---
title: 'CD-FSOD: Cross-Domain Few-Shot Detection'
url: https://www.emergentmind.com/topics/cross-domain-few-shot-object-detection-cd-fsod
type: topic
---

# CD-FSOD: Cross-Domain Few-Shot Detection

Cross-Domain Few-Shot Object Detection (CD-FSOD) studies object detection when a detector must be adapted from a source or base domain to a target domain that differs in visual distribution and offers only a few labeled target instances per class. A common formulation writes a source dataset $\mathcal{D}_S=\{I,y\},\, y\in\mathcal{C}_S$ and a target dataset $\mathcal{D}_T=\{I,y\},\, y\in\mathcal{C}_T$, with $\mathcal{C}_S\cap\mathcal{C}_T=\emptyset$ and $\mathcal{P}_S\neq\mathcal{P}_T$ [2402.03094]. In the benchmark setting that established the area, a detector is trained on a base dataset such as MS COCO and then adapted to target domains with $K\in\{1,5,10\}$ annotated instances per class [2210.05311]. The problem sits at the intersection of few-shot transfer, domain adaptation, and open-set or open-vocabulary detection, but the conjunction of label scarcity and domain shift produces its own characteristic bottlenecks: overfitting, weak foreground-background separation, feature confusion, and unstable fine-tuning [2210.05311][2505.00938][2603.28182].

## 1. Formal setting and neighboring problem formulations

In its canonical form, CD-FSOD assumes abundant source-domain supervision and extremely limited target-domain supervision. The detector is first trained on the source domain and then adapted to a target dataset in which each class has only $K$ labeled instances, typically for $K=1,5,10$ [2210.05311]. Later challenge formulations make the support/query protocol explicit: the target support set contains $K$ labeled instances per class, and the remaining unlabeled examples form the query set used for evaluation [2504.10685]. An important protocol detail from the NTIRE challenge reports is that “$K$-shot” means $K$ labeled instances, not $K$ fully annotated images; support images may therefore contain additional unlabeled target instances, a fact that later became methodologically consequential [2504.10685].

CD-FSOD differs from standard few-shot object detection because the scarcity of support data is compounded by a genuine distribution shift. It also differs from classical domain-adaptive detection because the target supervision is few-shot rather than abundant or predominantly unlabeled. A stricter adjacent formulation appears in “Few-Shot Object Detection in Unseen Domains,” which treats the setting as zero-shot domain adaptation or domain generalization for FSOD: source-domain few-shot support for novel classes is available, but target-domain images or labels of those novel classes are unavailable during training [2204.05072]. This stricter setting is not identical to the now-standard CD-FSOD benchmark, but it clarifies a boundary case in which the domain gap must be handled without any target-domain novel-class supervision.

The field also contains multimodal variants. “Cross-domain Multi-modal Few-shot Object Detection via Rich Text” extends the dataset from $\mathcal{D}_{\rm FSOD}=\{(x,y)_i\}$ to $\mathcal{D}_{\text{MM-FSOD}}=\{(x,\ell,y)_i\}$, where $\ell$ denotes text associated with object instances or categories, and frames the target problem as CDMM-FSOD [2403.16188]. This suggests that CD-FSOD is increasingly understood not only as an image-only transfer problem but also as a setting in which side information, especially language, can act as an auxiliary adaptation signal.

## 2. Benchmarks, datasets, and evaluation regimes

The first benchmark paper organized CD-FSOD around MS COCO as source data and three target domains: ArTaxOr, UODD, and DIOR. It reports ArTaxOr as a biology dataset with 7 classes, 13,991 train images, and 1,383 test images; UODD as an underwater dataset with 3 classes, 3,194 train images, and 506 test images; and DIOR as an aerial dataset with 20 classes, 18,463 train images, and 5,000 test images [2210.05311]. A later benchmark expansion used six target domains—ArTaxOr, Clipart1k, DIOR, DeepFish, NEU-DET, and UODD—and attached to each domain qualitative style labels and quantitative or semi-quantitative indicators of inter-class variance and indefinable boundaries [2402.03094].

| Setting | Domains / datasets | Notes |
|---|---|---|
| Initial CD-FSOD benchmark | ArTaxOr, UODD, DIOR | COCO source; 1/5/10-shot; mAP [2210.05311] |
| Expanded six-domain benchmark | ArTaxOr, Clipart1k, DIOR, DeepFish, NEU-DET, UODD | style, ICV, and IB analysis [2402.03094] |
| NTIRE 2025 challenge | Development on six domains above; final unseen DeepFruits, CARPK, CarDD | open-source and closed-source tracks [2504.10685] |
| NTIRE 2026 challenge | second public challenge under open-source and closed-source tracks | 128 registered participants and 696 submissions [2604.11998] |

Evaluation metrics are not entirely uniform across the literature. The benchmark paper reports mAP [2210.05311]. CD-ViTO’s benchmark tables also report mAP on the six target domains [2402.03094]. CDFormer explicitly reports **mAP@0.5:0.95** [2505.00938], while the NTIRE reports use mAP as the ranking metric and weight 1-shot performance more strongly in the final challenge score [2504.10685]. This implies that reported numbers are most reliable when interpreted within the protocol of each paper rather than as unqualified cross-paper absolutes.

