CSANet: Cross-Semantic Attention Network for Open-Set Object Recognition

With the increase of real-world scenarios such as robotics, urban rescue and autonomous driving, deep learning models are increasingly exposed to open-set scenarios where established methods should separate the known and unknown categories in the real world. However, most existing open-set recognition methods treat all features equally and focus on learning features that facilitate the discrimination of categories during the training, which is detrimental to the performance of models in the open world. In response to this challenge, we propose a novel framework based on a Cross-Semantic Attention Network (i.e., CSANet) to guide the model to explore more comprehensive features. In detail, we apply cross-semantic attention to guide both the high-level semantic features and a set of learnable category prototypes, which encourages the model to better characterise known categories and facilitates its ability to discriminate unknown categories in the open world. In addition, we develop a combined loss that widens the inter-category distance and narrows the intra-category distance, thus reserving the unknown categories a larger position in the feature space. Experimental results on several popular open-set recognition datasets demonstrate the effectiveness and efficiency of our method.