Towards Zero-Shot Alignment and Retrieval for Forensic Detection

Vidhi Jain, Adithya Pratapa, Amrith Setlur

Abstract: In a crime scene, shoeprints are considered as an important evidence. The probe shoeprint (query) needs to be matched against reference images from a large database of shoe impressions. This is a computationally challenging task as the probe image often contains partial patterns and is heavily induced with noise. In this paper, we investigate approaches for image retrieval with very limited labeled data. We propose our baselines and discuss various approaches including synthetic data generation for data augmentation. Particularly, we describe and compare different methods that (1) leverage class hierarchies (2) learn a distance metric (3) adaptively perform canonical correlation analysis (CCA) (4) learn to ignore noisy patterns through supervised alignments.