NTIRE 2026 CD-FSOD Challenge

• NTIRE 2026 CD-FSOD Challenge is a benchmark for cross-domain few-shot object detection that adapts detectors from a richly labeled source to novel, visually dissimilar target domains.
• It employs dual tracks—closed-source and open-source—with a weighted evaluation that emphasizes the particularly challenging 1-shot regime using mean Average Precision.
• Leading methods leverage foundation models, generative augmentation, and pseudo-labeling to overcome severe domain shifts and data scarcity, outperforming naive fine-tuning baselines.

The NTIRE 2026 CD-FSOD Challenge was the second NTIRE challenge devoted to Cross-Domain Few-Shot Object Detection (CD-FSOD), a setting in which a detector must adapt from a source domain to novel target domains under severe domain shift and with only 1, 5, or 10 labeled examples per novel class. In the challenge formulation, the source classes and target classes are disjoint, the source and target distributions differ substantially, and performance is measured on unseen target domains using mean Average Precision (mAP) and a score that emphasizes the hardest 1-shot regime (Qiu et al., 13 Apr 2026). The challenge inherits its conceptual basis from the original CD-FSOD benchmark, which showed that many standard FSOD methods degrade sharply, and can even underperform a naive fine-tuning baseline, when transferred across visually dissimilar domains (Xiong, 2022). It also extends the dual-track NTIRE framework introduced in 2025, where a closed-source setting isolates adaptation under source restrictions and an open-source setting probes the upper bound attainable with foundation models, external data, and broader pretraining (Fu et al., 14 Apr 2025).

1. Benchmark lineage and problem definition

CD-FSOD is an extension of standard few-shot object detection to settings where the source/base dataset and the target/novel dataset are visually dissimilar. The original benchmark paper defined this as a practically important departure from conventional FSOD, where source and target data usually come from a similar visual domain such as COCO/VOC. In CD-FSOD, the source domain is rich in labels, the target domain has only $K$-shot annotations per class, and the source/target domains are visually dissimilar; the benchmark used MS COCO as source and ArTaxOr, UODD, and DIOR as target domains, with evaluation at $K = 1, 5, 10$ using mAP (Xiong, 2022).

The empirical importance of this formulation lies in a negative result. The benchmark study evaluated both meta-learning FSOD approaches and fine-tuning FSOD approaches, including Meta-RCNN, H-GCN, Meta Faster R-CNN, TFA w/cos, FSCE, and DeFRCN, and reported that these methods “tend to fall, and even underperform the naive fine-tuning model” on cross-domain transfer (Xiong, 2022). The proposed distillation-based EMA teacher-student baseline reached 13.8 average mAP across the three target datasets and three shot settings, versus 11.8 for DeFRCN, a 2.0 mAP on average gain (Xiong, 2022).

This suggests that NTIRE’s CD-FSOD challenges are not merely competitions on a harder FSOD benchmark, but structured evaluations of whether adaptation mechanisms remain valid when domain mismatch, overfitting, and extreme data scarcity occur simultaneously. The first NTIRE challenge in 2025 formalized this into a public benchmark with open-source and closed-source tracks (Fu et al., 14 Apr 2025); the 2026 edition retained that dual structure and expanded the submission pool and method diversity (Qiu et al., 13 Apr 2026).

2. Task formulation, tracks, and scoring

The NTIRE 2026 challenge follows the CD-FSOD formulation in which the source classes $C_s$ and target classes $C_T$ are disjoint:

$C_s \cap C_T = \varnothing$

The goal is to train on a source domain and adapt to a novel target domain with only a few support annotations per class. The challenge uses the standard $N$-way $K$-shot protocol, where for each target class the support set contains $K$ labeled instances, with $K \in \{1, 5, 10\}$, and the remaining images form the query set (Qiu et al., 13 Apr 2026).

Two tracks structure the evaluation. In the closed-source CD-FSOD track, training is restricted to a fixed source domain, specifically MS-COCO. The organizers treat this as the special track, with one award for the top team. In the open-source CD-FSOD track, participants may use additional data sources, pretraining, foundation models, and other external resources; this is the main track, with awards for the top three teams (Qiu et al., 13 Apr 2026). The distinction mirrors the 2025 challenge design, where the closed-source track measured adaptation under tighter constraints and the open-source track measured what is achievable when large pretrained models and outside data are allowed (Fu et al., 14 Apr 2025).

Evaluation is based on nine mAP values:

$$D1_{1shot}, D1_{5shot}, D1_{10shot},\; D2_{1shot}, D2_{5shot}, D2_{10shot},\; D3_{1shot}, D3_{5shot}, D3_{10shot}$$

where $K = 1, 5, 10$0 correspond to the final test datasets. The official ranking score is the weighted average

$K = 1, 5, 10$1

with the 1-shot condition weighted twice because it is both the hardest and the most practically relevant regime (Qiu et al., 13 Apr 2026). This weighting principle had already been established in NTIRE 2025 (Fu et al., 14 Apr 2025).

