A semantic consistent object detection model for domain adaptation based on mixed-class distribution metrics

Abstract

Unsupervised domain adaptation is crucial for mitigating the performance degradation caused by domain bias in object detection tasks. In previous studies, the focus has been on pixel-level and instance-level shift alignment to minimize domain discrepancy. However, it is important to note that this method may inadvertently align single-class instance features with mixed-class instance features that belong to multiple categories within the same image during image-level domain adaptation. This challenge arises because each image in object detection tasks contains objects of multiple categories. To achieve the same category feature alignment between single-class and mixed-class, our method considers features with different mixed categories as a new class and proposes a mixed-classes $H$-divergence to reduce domain bias for object detection. To enhance both single-class and mixed-class semantic information, and to achieve semantic separation for the mixed-classes in $H$-divergence, we employ Semantic Prediction Models (SPM) and Semantic Bridging Components (SBC). Furthermore, we reweigh the loss of the pixel domain discriminator based on the SPM results to reduce sample imbalance. Our extensive experiments on widely used datasets illustrate how our method can robustly improve object detection in domain bias settings.