In recent person re-identification (Re-ID) approaches, combining global and local appearance-based features has been shown to increase performance effectively. These types of models are often characterized by multiple branches that act as experts for specific local regions or global high-level semantic features. We argue that attention mechanisms can be useful for multi-branch Re-ID models by creating more robust representations based on the interaction of informative image features. In this paper, we investigate this idea and propose a novel multi-branch architecture with experts that learn distinct representations based on (i) the global image appearance and (ii) the interaction between features. Unlike former methods with local experts acting on partitions that are fixed a-priori, our feature interaction expert uses a novel attention-based pooling to automatically extract semantically-rich and discriminative features from different regions of a person image. Compared with existing attention-based algorithms, our method maintains the feature interaction information separately in order to discriminate between identities. Our approach achieves state-of-the-art performance across three popular benchmarks - CUHK03, Market1501 and MSMT17. Furthermore, saliency visualizations show that appearance and interaction experts learn complementary representations that attend to multiple discriminant regions, leading to improved classification ability.
On the Importance of Appearance and Interaction Feature Representations for Person Re-identification
Murino, V
2022-01-01
Abstract
In recent person re-identification (Re-ID) approaches, combining global and local appearance-based features has been shown to increase performance effectively. These types of models are often characterized by multiple branches that act as experts for specific local regions or global high-level semantic features. We argue that attention mechanisms can be useful for multi-branch Re-ID models by creating more robust representations based on the interaction of informative image features. In this paper, we investigate this idea and propose a novel multi-branch architecture with experts that learn distinct representations based on (i) the global image appearance and (ii) the interaction between features. Unlike former methods with local experts acting on partitions that are fixed a-priori, our feature interaction expert uses a novel attention-based pooling to automatically extract semantically-rich and discriminative features from different regions of a person image. Compared with existing attention-based algorithms, our method maintains the feature interaction information separately in order to discriminate between identities. Our approach achieves state-of-the-art performance across three popular benchmarks - CUHK03, Market1501 and MSMT17. Furthermore, saliency visualizations show that appearance and interaction experts learn complementary representations that attend to multiple discriminant regions, leading to improved classification ability.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.