The challenge reports also enlarged the benchmarking function of CD-FSOD. The 2025 NTIRE challenge introduced open-source and closed-source tracks, used the six benchmark domains for development, and evaluated on three unseen final datasets—DeepFruits, CARPK, and CarDD—under 1-shot, 5-shot, and 10-shot conditions [2504.10685]. The 2026 NTIRE challenge continued the two-track structure and reported 128 registered participants, 696 submissions, 31 active teams, and 19 valid final results, underscoring the topic’s transition from a small benchmark niche to a public evaluation ecosystem [2604.11998].

## 3. Diagnosed failure modes and analytic frameworks

The earliest benchmark results established that same-domain FSOD intuitions transfer poorly to CD-FSOD. On the initial benchmark, tested meta-learning methods such as A-RPN, Meta-RCNN, and H-GCN underperformed or only marginally matched a naive Faster R-CNN fine-tuning baseline, while freezing parameters consistently hurt performance [2210.05311]. The benchmark paper explicitly attributed this to overfitting and to the inability of source-domain episodic training to mimic distant target domains [2210.05311]. This remains a foundational claim in the area: the central difficulty is not only data scarcity, but the interaction between scarcity and distributional mismatch.

CD-ViTO reframed the domain gap through three axes: style, inter-class variance (ICV), and indefinable boundaries (IB) [2402.03094]. ICV is computed from CLIP text features via
$$
S=\mathcal{F}\cdot \mathcal{F}^T,\qquad ICV=\frac{1}{N^2 * D}\sum_i\sum_j S_{ij},
$$
and was reported as small for ArTaxOr and UODD, medium for DIOR, and large for Clipart1k and NEU-DET [2402.03094]. IB was estimated through a user study and summarized as
$$
IB = w_{sli}p_{sli} + w_{mod}p_{mod} + w_{sig}p_{sig},
$$
with ArTaxOr, Clipart1k, and DIOR labeled slight, DeepFish moderate, and NEU-DET and UODD significant [2402.03094]. In that analysis, style alone did not explain the full degradation; small ICV and especially large IB were more predictive of poor transfer [2402.03094].

Later work moved from dataset-level diagnostics to representation-level failure modes. CDFormer argues that the decisive problem is **feature confusion**, divided into object-background confusion and object-object confusion, and introduces separate modules for each [2505.00938]. A still more recent analysis identifies a “target-domain Astigmatism” phenomenon in transformer detectors: attention becomes dispersed and unfocused in the target domain, producing imprecise localization and redundant predictions, while regular fine-tuning only partially remedies this effect [2603.18541]. A plausible implication is that CD-FSOD has progressed from being viewed mainly as a sample-efficiency problem to being studied as a problem of **representation geometry, attention allocation, and background modeling under shift**.

Fine-tuning behavior itself has become an analytical object. “A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps” explicitly identifies unstable optimization and weak generalization during adaptation as primary bottlenecks and evaluates robustness on an OOD-contaminated mixed-domain test set built from CD-FSOD [2603.28182]. This line of work suggests that CD-FSOD performance is sensitive not only to what representations are used, but also to how adaptation is staged and regularized.

## 4. Methodological development

The earliest methods in and around CD-FSOD were shaped by augmentation, selective adaptation, and regularization. “Few-Shot Object Detection in Unseen Domains” proposed a two-fold strategy: learn domain shift on base classes through meta-training and augment few-shot novel examples with color jittering, Gaussian blur, Gaussian noise, background replacement, and feature-level sampling, plus a contrastive foreground-class embedding loss to reduce background bias [2204.05072]. AcroFOD formalized an **adaptive optimization** view of augmentation: generate cross-domain candidates, score them by distance to target features, and keep only the closest fraction under a shrinkage ratio $k$; it combined this with image-level and box-level domain-aware augmentation [2209.10904]. The original CD-FSOD benchmark paper contributed an EMA teacher-student baseline with pseudo-label distillation, emphasizing overfitting control rather than complex source-task design [2210.05311].

