Title: Efficient Nearest Neighbor Recovery from Large Deep Descriptor Databases
Student: Rahman, M M Mahabubur toufik@txstate.edu
Advisor: Jelena Tešić
Date and Time: Thu Dec 8 2022 3:00 PM
Location: https://txstate.zoom.us/my/jtesic
Abstract: How do we discover if an object (e.g. vehicle) appeared in petabytes of video streams and archives? We use nearest neighbor search in the deep descriptor feature space. The k-nearest neighbors (k-NN) search identifies the top k nearest neighbors to the query and performs very well for retrieving exact solutions in smaller data sets with a lower dimension. However, the k-NN search can be sluggish in large datasets and higher dimensions because of the "curse of dimensionality". To address this problem, several approximate nearest-neighbor (ANN) methods were introduced that work very fast but sacrifice some accuracy by loosening the condition of exact nearest-neighbor retrieval. In this study, we present two novel methods for efficient and effective retrieval of similar deep descriptors. Both methods retrieve truly similar items in the database, even if the retrieval set is large, load items that are truly close to the incoming query at retrieval time, and outperform state-of-the-art methods in terms of precision and recall at depths of up to 100. The proposed methods offer consistent index loading and retrieval performance for a large set of deep descriptor databases: a crucial condition for the unknown class discovery task in large image archives and streams.
Deadline: Dec. 23, 2022, midnight