Cross-domain object detection by local to global object-aware feature alignment

Cross-domain object detection has attracted more and more attention recently. It reduces the gap between the two domains, where the source domain is labeled and the target domain is label-agnostic. The feature alignment is a fundamental step for cross-domain object detection. However, the complex background information may lead to an unreliable feature alignment and even negative transfer. It has been found that paying more attention to the object instances is more conducive to the feature alignment. Therefore, this paper proposes a local to global object-aware feature alignment (OFA) method for cross-domain object detection. The proposed network consists of three components: (1) local object-aware feature alignment (LOFA) module, which utilizes the classification map to enhance the local object representation and achieve the local object feature alignment. (2) cross-scale attention feature alignment (CAFA) module, which employs the local cross-channel interactive attention to highlight the foreground features and aligns the features of global cross-scale fusion in the multi-scale level. (3) memory-category guided alignment (MCGA) module, which records all category features from the source domain, and then uses the memory-category features to guide the category learning of target domain. By the proposed OFA, the enhanced object information can lead to better feature alignment. The proposed method is demonstrated by the cross-domain detection in different scenarios and shows state-of-the-art performance.