Dynamic Retraining-Updating Mean Teacher for Source-Free Object Detection

TL;DR

This paper tackles source-free object detection (SFOD) by diagnosing why Mean Teacher–based self-training degrades in the absence of source data and proposing a co-evolutionary solution. The Dynamic Retraining-Updating (DRU) mechanism dynamically retrains the student and updates the teacher based on the student’s evolution, and is augmented by a Historical Student Loss to dampen the influence of incorrect pseudo labels. DRU also incorporates Masked Image Modeling to encourage contextual learning in the target domain. Across Cityscapes/Foggy Cityscapes, BDD100k, and Sim10k→Cityscapes benchmarks, DRU achieves state-of-the-art SFOD performance and approaches or surpasses several UDA methods, demonstrating improved stability and adaptability in privacy-preserving domain adaptation settings.

Abstract

In object detection, unsupervised domain adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. However, UDA's reliance on labeled source data restricts its adaptability in privacy-related scenarios. This study focuses on source-free object detection (SFOD), which adapts a source-trained detector to an unlabeled target domain without using labeled source data. Recent advancements in self-training, particularly with the Mean Teacher (MT) framework, show promise for SFOD deployment. However, the absence of source supervision significantly compromises the stability of these approaches. We identify two primary issues, (1) uncontrollable degradation of the teacher model due to inopportune updates from the student model, and (2) the student model's tendency to replicate errors from incorrect pseudo labels, leading to it being trapped in a local optimum. Both factors contribute to a detrimental circular dependency, resulting in rapid performance degradation in recent self-training frameworks. To tackle these challenges, we propose the Dynamic Retraining-Updating (DRU) mechanism, which actively manages the student training and teacher updating processes to achieve co-evolutionary training. Additionally, we introduce Historical Student Loss to mitigate the influence of incorrect pseudo labels. Our method achieves state-of-the-art performance in the SFOD setting on multiple domain adaptation benchmarks, comparable to or even surpassing advanced UDA methods. The code will be released at https://github.com/lbktrinh/DRU

Figures (7)

• Figure 1: The training curves of different Mean Teacher training strategies on the validation set of Cityscapes $\to$ Foggy Cityscapes under the SFOD setting. These varied strategies consistently show a degradation phenomenon: the teacher model gradually degrades due to inappropriate updates from the student model, while the student model experiences performance deterioration due to inaccurate pseudo labels.
• Figure 2: The comparison of the conventional Mean Teacher (MT) framework (left) and our Dynamic Retraining-Updating (DRU) method (right). Left: In MT, the teacher model is continuously updated by a fixed interval $m$ ($m=1$Mt or $m=s, (s>1)$A2sfodIrg). Right: In DRU, the student model is dynamically retrained and the teacher model is dynamically updated based on prediction feedback. Additionally, the current student model is further supervised by the historical student model.
• Figure 3: Overview of our Dynamic Retraining-Updating (DRU) method, which is built upon the Mean Teacher framework. DRU employs dynamic retraining of the student model and dynamic updating of the teacher model based on the student's evolution assessment. The student model is further supervised by Historical Student Loss $\mathcal{L}_{his}$. The details of the training process will be presented in \ref{['alg:dru']}
• Figure 4: (a), (b), (c) The training curves for adding modules to MT on Cityscapes $\to$ Foggy Cityscapes. (d) The training curves of our method on Sim10k $\to$ Cityscapes.
• Figure 5: The training curves of student and teacher models with Dynamic Retraining-Updating mechanism on Cityscapes $\to$ Foggy Cityscapes