[Defense] AI-Driven Vehicle Re-Identification: Decoding Effective Features through Ablation Studies
Friday, April 18, 2025
1:30 pm - 3:00 pm
In Partial Fulfillment of the Requirements for the Degree of Master of Science
Shireesh Kumar Poral Ashok Kumar
will defend his thesis
AI-Driven Vehicle Re-Identification: Decoding Effective Features through Ablation Studies
Abstract
In today’s increasingly urbanized world, effective traffic monitoring is essential to improve road safety, manage congestion, and support law enforcement. One of the critical challenges in traffic monitoring is vehicle re-identification (ReID), which involves recognizing the same vehicle at different locations, from varying viewpoints, and under diverse environmental conditions. Human beings can effortlessly identify vehicles based on unique features such as shape, color, and distinctive markings. However, translating this ability into an automated system poses significant challenges due to variations in vehicle appearance caused by changes in lighting, angle, and background. This research explores the potential of both traditional Convolutional Neural Networks (CNNs) and modern Transformer-based models for vehicle re-identification by conducting an ablation study on various architectural modifications. Specifically, we experiment with ResNet-50 as a representative CNN model and ViT-Base (16-patch) as a Transformer-based model to evaluate how these modifications affect their ability to learn crucial factors necessary for re-identifying the reoccurrence of the same vehicle under different conditions. By analyzing the results obtained from different modifications, we assess their impact on the overall performance of vehicle re-identification. This study contributes to the development of more robust vehicle re-identification systems, which can enhance the efficiency of automated traffic monitoring and security applications.
Friday, April 18, 2025
1:30 PM - 3:00 PM
PGH 550 and MS Teams (Meeting ID: 246 961 642 723 Passcode: bB9yV7mD)
Dr. Shishir Shah, thesis advisor
Faculty, students, and the general public are invited.