A second phase of the literature reoriented the problem around open-set and open-vocabulary detectors. CD-ViTO showed that DE-ViT, despite strong same-domain FSOD results, generalized poorly to CD-FSOD, and addressed this with learnable instance features, instance reweighting, and a domain prompter [2402.03094]. The multimodal extension “Cross-domain Multi-modal Few-shot Object Detection via Rich Text” argued that rich textual descriptions could stabilize support-query transfer under severe domain shift; it fused support visual and text features and used a bidirectional rich text rectification loss [2403.16188]. This line of work made two claims that continued to recur: first, language can supply domain-invariant semantics; second, semantics alone are not sufficient when localization depends on target-domain appearance.

Recent work has been dominated by foundation-model adaptation. “Enhance Then Search” treats CD-FSOD as a search problem over augmentation policies and fine-tuning configurations on top of GroundingDINO, using a mixed augmentation pipeline with CachedMosaic, YOLOXHSVRandomAug, RandomFlip, CachedMixUp, RandomResize, and RandomCrop, plus validation-guided grid search [2504.04517]. Domain-RAG shifts attention to training-free generative augmentation: it preserves the foreground object, retrieves domain-aware backgrounds using CLIP and shallow ResNet statistics, synthesizes new backgrounds with FLUX-based generation, and recomposes support images, yielding a “fix the foreground, adapt the background” strategy [2506.05872]. GiPL combines iterative pseudo-label self-training on the support set with generative data augmentation using large vision-language models to enrich training samples and suppress overfitting [2605.29539]. Challenge-winning systems further amplified the role of self-training and annotation completion, often because support images in the challenge protocol were only partially annotated [2504.10685].

A parallel thread has concentrated on representation design rather than data generation. CDFormer uses a DETR-style single-stage architecture with an Object-Background Distinguishing module, driven by a learnable background token, and an Object-Object Distinguishing module with an InfoNCE objective over learned class embeddings [2505.00938]. FSOD-VFM discards fine-tuning entirely: it combines a Universal Proposal Network, SAM2 masks, DINOv2 features, and graph-based confidence diffusion to refine fragmented proposals, and reports strong training-free CD-FSOD performance [2602.03137]. LMP returns to multimodality but with a different decomposition from text-enrichment work: it builds class-level visual prototypes from support RoIs, generates hard-negative prototypes from jittered query boxes, and couples a visual-guided branch to a parallel text-guided branch on top of GroundingDINO [2602.18811]. The fine-tuning-centric line represented by the hybrid ensemble decoder argues that stronger adaptation can be obtained without elaborate synthetic pipelines, by combining a shared hierarchical decoder layer with multiple parallel branches and a progressive freeze-then-unfreeze schedule [2603.28182]. The newest attention-centric proposal frames target-domain error as a center-periphery failure and introduces Positive Pattern Refinement, Negative Context Modulation, and Textual Semantic Alignment to convert “astigmatic” attention into more focused patterns [2603.18541].

## 5. Empirical progression and challenge outcomes

The initial benchmark established a low baseline ceiling. Averaged across ArTaxOr, UODD, and DIOR, A-RPN reached 10.2 mAP, Meta-RCNN 10.6, H-GCN 10.5, FRCN-ft 11.2, and the benchmark’s EMA teacher-student method 13.8, yielding a reported average gain of 2.0 mAP over the second-best approach [2210.05311]. This result was important less for its absolute number than for its conclusion: CD-FSOD did not behave like standard FSOD, and much of the design space inherited from VOC/COCO few-shot detection was inadequate [2210.05311].

On the six-domain benchmark, CD-ViTO raised average mAP from DE-ViT’s 1.0/5.1/5.9 to 13.9/26.1/29.6 for 1/5/10-shot, respectively [2402.03094]. CDFormer then reported 26.8/37.1/40.0 average mAP and improvements of 12.9, 11.0, and 10.4 mAP over the previous best under 1/5/10-shot settings [2505.00938]. Domain-RAG, built on GroundingDINO, reported 33.6/42.7/45.4 average mAP and gains of +7.3, +2.3, and +1.5 over the corresponding GroundingDINO baseline [2506.05872]. LMP reported 34.3/44.0/46.6 average mAP on the same six-domain benchmark, improving over fine-tuned GroundingDINO from 26.3/40.4/43.9 [2602.18811]. The fine-tuning-centric hybrid ensemble decoder further reported 34.9/45.0/47.9 on CD-FSOD, together with gains on ODinW-13 and RF100-VL, indicating that careful adaptation of pretrained open-vocabulary detectors remains highly competitive [2603.28182].

