Source-free Domain Adaptive Object Detection in Remote Sensing Images

TL;DR

The paper tackles the challenge of domain shift in remote sensing object detection when source-domain data cannot be accessed during adaptation. It introduces a source-free domain adaptive object detection (SFOD) framework that generates a perturbed target domain via Multilevel Mixed Sample Perturbation (MSP) and Adversarial Feature Style Perturbation (AFSP), and aligns detector behavior through a Mean-Teacher setup with Prototype-based Feature Distillation (PFD). Empirical results across synthetic-to-real and cross-sensor RS benchmarks show significant improvements over existing SFOD methods and competitive performance relative to UDAOD baselines, with demonstrated transferability to natural-scene datasets. The work offers practical, privacy-friendly guidance for RS object detection and yields techniques potentially applicable to other vision domains plagued by data-access constraints.

Abstract

Recent studies have used unsupervised domain adaptive object detection (UDAOD) methods to bridge the domain gap in remote sensing (RS) images. However, UDAOD methods typically assume that the source domain data can be accessed during the domain adaptation process. This setting is often impractical in the real world due to RS data privacy and transmission difficulty. To address this challenge, we propose a practical source-free object detection (SFOD) setting for RS images, which aims to perform target domain adaptation using only the source pre-trained model. We propose a new SFOD method for RS images consisting of two parts: perturbed domain generation and alignment. The proposed multilevel perturbation constructs the perturbed domain in a simple yet efficient form by perturbing the domain-variant features at the image level and feature level according to the color and style bias. The proposed multilevel alignment calculates feature and label consistency between the perturbed domain and the target domain across the teacher-student network, and introduces the distillation of feature prototype to mitigate the noise of pseudo-labels. By requiring the detector to be consistent in the perturbed domain and the target domain, the detector is forced to focus on domaininvariant features. Extensive results of three synthetic-to-real experiments and three cross-sensor experiments have validated the effectiveness of our method which does not require access to source domain RS images. Furthermore, experiments on computer vision datasets show that our method can be extended to other fields as well. Our code will be available at: https://weixliu.github.io/ .

Figures (11)

• Figure 1: Examples of synthetic-to-real adaptation (top) and cross-sensor adaptation (bottom).
• Figure 2: Domain differences in channel statistics of low-level features. For the source domain UCAS-AOD and the target domain xView, we calculate the channel average at the first stage of the backbone. We show the first 64 (out of 256) channels for better visualization.
• Figure 3: Adaptation settings for UDAOD (top) and SFOD (bottom). During adaptation on the target domain, the SFOD setting uses a source pre-trained detector but does not have access to the data in the source domain.
• Figure 4: The framework of our proposed SFOD method. Our method can be divided into two parts: perturbed target domain generation and alignment. The image-level mixed-based sample perturbation (MSP) module and the feature-level adversarial feature style perturbation (AFSP) module aim to obtain meaningful perturbation in the target domain. To better align the behavior of detector between the target domain and the perturbed target domain, prototype-based feature distillation (PFD) is adopted at the feature level, and pseudo labeling is used at the label level.
• Figure 5: The schema of AFSP module.