Test-time adaptation for object detection via Dynamic Dual Teaching

Abstract

Test-Time Adaptation (TTA) is a practical setting in real-world applications, which aims to adapt a source-trained model to target domains with online unlabeled test data streams. Current approaches often rely on self-training, utilizing supervision signals from the source-trained model, suffering from poor adaptation due to diverse domain shifts. In this paper, we propose a novel test-time adaptation method for object detection guided by dual teachers, termed Dynamic Dual Teaching (DDT). Inspired by the generalization potentials of Vision-Language Models (VLMs), we integrate the VLM as an additional language-driven instructor. This integration exploits the domain-robustness of language prompts to mitigate domain shifts, collaborating with the teacher of source information within the teacher–student framework. Firstly, we utilize an ensemble prompt to guide the prediction process of the language-driven instructor. Secondly, a dynamic fusion strategy of the dual teachers is designed to generate high-quality pseudo-labels for student learning. Moreover, we incorporate a dual prediction consistency regularization to further mitigate the sensitivity of the adapted detector to domain shifts. Experiments on diverse domain adaptation benchmarks demonstrate that the proposed DDT method achieves state-of-the-art performance on both online and offline domain adaptation settings.