A Meta-learning Approach for Domain Generalisation across Visual Modalities in Vehicle Re-identification

Recent advances in imaging technologies have enabled the usage of infrared spectrum data for computer vision tasks previously working with traditional RGB data, such as re-identification. Infrared spectrum data can provide complementary and consistent visual information in situations of low visibility such as night-time, or adverse environments. However, the main issue that prevents the training of multi-modal systems is the lack of available infrared spectrum data. To this end, it is important to create systems that can easily adapt to data of multiple modalities, at inference time. In this paper, we propose a domain generalisation approach for multi-modal vehicle re-identification based on the recent success of meta-learning training approaches, and evaluate the ability of the model to perform to unseen modality data at testing time. In our experiments we use RGB, near-infrared and thermal-infrared modalities using the RGBNT100 dataset and prove that our meta-learning training configuration can improve the generalisation ability of the trained model compared to traditional training settings.

1. Introduction

Vehicle Re-identification (ReID) is an important and challenging task in the computer vision literature for visual surveillance applications. A large number of models for ReID problem have been proposed [1], [16], [20], [26], [41], exploring various architectures, deep metric learning methods and image enhancement techniques. The vast majority of ReID approaches have been based on visible spectrum visual data, as traditional RGB sensors have been the most common ones for surveillance scenarios. However, one of the main factors preventing the applicability of such systems is their poor performance under low illumination conditions, at night time, in foggy weather or dark scenes. To this end, infrared spectrum imaging sensors -including near infrared and thermal infrared- have been recently deployed in surveillance applications. For instance, unlike RGB sensors, thermal sensors can provide consistent 24h high quality visual imagery, overcoming the aforementioned challenges.