Large Scale Vehicle Re-Identification by Knowledge Transfer from Simulated Data and Temporal Attention

Automated re-identification (re-id) of vehicles is the foundation of many traffic analysis applications across camera networks, e.g. vehicle tracking, counting, or traffic density and flow estimation. The re-id task is made difficult by variations in lighting, viewpoint, image quality and similar vehicle models and colors that can occur across the network. These influences can cause a high visual appearance variation for the same vehicle while different vehicles may appear near identical under similar conditions. However, with a growing number of available datasets and well crafted deep learning models, much progress has been made. In this work we summarize the results of our participation in the NVIDIA AI City Challenge 2020 for vehicle reid. We address the re-id task by relying on well-proven design choices from the closely related person re-id literature. In addition to this, we explicitly address viewpoint and occlusions variation. The former is addressed by incorporating vehicle viewpoint classification results into our matching distance. The required viewpoint classifier is trained predominantly on simulated data and we show that it can be applied to real-world imagery with minimal domain adaptation. We address occlusion by relying on temporal attention scores which emphasize video frames that contain minimal occlusion. Finally, we further boost re-id accuracy by applying video-based re-ranking and an ensemble of complementary models. Our models, code, and simulated data is available at https://github.com/corner100/2020-aicitychallenge-IOSB-VeRi.