Independent Study Presentation

TITLE:  Object-level Matching In Large Video Streams Using Deep Features

PRESENTER:  George Strauch

ADVISOR:  Dr. Jelena Tesic

DATE/TIME:  Thursday, December 2nd @5:00 p.m.

LOCATION:  Comal 310


Deep feature extraction using convolutional neural networks (CNNs) provides an effective method of similarity search for objects in petabytes of video feeds.  We have implemented and evaluated a pipeline to extract features from multiple datasets using 2 state-of-art deep architectures and modeling approaches (out-of-box model and specifically trained model for the domain. We have evaluated the performance of the similarity search in deep feature vector space on large computer vision benchmarks:  DOTA and VisDrone; and implemented a feature extraction, indexing, and search processing pipeline that can be easily extended for different deep frameworks, features (object, track, re-identification), and indexing methods.


Deadline: Dec. 3, 2021, midnight