3. Domain configuration and dataset regime

The 2026 challenge retains the six CD-ViTO validation target domains:

• ArTaxOr
• Clipart1K
• DIOR
• DeepFish
• NEU-DET
• UODD

It then adds three new unseen test domains for final evaluation:

• RUOD
• CARPK
• CarDD

These target domains differ from MS-COCO in visual style, inter-class variance (ICV), and ambiguous category boundaries (IB) (Qiu et al., 13 Apr 2026). The three final test domains were selected to expose distinct kinds of domain shift: RUOD is an underwater-object domain, CARPK is an aerial vehicle counting/detection domain, and CarDD is a car-damage detection domain (Qiu et al., 13 Apr 2026).

The 2025 challenge used a related but not identical final test configuration: DeepFruits, CARPK, and CarDD (Fu et al., 14 Apr 2025). This suggests that NTIRE 2026 preserved aerial and car-damage evaluation while refreshing the third hidden domain toward underwater imagery.

The importance of the domain regime is inseparable from the original CD-FSOD diagnosis. In the benchmark paper, the source-to-target similarity order from high to low was ArTaxOr, UODD, then DIOR, and the paper argued that cross-domain transfer difficulty arises because features learned from COCO may not align with target-domain appearance, the support examples are too scarce to correct that mismatch, and methods that rely on tight source-domain assumptions overfit (Xiong, 2022). NTIRE 2026 generalizes this logic into a broader multi-domain evaluation in which texture, scale, object appearance, and annotation ambiguity all vary across the final test domains (Qiu et al., 13 Apr 2026).

4. Participation and final standings

The second challenge received 128 registered participants and 696 total submissions. Among these, 31 teams actively participated, and 19 teams submitted valid final results. In the final testing stage, 15 teams submitted results in the open-source track and 4 teams submitted results in the closed-source track (Qiu et al., 13 Apr 2026).

The most important leaderboard pattern was the separation between tracks: open-source methods clearly outperformed closed-source methods, which the report attributes to the use of stronger priors, more data, and larger foundation models (Qiu et al., 13 Apr 2026).

Track	Team	Score
Open-source	FDUROILab_Lenovo	217.21
Open-source	CDiscover	192.79
Open-source	NJUST-KMG	191.38
Closed-source	FewShotEverything	134.31
Closed-source	Fusion-Few	108.48
Closed-source	nudt_0110Dplter	73.71

In the open-source track, the top three teams were FDUROILab_Lenovo with 217.21, CDiscover with 192.79, and NJUST-KMG with 191.38. In the closed-source track, FewShotEverything won with 134.31, followed by Fusion-Few with 108.48, nudt_0110Dplter with 73.71, and freav with 69.82 (Qiu et al., 13 Apr 2026). The report notes that nearly all open-track teams improved significantly over the CD-ViTO baseline, while the large gap between the best open-source and closed-source scores underscores the value of generative models, pseudo-labeling, and external pretraining in this task (Qiu et al., 13 Apr 2026).

A plausible implication is that NTIRE 2026 functions as two benchmarks at once: one for the practical ceiling of contemporary foundation-model ecosystems, and one for the algorithmic efficiency of adaptation under source-constrained few-shot transfer.

5. Leading methods and representative solution families

The top-performing open-source method, FDUROILab_Lenovo, combined open-vocabulary detection, few-shot fine-tuning, target-object cropping and augmentation, LLM/VLM-guided label generation, and dataset-specific post-processing. Its three-stage pipeline consisted of object cropping and augmentation, fine-tuning with augmented data, and Qwen3-VL-235B-A22B-assisted label generation and post-processing. The team reported that the detector often localized correctly but misclassified objects in the target domain, and therefore used Qwen3-VL as an auxiliary classifier for filtering and reclassification (Qiu et al., 13 Apr 2026).

CDiscover, the second-ranked open-source team, used a domain-adaptive hybrid strategy with two branches: a GLIP-based iterative self-training branch for dataset2, and a Qwen-based generative augmentation branch for dataset1 and dataset3, trained with GroundingDINO. For dataset2, the team treated the setting as one with incomplete annotations and used GLIP pseudo-labeling to recover missing objects. For datasets 1 and 3, it used Qwen image generation to synthesize domain-specific training data for GroundingDINO. The reported optimization included classification / contrastive loss, box $K = 1, 5, 10$2 loss, and GIoU loss, with Hungarian matching weights 2.0, 5.0, and 2.0, respectively (Qiu et al., 13 Apr 2026).

NJUST-KMG proposed ASTER, a hybrid teacher-guided adaptation framework with a training-free teacher branch (FSOD-VFM), a trainable student branch (ETS based on GroundingDINO), a pseudo-label bridge from teacher to student, and Domain-RAG augmentation. The teacher side used UPN proposals, SAM2 masks, DINOv2 features, prototype matching, and graph diffusion reweighting; the student side used GroundingDINO with Swin-B and BERT and strong mixed-image augmentation. The team first ran the training-free branch on unlabeled target images, converted high-confidence predictions into COCO-format pseudo annotations, and then continued fine-tuning the student with them (Qiu et al., 13 Apr 2026).

