cs.CVApr 9, 2026

Generalization Under Scrutiny: Cross-Domain Detection Progresses, Pitfalls, and Persistent Challenges

Authors: Saniya M. Deshmukh, Kailash A. Hambarde, Hugo Proença

Organizations: IT: Instituto de Telecomunicações, University of Beira Interior, Covilhã, Portugal.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Object detection models trained on a source domain often exhibit significant performance degradation when deployed in unseen target domains, due to various kinds of variations, such as sensing conditions, environments and data distributions. Hence, regardless the recent breakthrough advances in deep learning-based detection technology, cross-domain object detection (CDOD) remains a critical research area. Moreover, the existing literature remains fragmented, lacking a unified perspective on the structural challenges underlying domain shift and the effectiveness of adaptation strategies. This survey provides a comprehensive and systematic analysis of CDOD. We start upon a problem formulation that highlights the multi-stage nature of object detection under domain shift. Then, we organize the existing methods through a conceptual taxonomy that categorizes approaches based on adaptation paradigms, modeling assumptions, and pipeline components. Furthermore, we analyze how domain shift propagates across detection stages and discuss why adaptation in object detection is inherently more complex than in classification. In addition, we review commonly used datasets, evaluation protocols, and benchmarking practices. Finally, we identify the key challenges and outline promising future research directions. Cohesively, this survey aims to provide a unified framework for understanding CDOD and to guide the development of more robust detection systems.

