Source-linked AI summary
The Second Challenge on Cross-Domain Few-Shot Object Detection at NTIRE 2026: Methods and Results
Xingyu Qiu, Yuqian Fu, Jiawei Geng, Bin Ren, Jiancheng Pan, Zongwei Wu, Hao Tang, Yanwei Fu, Radu Timofte, Nicu Sebe, Mohamed Elhoseiny, Lingyi Hong, Mingxi Cheng, Xingqi He, Runze Li, Xingdong Sheng, Wenqiang Zhang, Jiacong Liu, Shu Luo, Yikai Qin, Yaze Zhao, Yongwei Jiang, Yixiong Zou, Zhe Zhang, Yang Yang, Kaiyu Li, Bowen Fu, Zixuan Jiang, Ke Li, Hui Qiao, Xiangyong Cao, Xuanlong Yu, Youyang Sha, Longfei Liu, Di Yang, Xi Shen, Kyeongryeol Go, Taewoong Jang, Saiprasad Meesiyawar, Ravi Kirasur, Rakshita Kulkarni, Bhoomi Deshpande, Harsh Patil, Uma Mudenagudi, Shuming Hu, Chao Chen, Tao Wang, Wei Zhou, Qi Xu, Zhenzhao Xing, Dandan Zhao, Hanzhe Xia, Dongdong Lu, Zhe Zhang, Jingru Wang, Guangwei Huang, Jiachen Tu, Yaokun Shi, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Liwei Zhou, Bei Dou, Tao Wu, Zekang Fan, Junjie Liu, Adhémar de Senneville, Flavien Armangeon, Mengbers, Yazhe Lyu, Zhimeng Xin, Zijian Zhuang, Hongchun Zhu, Li Wang
TL;DR
CD-FSOD asks detectors to recognize novel objects across substantial domain shifts with only a few labeled examples, a setting that remains underexplored. This report presents the second NTIRE CD-FSOD Challenge, expanding target domains and supporting open- and closed-source evaluation. In the open-source track, FDUROILab Lenovo achieved the highest reported score, while FewShotEverything led the closed-source track.
Problem
CD-FSOD must generalize to novel target domains with limited labeled data, while conventional few-shot detection commonly assumes source and target data come from the same domain.
Method
The challenge evaluates open- and closed-source CD-FSOD tracks across six validation domains and three newly introduced testing datasets, using few-shot target support sets.
Results
FDUROILab Lenovo achieved 217.21 in the open-source track, while FewShotEverything achieved 134.31 in the closed-source track.
Takeaways & Limitations
The results show strong improvements over the CD-ViTO baseline in the open-source track and meaningful baseline improvements under stricter closed-source constraints.
Takeaways & Limitations
Vision-language detectors can remain limited by fine-grained classification errors, while pseudo-label-driven adaptation depends on pseudo-label quality and threshold tuning.
Abstract
from arXiv · showhide
Cross-domain few-shot object detection (CD-FSOD) remains a challenging problem for existing object detectors and few-shot learning approaches, particularly when generalizing across distinct domains. As part of NTIRE 2026, we hosted the second CD-FSOD Challenge to systematically evaluate and promote progress in detecting objects in unseen target domains under limited annotation conditions. The challenge received strong community interest, with 128 registered participants and a total of 696 submissions. Among them, 31 teams actively participated, and 19 teams submitted valid final results. Participants explored a wide range of strategies, introducing innovative methods that push the performance frontier under both open-source and closed-source tracks. This report presents a detailed overview of the NTIRE 2026 CD-FSOD Challenge, including a summary of the submitted approaches and an analysis of the final results across all participating teams. Challenge Codes: https://github.com/ohMargin/NTIRE2026_CDFSOD.
1. Introduction
CD-FSOD extends few-shot detection to unseen target domains, where the usual same-domain assumption fails. The challenge addresses this underexplored and difficult setting by adding target domains and evaluating both open- and closed-source tracks.
- Few-shot object detectors must recognize and localize novel categories from only a handful of labeled examples, but commonly assume source and target data share a domain.
- Cross-domain few-shot object detection remains relatively underexplored despite broader study of cross-domain few-shot classification.
- The challenge is motivated by the need to further advance a new and difficult CD-FSOD task.
- The challenge adds RUOD, CARPK, and CarDD as unseen target domains exhibiting visual-style variation, inter-class variance, and ambiguous category boundaries.
- The evaluation includes an open-source primary track and a closed-source supplementary track, with MS-COCO as the exclusive closed-source training domain.
2. NTIRE 2026 CD-FSOD Challenge
The NTIRE 2026 CD-FSOD Challenge evaluates object detection across domain shifts under limited target-domain labels through closed-source and open-source tracks, development and testing stages, and weighted few-shot scoring. It introduces new target datasets and establishes baseline methods for both tracks.
- Challenge scope: The challenge targets cross-domain few-shot object detection with limited labeled data, using six published target domains for validation and three newly constructed datasets for final testing.It also introduces an open-source track in which participants may select source datasets and pretrained models.
- Challenge settings: In the closed-source track, source and target class sets are disjoint, domain distributions differ, and models trained on the source dataset are tested with few labeled examples per target class.The track typically uses a single source dataset, as in CD-ViTO.
- Challenge settings: The challenge uses an open-source main track and a closed-source special track, with N-way K-shot evaluation based on target support and query sets.The open-source track awards the top three teams, while the closed-source track awards one top-performing team.
