Using Synthetic Data for Deep Person Re-Identification
Ongoing Project
Abstract Person re-identification is one of the major research topics in computer vision that has seen significant progress, with the advent of deep learning. The automation of this task is of utmost importance for visual surveillance systems, where the goal is to simply identify a person in a given image in a large gallery of images captured across multiple cameras. The rising interest in this topic is not merely because it poses a challenging real-world problem but also due to the introduction of, as per the aforementioned trend, much larger person re-identification benchmark datasets lately. While the progress has been extremely significant, the available datasets in the literature are still not of the desired quality in terms of size and variety for effectively training deep models.

The aim of the proposed project is to produce new large-scale datasets which will allow construction of more powerful deep models for person re-identification and are much larger and more comprehensive than the existing ones. For this purpose, instead of manually labeling images taken from cameras placed at different locations under specific scenarios, in our project, we will investigate novel synthetic data generation methods.

Sponsor: The Scientific and Technological Research Council of Turkey (TUBITAK) The Support Program for Scientific and Technological Research Projects (Award# 217E029)

Related Publications Language Guided Fashion Image Manipulation with Feature-wise Transformations
ECCV 2018 Workshop on Computer Vision in Art, Fashion and Design
Mehmet Gunel, Erkut Erdem, Aykut Erdem