.

Evaluations

Reference

[1] Cross-domain forensic shoeprint matching
Kong, Bailey, Deva Ramanan, and Charless Fowlkes. In British Machine Vision Conference (BMVC)

[2] Compact bilinear pooling
Gao, Yang, Oscar Beijbom, Ning Zhang, and Trevor Darrell. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR)

[3] Spatial transformer networks
Jaderberg, Max, Karen Simonyan, and Andrew Zisserman. In Advances in neural information processing systems (NIPS)