Unsupervised Cross-domain Object Detection via Multiple Domain Randomization

Cross-domain detection refers to the challenge of detecting objects or patterns belonging to different domains or contexts. The cross-domain detection problem arises when the training data and test data do not subject to independent and identical distribution, leading to a significant decrease in the performance of existing object detection methods. In order to address the aforementioned cross-domain detection problem, this paper proposes an unsupervised cross-domain object detection method based on multi-domain randomization. Firstly, the method utilizes Cycle-GAN to generate multiple randomized domains, enabling the comprehensive learning of the target domain overall's feature distribution. Then, a domain randomization parameter callback module is devised to retain the key detection information of the object, thereby improving the model's stability. Additionally, to alleviate the problem of domain bias and inconsistency between data and labels, a source domain consistency loss is incorporated to enhance the convergence speed of the model and amplify the semantic information embedded within the features. The experimental results on multiple cross-domain datasets show that the proposed method outperforms existing unsupervised cross-domain object detection algorithms in terms of cross-domain detection performance.