Figures & tables

SurveyYearFormal ProblemPipelineFailureUnified Framework
Pan & Yang Pan and Yang (2009)2009✓\checkmark×\times×\times×\times
Weiss et al. Weiss et al. (2016)2016✓\checkmark×\times×\times×\times
Csurka Csurka (2017)2017✓\checkmark×\times×\times×\times
Li et al. Li et al. (2020)2020△\triangle△\triangle△\triangle×\times
Muzammul & Li Muzammul and Li (2021)2021△\triangle×\times×\times×\times
Oza et al. Oza et al. (2023)2023△\triangle△\triangle△\triangle×\times
Table 1: Comparison of representative surveys relevant to cross-domain object detection with respect to this work ( ✓\checkmark = explicit coverage, △\triangle = partial coverage, ×\times = no/unknown coverage).
Figure 1: Overview and motivation of this survey. Top: a source-trained detector degrades on the target domain under domain shift, while a cross-domain adapted detector improves target detections. Bottom: existing surveys mainly emphasize method taxonomy, benchmarks, and applications, whereas our survey focuses on four complementary pillars formal problem formulation, pipeline-centric analysis, failure-mode analysis, and a unified design space with stage-wise diagnostics.
Figure 2: Overview of the object detection pipeline highlighting three sources of misalignment: feature alignment (style-dominated features and foreground–background confusion), proposal alignment (missing or noisy proposals and class-agnostic mismatch), and proposal-to-label alignment (mis-calibrated regression and classification–localization conflict). These misalignments propagate through the pipeline, leading to poor proposals and ultimately degraded detection performance.
Figure 3: The alignment-discriminativity tension: strong alignment can erase the feature structure that detection needs, so objectives must balance alignment with task-relevant discriminativity.
Figure 4: Instance-level versus scene-level adaptation: adapting localized instances versus global scene context.
Figure 5: Progression from closed-set to open-set to universal domain shift: closed-set assumes identical class sets; open-set allows target-private classes; universal handles all scenarios including partial-set.
MethodYearStageSetParRepModLblGranRsn
DA-Faster Chen et al. (2018)2018featureUDA↔\leftrightarrow△\triangle∘\circ■\blacksquareϕ\phi∼\sim
Strong-Weak Saito et al. (2019)2019feat.+headUDA↔\leftrightarrow△\triangle∘\circ■\blacksquareϕ\phi∼\sim
Yu et al. Yu et al. (2019)2019feat.+headUDA↔\leftrightarrow△\triangle∘\circ■\blacksquareϕ\phi∼\sim
Category anchor Zhang et al. (2019)2019featureUDA↔\leftrightarrow⋄\diamond∘\circ■\blacksquareϕ\phi∼\sim
Selective Zhu et al. (2019)2019featureUDA↔\leftrightarrow△\triangle∘\circ■\blacksquare∘\circ∼\sim
Nguyen et al. Nguyen et al. (2020)2020featureUDA↔\leftrightarrow⋄\diamond∘\circ■\blacksquare∘\circ∼\sim
Table 2: Taxonomy coding of representative CDOD methods. Columns and values: Stage (feature/head/regression), Set = Setting (UDA: unsupervised domain adaptation, DG: domain generalization, s-free: source-free, few-shot, univ: universal), Par = Paradigm ( ↔\leftrightarrow : alignment, R\mathcal{R} : robustness, ϕ\phi : hybrid), Rep = Representation ( △\triangle : semantic, ⋄\diamond : geometric), Mod = Modeling ( ∘\circ : implicit, ∙\bullet : explicit), Lbl = Label Space ( ■\blacksquare : closed-set, □\square : open-set, ϕ\phi : hybrid), Gran = Granularity ( ∘\circ : instance-level, □\Box : scene-level, ϕ\phi : mixed/uncertain), and Rsn = Reasoning ( ∼\sim : correlational, →\rightarrow : causal, ϕ\phi : mixed/uncertain).
MethodSeparationRecallECEFailure
DA-Faster Chen et al. (2018)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Strong-Weak Saito et al. (2019)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Yu et al. Yu et al. (2019)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Category anchor Zhang et al. (2019)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Selective Zhu et al. (2019)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Nguyen et al. Nguyen et al. (2020)✓\checkmark×\times×\times↑ ⁣B\uparrow\!\mathcal{B}
Table 3: Method-level diagnostic matrix showing whether each method explicitly addresses separation, recall, and calibration (ECE), with the observed dominant failure mode. Symbols: ✓\checkmark = addressed, ×\times = not addressed, △\triangle = partial tendency, ↑ ⁣B\uparrow\!\mathcal{B} = bias amplification, Freg△\mathcal{F}_{\text{reg}}\triangle = regression-focused partial effect, and τ\tau -sensitive = threshold-sensitive behavior.
DatasetYearModality#Images#Cls#AnnoRoleDomain Shift
PASCAL VOC Everingham et al. (2010)2007–2012RGB∼\sim 16.5K∼\sim 20∼\sim 40KS/Tmild scene shift
MS COCO Lin and others (2014)2014RGB∼\sim 330K∼\sim 80∼\sim 2.5MSscene diversity
ImageNet DET Russakovsky et al. (2015)2013RGB∼\sim 450K∼\sim 200∼\sim 500KSfine-grained category
Cityscapes Cordts et al. (2016)2016RGB∼\sim 3.0K∼\sim 8∼\sim 65KTurban scene shift
Foggy Cityscapes Sakaridis et al. (2018)2018RGB∼\sim 3.0K∼\sim 8∼\sim 65KTweather (clear →\rightarrow fog)
SIM10K Richter et al. (2016)2018RGB (Synthetic)∼\sim 10K∼\sim 1∼\sim 58KSsynth →\rightarrow real
Table 4: Common datasets used in CDOD benchmarks, summarizing modality, scale, annotation volume, typical role, and dominant shift type. Acronyms: S = Source, T = Target. Symbol: ∼\sim indicates approximate counts.
Figure 6: Adversarial domain adaptation for CDOD: a domain discriminator drives feature alignment between source and target.
Figure 7: Pseudo-label self-training in CDOD: a teacher generates target pseudo-labels to supervise a student, with risk of confirmation bias.
Figure 8: Distillation for CDOD: transferring teacher knowledge for example, foundation-model or EMA teacher signals to improve target robustness.
More in this topic

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Explore similar work

CardsList
Mar 17, 2026

CD-FKD: Cross-Domain Feature Knowledge Distillation for Robust Single-Domain Generalization in Object Detection

