Camera Re-ID

In machine vision, re-identification is essential for tracking persons or moving objects over time within a single camera or across multiple cameras with overlapping or non-overlapping field-of-views. Person re-identification is the task of associating the same person’s images across different non-overlapping or overlapping cameras or from the same camera. A person should be associated with a unique global ID no matter what camera captures this person.

Surveillance and supervision service are deployed in enterprise and industrial environments. Accurate re-identification of people and objects is a central capacity in various people and object tracking & localization solutions and Industry 4.0 solutions.

Real-time re-identification is a challenging task due to large variations in image styles and contents such as strong illumination changes across cameras and over time, different viewpoints, occlusions, variations in color responses across cameras, unconstrained person poses, different persons in similar clothes, the same person changing clothes or showing different colors/patterns from front and back views, background clutters (e.g., partial body of another person in bounding box), low resolution for small-sized or faraway persons, reflections on door or window glasses, unreliable detections (bounding box not accurate enough). Because of these challenges, re-ID models suffer from significant domain shift that causes poor generalization of the models when deployed to unseen environments.

The goal of this work is to develop domain-generalizable re-ID models. These re-ID models are trained on available datasets and can be deployed to unseen camera systems without re-training to deliver acceptable Re-ID performance in real-time object tracking & localization solutions. In this work, network architectures and loss functions are designed for training re-ID models on existing re-ID datasets as well as other large-scale datasets to improve the models’ domain generalizability.