SFDATrack Generalized Source-Free Domain Adaptive Tracking Under Adverse Weather Conditions

Jul 16, 2026

Upload by Finn

Domain adaptive visual object tracking under adverse weather targets robust target trajectory estimation despite severe appearance shifts. Existing approaches depend on large-scale labeled source video frames, which becomes impractical when source data cannot be stored, transmitted, or accessed during adaptation. SFDATrack addresses this by using only unlabeled target-domain adverse weather samples for generalized source-free adaptation. A mean-teacher backbone with Dual Interactive Mamba distills resilient tokens, and hyperspherical prototype projection aligns multi-domain prototypes to enable adaptation across diverse weather conditions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

Introduction

Problem: domain shift under adverse weather

Proposed method: SFDATrack (source-free adaptation)

Key components and contributions

Experimental setup and results

arXiv :2607 .00369v 1 [ cs .CV] 1 Jul 2026 SFDATrack: Generalized Source-Free Domain Adaptive Tracking Under Adverse Weather Conditions Siyuan Yao 1 , Ziqi Wang 1 ,2 , Ruiqi Yu3 , Junqi Huang 1 , Wenqi Ren 1†, and Xiaochun Cao 1† 1 Shenzhen Campus of Sun Yat-sen University 2 Beijing University of Posts and Telecommunications 3 Nanyang Technological University {yaosiyuan04,zqwatcherbr0,yuruiqi422,huangjq3632,[rwq.renwenqi}@gmail.com](rwq.renwenqi}@gmail.com) , [caoxiaochun@mail.sysu.edu.cn](caoxiaochun@mail.sysu.edu.cn) Abstract. Domain adaptive visual object tracking under adverse weather conditions has garnered significant attention in recent years. Despite the impressive performance, existing methods heavily rely on the largescale video frames from both source and target domains, which is impractical under rigid resource constraints where source data is unavailable. To overcome this limitation, we propose SFDATrack, a generalized source-free domain adaptive tracker that merely leverages adverse weather samples from the target domain for robust state estimation. Specifically, SFDATrack first employs a mean-teacher backbone with Dual Interactive Mamba (DIM) blocks to distill the candidate target tokens that are resilient to weather variations from classified, augmented samples. Afterwards, we introduce a hyperspherical prototype projection (HPP) module to project these tokens onto multi-domain prototypes within a latent hyperspherical space. By enforcing both domainspecific and domain-invariant properties of the multi-domain prototypes, SFDATrack can be seamlessly adapted to diverse weather conditions with powerful generalizability. Extensive experiments evaluated on various benchmarks demonstrate that SFDATrack achieves superior performance compared to state-of-the-art approaches. The code is available at [https://github.com/watcherBR0/sfdatrack](https://github.com/watcherBR0/sfdatrack). Keywords: Visual Object Tracking · Source-Free Domain Adaptation · Hyperspherical Embedding 1 Introduction Visual object tracking (VOT) is an essential computer vision task, which aims to estimate the trajectory of an arbitrary target in video sequences under the guidance of first frame initialization. It serves as a critical component for a † Corresponding authors. 2 S. Yao et al. Fig. 1: Comparison of the domain adaptive tracking frameworks under adverse weather conditions. (a) Single domain adaptive tracking [50] . (b) Multi-domain adaptive tracking [48] . (c) The proposed source-free domain adaptive tracker (SFDATrack) . SFDATrack adapts a pre-trained model to multiple target domains via latent space projection, enabling seamless adaptation to diverse weather conditions without requiring any original source data. wide range of applications, spanning from autonomous driving, video surveillance and embodied intelligence. Driven by the powerful deep learning architectures, modern trackers have shown remarkable capability to predict the target state in controlled environments. While the deep trackers have achieved impressive performance on standard benchmarks, a critical challenge persists for modern VOT, i.e. , when a tracking model trained on a curated source domain (e.g., daytime) is deployed to an unseen target domain (e.g., nighttime or fog), it often suffers severe performance degradation. Such adverse weather conditions introduce complex, non-uniform perturbations like low contrast and atmospheric noise, which severely disrupt the model’s identification ability to track the target object. To overcome this challenge, recent studies have proposed Domain Adaptive Visual Object Tracking (DAVOT) [13, 39, 50, 51, 53] as a promising solution. As shown in Fig. 1, the prevailing DAVOT methods typically follow an unsupervised single domain adaptation or multi-domain adaptation paradigms, which utilizes annotated source data and unlabeled target data to learn domain-invariant features across different wea

FAQs

What are the main components of SFDATrack?

It uses a mean-teacher backbone with Dual Interactive Mamba blocks for multi-domain knowledge distillation, followed by a hyperspherical prototype projection module to align tokens with multi-domain prototypes in a latent hyperspherical space.

Jul 16, 2026

Upload by Finn