Wasserstein Distance for Attention based cross modality Person Re-Identification

Cross-modality based person re-identification is a challenging task because of the high inter modality gap present between the RGB and the IR data. These systems have to learn how to discriminate between different identities as well as how to match two modalities. This has been handled using various distance metrics and custom loss functions. But the already used distance metrics might not perform well in certain cases.Hence Optimal Transport Theory, which uses the Wasserstein distance is used for the person Re-ID problem. This distance measures the effort of aligning different distributions. Attention mechanism is used to enhance the model and also to deal with noisy data. Multiple loss functions are used in this Wasserstein distance based cross-modality person Re-Id with attention. This combination of loss functions also helps with dealing the low resolution images. A model is proposed with attention mechanism where the Earth Mover’s distance is used as the distance metric instead of the usual euclidean distance. This has showed great improvement to the euclidean distance based models. The proposed model is able to achieve a rank 10 accuracy of 92.3%