Structural consistency learning for unsupervised domain adaptive object detection

Abstract

Unsupervised domain adaptive object detection aims to facilitate the transfer of trained object detection models from the source domain to an unlabeled target domain. Although existing methods have made strides in feature alignment through adversarial learning, they tend to ignore the issue of category imbalance, leading to inadequate generalization of the model for rare categories. In addition, they fail to adequately address the background information embedded in the features, limiting the extraction of crucial object features. In order to overcome these limitations, this work proposes a structural consistency learning framework for unsupervised domain adaptive object detection. The framework enhances foreground feature representation through an Enhanced Dual Attentional Feature Alignment (EFA) mechanism and accomplishes comprehensive cross-domain feature alignment through the Structural Feature Consistency Module (SFC). The EFA introduces an attention mechanism in the image-level and instance-level feature alignment phases, enhancing the recognition of foreground objects. The SFC integrates information from multiple batches to obtain global prototypes and constructs a structure matrix based on the distances between these global prototypes. This process comprehensively reduces the structural differences between the source and target domains. The effectiveness of the approach has been validated through comprehensive experimentation on multiple cross-domain object detection benchmark datasets. The method achieves significant performance gains over current state-of-the-art techniques.