Frequency-based pseudo-domain generation for domain generalizable object detection

Siqi Zhang, Lu Zhang, Zhiyong Liu

Abstract: Highlights•In this paper, we propose a simple yet effective framework for domain generalizable object detection from a novel pseudo-domain generation perspective, including two stages: distribution diversification and domain-invariant feature learning.•In the distribution diversification stage, we present a Frequency-based Pseudo-domain Generator to construct a pseudo domain by excavating latent style information and enhancing semantic information in frequency space, which provides diverse training distributions to facilitate domain-invariant feature learning in the following stage.•In the domain-invariant feature learning stage, we introduce Rotation Prediction and Semantic Consistency learning, including an auxiliary self-supervised task rotation prediction that encourages generalized feature learning and a semantic consistency loss that enforces the feature extractor to overlook style differences among domains.•We conduct extensive experiments on various object detection benchmarks, and our method outperforms state-of-the-art approaches in both single-source and multi-source settings.