The strongest closed-source method, FewShotEverything, proposed AIPR, combining training-set data augmentation, prototype refinement, and iterative pseudo-labeling. Its three main modules were TDAM, which used Qwen-Image 2.0 + Qwen-VL to synthesize support-like images and generate pseudo labels; PRM, which improved foreground prototype extraction by focusing on semantically relevant object regions; and IPLM, which iteratively mined unlabeled target instances, especially for dataset2 where annotations are incomplete. The architecture used a DINOv2 ViT-L backbone, a Faster R-CNN detector, and SGD with learning rate 0.001 and batch size 16 (Qiu et al., 13 Apr 2026).

Beyond the top entries, the challenge report documents a wide range of method families. earth-insights emphasized support-label quality with SAM3, GroundingDINO, Weighted Boxes Fusion (WBF), and Object-Centric Mosaic augmentation. Intellindust AI Lab introduced ZAP, which selected the better pseudo-label generator per dataset between SAM3 and Qwen3.5-35B-A3B via a custom FSOD-mAP criterion. SAIDA built a four-stage pipeline spanning shot-agnostic domain adaptation, shot-dependent iterative fine-tuning, diffusion-based data augmentation, and final optimization, using ZERO, GroundingDINO, YOLO-E, SAM3, Qwen3-VL, CLIP, CapPa, LoRA, and Qwen-Image (Qiu et al., 13 Apr 2026).

Several teams explored training-free or weakly trained variants of foundation-model pipelines. NTR used a fully training-free combination of UPN, SAM2, and DINOv2, reporting that this setup generalized surprisingly well and sometimes outperformed fine-tuned detectors on unseen domains. J_G_team added negative support features and query-adaptive cross-attention prototypes to a training-free FSOD-VFM variant. WRC combined GroundingDINO, a visual query library, CoOp-style learnable prompts, and LoRA, while also using an EMA Mean Teacher for pseudo labels (Qiu et al., 13 Apr 2026).

Other entries concentrated on explicit classification repair. French Borelli decoupled localization from classification, refining class predictions with a separate vision-only model, DINOv3, and a Nearest Centroid Classifier (NCC) loss. Fusion-Few introduced FusionFormer, with Object-Background Discrimination (OBD), Object-Object Discrimination (OOD), and ensemble classification heads. nudt_0110Dplter extended CD-ViTO through multi-scale prototype fusion and an enhanced training strategy (ETS) while adding only +0.8M trainable parameters over CD-ViTO (Qiu et al., 13 Apr 2026).

6. Recurring themes, limitations, and research significance

The challenge report identifies several clear patterns. First, foundation models are central: many top teams relied on GroundingDINO, DINOv2 / DINOv3, SAM2 / SAM3, Qwen-family VLMs, and large generative models such as Qwen-Image or FLUX (Qiu et al., 13 Apr 2026). Second, pseudo-labeling is extremely important for recovering missing annotations, augmenting sparse support data, bootstrapping second-stage detectors, and refining localization or classification (Qiu et al., 13 Apr 2026). Third, generative augmentation helps, especially when it synthesizes backgrounds, support-like images, compositional scenes, or synthetic labels (Qiu et al., 13 Apr 2026). Fourth, classification is often harder than localization, motivating multimodal reclassification, auxiliary vision-only classifiers, prompt refinement, and prototype alignment (Qiu et al., 13 Apr 2026). Fifth, one-shot remains the hardest regime, which explains the challenge’s weighted score (Qiu et al., 13 Apr 2026).

The limitations are equally explicit. Pseudo-label noise can severely hurt training, especially on fine-grained categories or incomplete labels. Prompt engineering is fragile: good prompts help, but synonym expansion can increase false positives. Domain shift is dataset-dependent, so a method effective on one target domain may fail on another. Overfitting remains a major risk in low-shot fine-tuning; the report notes that training-free methods sometimes generalize better than carefully tuned detectors. Finally, closed-source constraints remain hard, with much lower performance when external priors are unavailable (Qiu et al., 13 Apr 2026).

These findings are consistent with the original CD-FSOD benchmark diagnosis. The 2022 study argued that standard FSOD methods fail because of domain mismatch, overfitting to the source or small target support set, the fact that freezing parameters can hurt, and the inability of meta-learning to solve distant-domain transfer by itself (Xiong, 2022). NTIRE 2025 and NTIRE 2026 show that the strongest contemporary responses to these issues are no longer limited to detector-head modifications; they combine pretrained detectors, pseudo-labeling, prototype recalibration, domain adaptation, multimodal prompting, and, in some cases, training-free vision foundation models (Fu et al., 14 Apr 2025, Qiu et al., 13 Apr 2026).

This suggests that the NTIRE 2026 CD-FSOD Challenge marks a mature stage in the field’s evolution. The central research question is not simply how to design a few-shot detector, but how to coordinate foundation models, support-set exploitation, and adaptation strategies that remain stable under severe domain mismatch and extreme label scarcity.