Specialise to Generalise: The Person Re-identification Case

Person re-identification (Re-Id) is a beneficial computer vision functionality in security-related applications based on video surveillance systems. It is a challenging cross-camera matching problem, which makes it prone to domain shift issues. To mitigate them, supervised and unsupervised domain adaptation and domain generalisation (DG) methods have been proposed. All such methods tend to favour performance improvements on the target data set at the expense of performance on the source data set(s), on which they generally deteriorate. In this work, instead, we propose an alternative method for DG that does not involve any re-training or fine-tuning of the Re-Id model and thus has no adverse effect on the performance of the source data set. It exploits Generative Adversarial Networks, trained on the source data set only with a one-vs-all mapping that simulates the target data set images, with the aim of transferring the style of the source data set into the target images. Finally, an ad hoc ranking process combines the features extracted from the original and generated images and produces the final ranked list. The proposed method can be used on top of any Re-Id model, making it a possible alternative method against domain shift and also complementary to other approaches. The considered solution is evaluated on a challenging cross-data set scenario on two benchmark data sets and a deep learning baseline for Re-Id. The obtained results demonstrate that the proposed solution improves performance, especially when the Re-Id model is specialised in the source domain.