[Defense] Unsupervised Representation Learning for Person Re-Identification in Video Surveillance
Wednesday, April 12, 2023
5:00 pm - 6:00 pm
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Khadija
Khaldi
will
defend
her
dissertation
Unsupervised
Representation
Learning
for
Person
Re-Identification
in
Video
Surveillance
Abstract
The dissertation focuses on the challenging problem of person re-identification (ReID), which has gained significant attention from the research community due to its wide range of applications in video surveillance, law enforcement, and smart city management. However, most previous work in ReID has been done in a supervised learning setting, which requires expensive data annotations. Therefore, this dissertation proposes three purely unsupervised learning methodologies to solve the ReID task, significantly contributing to the field. The first proposed method, CUPR, is a novel pure unsupervised learning approach using contrastive learning. This iterative approach learns strong high-level features from raw pixels using contrastive learning and performs clustering to generate pseudo-labels. The proposed method outperforms unsupervised and semi-supervised state-of-the-art methods on two widely used benchmark datasets, Market-1501 and DukeMTMC-reID. Despite the first proposed method鈥檚 success, person ReID remains challenging, especially when individuals wear similar clothing or certain parts of their body are occluded. To address this issue, the second proposed method leverages the effectiveness of modeling the spatiotemporal information of pedestrian video by mining the relationships between human body joints. Specifically, the framework proposes learning inter-frame and intra-frame relationships for discriminative feature learning via two GCN modules: spatial and temporal. The third proposed method addresses the challenges posed by aerial-based datasets, which have become increasingly popular in various applications, including person re-identification. Aerial-based datasets pose unique challenges, such as viewpoint variations, occlusion, and low resolution, making it challenging to extract discriminative features for person ReID. To address these challenges, the proposed method includes a generative stage, a contrastive stage, and a final clustering stage. Finally, we evaluate the proposed method on two recently released UAV-based datasets, where it outperforms state-of-the-art methods with an improvement of up to 4% in rank-1 score. Overall, this dissertation contributes to person re-identification by proposing three novel and effective unsupervised learning methodologies that address different challenges in person ReID. These proposed methods are evaluated on widely used benchmarks, and their effectiveness is demonstrated, making them promising candidates for practical applications in the real world.
Wednesday,
April
12,
2023
5:00PM
CT
Online
via
Department Chair Dr. Shishir Shah, Faculty Advisor
Faculty, students and the general public are invited.
