Vehicle Reidentification in Thermal Infrared Spectrum

In this work we address the problem of reidentifying vehicles moving on a highway in thermal infrared spectrum. The vehicle images are collected at the different locations on the highway, and the collection timestamp is served as the weak supervisory signal to allow the training of the vehicle reidentification network. Combined with the vehicle detection network trained on RGB images, the system allows fully automatic training and testing of the reidentification network in thermal infrared spectrum. The experiments confirm the advantage of the proposed method over using identical reidentification network trained on RGB images. This approach can be used for the tasks where the labelled training data is sparse or unavailable, as in the examples of thermal infrared or multispectral infrared sensors.