Attention Fusion Mechanism for Domain Adaptive Object Detection

Its purpose is to alleviate performance degradation caused by domain-shift. However, most previous methods in the design of domain classifiers is often too simple. These methods input different scale feature maps from the network to independent domain classifiers, which do not effectively utilize the relationships between feature maps. Based on the shortcomings of the above methods, we designed a new attention network for adaptive object detection. Our method proposes an attention-based fusion domain classifier. This classifier inputs multi-scale feature maps and utilizes an attention mechanism to generate an attention map that fuses deep-layer feature maps with shallow-layer feature maps, thereby enhancing the domain classifier's discriminative ability. In this way, the network can obtain different levels of global structure representation and local texture mode. We test the target detection tasks on different challenging datasets. The experimental results prove the effectiveness of the method.