Learning Attentive Deep Representations for Object Re-Identification and Beyond

Object re-identification is a common task in computer vision, a certain object's (person, wildlife, vehicle, etc.) image of interest is used to match against a large gallery of images to determine if the object has already been observed by a network of cameras with non-overlapping viewpoints at different time and location. Conventionally, researchers in the computer vision community have proposed various hand-crafted features for the re-identification problem. Despite all these research efforts, re-identification remains a challenging problem due to the fact that an object's appearance can change significantly when uncontrolled variations in illumination, occlusion, resolution, pose, view angle, and background clutter are involved. As an instance of object re-identification, person re-identification has drawn increasing interest due to its wide applications. In this dissertation, we explore general mechanisms to model data relationship for the problem.

The ability for wild animal researchers to re-identify an animal identity upon re-encounter is crucial for wildlife conservation. The accurate re-identification of individual animal identities requires a considerable amount of training time for them to obtain expert skills. To expedite the process, we leverage the advancement in person re-identification research to develop a novel computer vision and cloud-based serverless wildlife identification framework and build a painted dog identification system based on the framework.

However, the designed attention mechanisms are designed from the root of solving object re-identification problems, they are also deep learning mechanisms with the capability of solving other problems. The attention mechanisms are versatile for solving a multi-label text classification problem, which takes advantage of the attention mechanism's capability of modeling pair-wise relationships to model the text representations and label representations.

Overall, this dissertation is motivated by wildlife conservation to address the object re-identification problem. The goal of the research is to advance the state-of-the-art object re-identification research, which not only has an impact on harnessing the power of AI for better wildlife conservation, but also has broader impacts on other machine learning problems beyond object re-identification.