Training-free performance has also improved materially. FSOD-VFM reported 25.3/31.3/31.6 nAP on CD-FSOD for 1/5/10-shot and emphasized that its 10-shot result of 31.6 AP substantially exceeded previous training-free methods reported at 21.4 AP [2602.03137]. This does not overtake the best fine-tuned open-source systems, but it demonstrates that part of the CD-FSOD difficulty can be shifted from optimization to proposal quality, foreground masking, and prototype matching [2602.03137].

The NTIRE challenges clarified larger ecosystem trends. The 2025 report states that top open-source methods were dominated by foundation-model transfer, self-training or pseudo-label completion, augmentation, and, in some cases, ensembling of GLIP and GroundingDINO variants [2504.10685]. The winning open-source system, MoveFree, combined Grounding DINO, self-training, and Mixture-of-Experts inside the decoder; AI4EarthLab’s ETS ranked second; IDCFS combined GLIP, iterative pseudo-labeling, LoRA-adapted GroundingDINO, and ensembling [2504.10685]. In the closed-source track, prototype and memory mechanisms remained central: X-Few’s Instance Feature Caching topped the track with a score of 125.90 against a CD-ViTO baseline of 91.00 [2504.10685]. This suggests a bifurcation of the field: in open-source settings, large pretrained grounding detectors and pseudo-label refinement dominate; in closed-source settings, prototype calibration, caching, and feature refinement remain the main levers.

## 6. Open questions, controversies, and future directions

One recurring controversy concerns what exactly is being measured in open-source CD-FSOD. Challenge reports explicitly note that once large foundation models are permitted, strict non-overlap between source and target classes becomes difficult to guarantee [2504.10685]. This does not invalidate the task, but it changes its interpretation: open-source CD-FSOD often becomes a problem of **few-shot specialization under domain shift** rather than a strict disjoint-class transfer experiment. A related ambiguity appears in support supervision itself: because challenge support images may contain unlabeled target objects, pseudo-label completion can improve performance partly by repairing annotation incompleteness rather than only by learning better detectors [2504.10685]. This suggests that evaluation protocols and annotation assumptions are not neutral implementation details; they actively shape which methods look best.

A second tension concerns **fine-tuning versus training-free adaptation**. The strongest averages on current six-domain benchmarks come from careful fine-tuning of large pretrained detectors [2603.28182][2602.18811], yet FSOD-VFM shows that training-free pipelines can already be competitive, especially when proposal generation, masking, and graph-based reweighting are strong [2602.03137]. This suggests that the field is not converging on a single paradigm. Instead, it is separating into optimization-heavy and optimization-light branches, each with different assumptions about compute budget, robustness, and deployment constraints.

A third unresolved issue is how best to use multimodality. Rich text improves CDMM-FSOD substantially in severe low-shot regimes, but richer text is not automatically better: on NEU-DET, LLM-generated descriptions underperformed manually designed text [2403.16188]. LMP, by contrast, argues that the missing ingredient in text-dominant open-vocabulary detectors is domain-specific visual information, especially hard negatives [2602.18811]. This suggests a broader lesson: language is valuable as a semantic prior, but CD-FSOD remains a localization problem, and localization under domain shift often depends on target-domain appearance cues that text alone does not encode.

The persistent difficulty of high-IB domains such as NEU-DET and UODD also remains unresolved [2402.03094]. Across multiple papers, these datasets continue to expose failures in background suppression, precise boundary modeling, and fine-grained support quality assessment [2402.03094][2505.00938]. Recent attention-based analyses and center-periphery refinement mechanisms imply that future progress may come from finer control of where detectors attend in target-domain images rather than from larger support sets alone [2603.18541].

The present trajectory of the field points toward several concrete research directions. One is better OOD-aware evaluation and adaptation, motivated by mixed-domain tests such as CD-Mixed [2603.28182]. Another is more principled support-set completion and pseudo-label refinement for partially annotated few-shot images [2504.10685]. A third is adaptive prototype management—selection, pruning, negative mining, and multimodal fusion—which appears across LMP, rich-text CDMM-FSOD, and closed-source challenge methods [2602.18811][2403.16188][2504.10685]. A plausible synthesis is that future CD-FSOD systems will combine robust pretrained grounding detectors, explicit target-domain prototype construction, selective pseudo-label expansion, and finer-grained attention control, while evaluation protocols continue to distinguish more clearly between closed-source scientific comparison and open-source practical transfer.

Source: https://www.emergentmind.com/topics/cross-domain-few-shot-object-detection-cd-fsod