- Evaluation stages: Development uses ArTaxOr, Clipart1K, DIOR, DeepFish, NEU-DET, and UODD, while testing uses three previously unseen datasets with the same source-data policy for each track.The supplied passage identifies the development datasets and begins listing the unseen testing datasets.
- Baseline methods: The closed-source baseline is CD-ViTO, while the open-source baseline is Domain-RAG, which retrieves and generates domain-aligned backgrounds before composing them with foreground objects.CD-ViTO adds learnable instance features, instance reweighting, a domain prompter, and finetuning; Domain-RAG is training-free and preserves foreground semantics during augmentation.
- Scoring: Final ranking averages mAP across three testing datasets and 1-shot, 5-shot, and 10-shot settings, assigning twice the weight to 1-shot performance.The weighting emphasizes robustness under extremely limited labeled data.
3. Challenge Results
The challenge results show that open-source participation substantially improved CD-FSOD performance, while closed-source teams also clearly surpassed the baseline under stricter constraints. The leading open-source teams achieved the highest scores, and FewShotEverything led the special closed-source track.
- Open-Source Track Results: Nearly all open-source teams clearly improved over the CD-ViTO baseline, indicating an advantage from relaxing source-data restrictions.The open-source track allowed greater freedom in selecting source data, which the results associate with stronger CD-FSOD performance.
- Open-Source Track Results: 217.21, 192.79, and 191.38 were the final scores of FDUROILab Lenovo, CDiscover, and NJUST-KMG, respectively, leading the open-source track.These teams significantly outperformed the baseline and other competing teams in the same track.
- Closed-Source Track Results: Closed-source performance was generally lower than open-source performance, but participants still clearly improved the baseline under stricter constraints.The closed-source setting limits source-data flexibility compared with the open-source track.
- Closed-Source Track Results: 134.31 was FewShotEverything’s final score, making it the leading result in the special closed-source track.Its framework generated underwater images from text prompts and used a vision-language model to produce pseudo labels for those samples.
4. Main Open-Source Track Methods
Main open-source-track methods combine pseudo-labeling, generative augmentation, foundation-model adaptation, and dataset-specific strategies to address sparse supervision and domain shift. Reported analyses emphasize that performance depends on domain characteristics, label quality, and the balance between zero-shot and fine-tuned detection.
- Generative augmentation: Generative augmentation expands scarce target-domain data, including synthetic samples, domain-specific backgrounds, and diffusion-based image generation.Methods use Qwen, Domain-RAG, or Qwen-Image to diversify training data while preserving or enriching target-domain appearance.
- Hybrid adaptation: Hybrid frameworks combine training-free proposal or foundation-model branches with trainable detectors and supplementary pseudo supervision.ASTER pairs FSOD-VFM proposals and support-prototype matching with an ETS GroundingDINO branch, while other methods combine foundation models with parameter-efficient adaptation.
- Dataset-specific strategies: Dataset-specific adaptation is necessary because zero-shot performance varies sharply with object properties and domain difficulty.SAM3 exceeds 50% zero-shot mAP on CarPK but falls below 10% on CarDD, motivating tailored processing and augmentation.
- Pseudo-labeling: Pseudo-labeling frameworks use high-confidence detections to refine supervision and adapt vision-language detectors without additional manual annotations.GLIP-based iterative refinement and related pseudo-label-driven pipelines target false negatives and localization limitations under domain shift.
- Limitations: Pseudo-label quality and threshold selection remain practical constraints, with iterative fine-tuning adding computational overhead and rare categories remaining difficult.The KLETech-CEVI analysis identifies dependence on pseudo-label quality, threshold tuning, and weak semantic alignment as limitations.
5. Special Closed-Source Track Methods
The closed-source track methods build on CD-ViTO with data augmentation, prototype refinement, pseudo-labeling, multi-scale prototypes, or feature-discrimination ensembles to improve CD-FSOD under limited target annotations.
- Training-Set Data Augmentation synthesizes support-like images and detection annotations to increase sample diversity and robustness to domain shift.
- Prototype Refinement focuses foreground prototype extraction on semantically relevant object regions, reducing surrounding noise and irrelevant background.
- Iterative Pseudo-Labeling progressively identifies unlabeled target instances in incompletely annotated support images, enriching supervision.
- FusionFormer combines Object-Background Discrimination, Object-Object Discrimination, and an offline ensemble of linear, graph-diffusion, and feature-cache predictions.
- Multi-scale prototype fusion uses 0.9x, 1.0x, 1.1x, and 1.2x image scales, while Enhanced Training Strategy applies augmentation across 11 resize scales.
- The multi-scale optimization preserves CD-ViTO parameter efficiency with only +0.8M trainable parameters.
CDiscover
CDiscover is identified by its affiliation with Huazhong University of Science and Technology.
- CDiscover is affiliated with Huazhong University of Science and Technology.
SAIDA
SAIDA is presented as a synthetic-augmented iterative domain adaptation method by Superb AI members Kyeongryeol Go and Taewoong Jang.
- SAIDA is titled Synthetic-Augmented Iterative Domain Adaptation and includes members Kyeongryeol Go and Taewoong Jang from Superb AI.
KLETech-CEVI
KLETech-CEVI is associated with a zero-shot GLIP transfer method that uses multi-scale test-time augmentation and researchers from KLE Technological University and CEVI.
- KLETech-CEVI uses zero-shot transfer with GLIP and multi-scale test-time augmentation.
- The listed affiliations include KLE Technological University and the Center for Visual Intelligence.