Master Thesis Defense: Haotao Lai

Speaker: Haotao Lai

Supervisors: Drs. S. Mokhov, J. Paquet

Examining Committee:
Drs. N. Shiri, J. Yang, A. Hanna (Chair)

Title: An OpenISS Framework Specialization for Deep Learning-based Person Re-identification

Date:Wednesday, August 21, 2019

Time: 13:00

Place: EV 1.162

ABSTRACT

Person detection, person re-identification, each on their own, is a rapidly increasing and independent research task of computer vision. In fact, the output from person detection is naturally the input of person re-identification task, which indicates that these two tasks are highly related, even though there are a certain number of solutions for each of the individual task. But currently, there is no existing solution that can combine them to form a integrated working pipeline.

To fill the gap, we propose a highly modular and structural framework solution that provides the functionalities including not only cross-language invocation and pipeline execution mechanism but also viewer, device, tracker, detector, and recognizer abstraction. We instantiate the proposed framework to achieve our goal of tracking the same person across multiple cameras, which essentially is the combination of person detection and person re-identification. Besides the main person ReID task, we also support skeleton tracking, as well as camera calibration, image alignment and green screen image which commonly comes with a computer vision framework. We first evaluate our solution from the framework’s point of view according to the requirements and scenarios then also report the major metrics used by the research community for person detection and person re-identification tasks respectively.