Single-domain generalization is essential for object detection, particularly when training models on a single source domain and evaluating them on unseen target domains. Domain shifts, such as changes in weather, lighting, or scene conditions, pose significant challenges to the generalization ability of existing models. To address this, we propose Cross-Domain Feature Knowledge Distillation (CD-FKD), which enhances the generalization capability of the student network by leveraging both global and instance-wise feature distillation. The proposed method uses diversified data through downscaling and corruption to train the student network, whereas the teacher network receives the original source domain data. The student network mimics the features of the teacher through both global and instance-wise distillation, enabling it to extract object-centric features effectively, even for objects that are difficult to detect owing to corruption. Extensive experiments on challenging scenes demonstrate that CD-FKD outperforms state-of-the-art methods in both target domain generalization and source domain performance, validating its effectiveness in improving object detection robustness to domain shifts. This approach is valuable in real-world applications, like autonomous driving and surveillance, where robust object detection in diverse environments is crucial.

Junseok Lee, Sungho Shin, Seongju Lee +1

Junseok Lee is with the Advanced Robotics Lab, LG Electronics, Seoul, Republic of Korea. · Sungho Shin is with the Place AI Flatform, Naver, Gyeonggi-do, Republic of Korea · Seongju Lee and Kyoobin Lee are with the Department of AI Convergence, Gwangju Institute of Science and Technology (GIST), Gwangju, Republic of Korea.

Object DetectionCross-Domain GeneralizationFeature-Based Knowledge DistillationKnowledge Distillation
AskPDF
Sep 2, 2026

Domain generalization and synthetic data in object detection: the enabler, the probe, and the gap

Object detection models often experience performance degradation when deployed under distribution shifts, caused by for example changes in weather type, operational environment, or object appearance. Domain Generalization (DG) aims to develop models that remain robust under such shifts and generalize well to unseen domains. DG research specifically focused on object detection models is scarce, although these models face additional challenges around localization and multi-scale representations. Synthetic data is a promising tool to support in DG, by enabling large-scale generation of diverse new samples. In this paper, we present an object detection-centric review of DG and examine the role of synthetic data from three complementary perspectives. First, synthetic data acts as an enabler of DG through diversification and alignment strategies that aim to improve robustness to distribution shifts. Second, it serves as a probe that enables controlled experimentation to identify and understand failure modes. Third, we discuss the synthetic-to-real gap, a particularly challenging form of domain shift that arises when models trained on synthetic imagery are deployed on real-world data. Through reviewing these perspectives, we identify limitations of current DG approaches for object detection and argue that future research requires representation-aware methods that explicitly address both localization and classification under domain shift.

Elfi I. S. Hofmeijer, Ella P. Fokkinga, Friso G. Heslinga +2

TNO - Defence, Security and Safety, the Hague, the Netherlands · FOI - Swedish Defence Research Agency, Linköping, Sweden

Cross-Domain GeneralizationDistribution Shift RobustnessObject DetectionSynthetic Data AugmentationSynthetic-to-Real Domain Adaptation
AskPDF
Mar 12, 2026

ABRA: Teleporting Fine-Tuned Knowledge Across Domains for Open-Vocabulary Object Detection

Although recent Open-Vocabulary Object Detection architectures, such as Grounding DINO, demonstrate strong zero-shot capabilities, their performance degrades significantly under domain shifts. Moreover, many domains of practical interest, such as nighttime or foggy scenes, lack large annotated datasets, preventing direct fine-tuning. In this paper, we introduce Aligned Basis Relocation for Adaptation(ABRA), a method that transfers class-specific detection knowledge from a labeled source domain to a target domain where no training images containing these classes are accessible. ABRA formulates this adaptation as a geometric transport problem in the weight space of a pretrained detector, aligning source and target domain experts to transport class-specific knowledge. Extensive experiments across challenging domain shifts demonstrate that ABRA successfully teleports class-level specialization under multiple adverse conditions. Our code will be made public upon acceptance.

Mattia Bernardi, Chiara Cappellino, Matteo Mosconi +3

University of Modena and Reggio Emilia

Open-Vocabulary Object DetectionVision Foundation Model Adaptation
AskPDF