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

cs.CVApr 13, 2026

The Second Challenge on Cross-Domain Few-Shot Object Detection at NTIRE 2026: Methods and Results

Authors: 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

Abstract

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.

Figures & tables

Figure 1: Illustration of the challenge settings, including the closed-source and open-source CD-FSOD tracks. The three newly introduced target datasets used in the final testing phase are also shown.

Figure 2: Overall framework of CD-ViTO baseline method.

Figure 3: Overview of Domain-RAG. Domain-RAG consists of three stages: (1) Domain-aware background retrieval, (2) Domain-guided background generation, and (3) Foreground-background composition. The whole pipeline is training-free and follows the principle of fix foreground, adapt background .

Main Open-Source Track (2026)
Rank	Team Name	Score	D1_1shot	D1_5shot	D1_10shot	D2_1shot	D2_5shot	D2_10shot	D3_1shot	D3_5shot	D3_10shot
1	FDUROILab_Lenovo	217.21	57.04	57.15	58.08	59.23	59.23	59.23	45.23	46.17	48.77
2	CDiscover	192.79	34.61	41.14	42.06	63.26	63.00	61.29	39.71	47.43	48.30
3	NJUST-KMG	191.38	35.62	47.51	46.22	60.41	60.51	61.12	40.09	42.01	44.54
4	earth-insights	190.09	38.20	44.95	46.59	58.73	62.78	63.63	33.95	40.10	50.48
5	Intellindust_AI_Lab	188.05	39.61	43.05	45.25	53.42	53.60	53.29	44.86	45.82	47.37

Table 1: Open-source and closed-source results on CD-FSOD (2026). D1, D2, and D3 represent RUOD, CARPK, and CarDD, respectively. Mean Average Precision (mAP) on 1-shot, 5-shot, and 10-shot are reported. Teams achieving top results are highlighted.

Figure 4: Overview of their efficient tuning and inference.

Figure 6: Proposed framework for CDFSOD. Branch 1 iteratively refines GLIP with pseudo labels, while Branch 2 uses Qwen image generation to augment training data for Grounding DINO.

Figure 7: Team earth-insights: overview of the training process.

Figure 10: Overview of the proposed pipeline.

Figure 12: Overall pipeline of Multimodal Prompt-Driven Diffusion Augmentation (MPDA) .

Figure 13: Overall pipeline of GDino-FT. For D1 (underwater) and D2 (car), they use pretrained GroundingDINO with optimized text prompts (zero-shot). For D3 (car damage), they fine-tune the model on the 10-shot support set while freezing the BERT text encoder.

Configuration	D1	D2	D3	Total
(a) ZS, basic prompts	11.23	57.06	9.12	103.21
(b) ZS, optimized prompts	23.42	41.04	7.14	95.47
(c) Best ZS per domain	23.42	57.06	9.12	119.47
(d) (c) + FT D3	23.42	57.06	33.78	152.35
(e) (d) + lower threshold	23.42	57.06	36.08	155.42

Table 2: Ablation study showing progressive improvements from prompt optimization and fine-tuning.

Figure 14: Ablation results. Prompt optimization boosts D1 (+12 mAP), fine-tuning dramatically improves D3 (+27 mAP), while D2 is near-optimal with simple zero-shot detection.

Figure 15: Overall pipeline of the AIRCAS MILab team’s method. Each dataset-shot pair is treated as an independent few-shot task. Support annotations are expanded with semantically filtered Domain-RAG generated samples, used to build task-specific GroundingDINO configurations, and detector outputs are filtered with dataset-specific post-processing for submission.

Figure 16: Overall pipeline of Negative Prompting for Few-Shot Object Detection.

Figure 17: Overall pipeline of Multimodal Query-based Object Detection.

Figure 18: Overall pipeline of DDR .

Figure 21: Overall pipeline of the data generation process for data augmentation.

Figure 22: Overview of the FusionFormer framework. The architecture leverages a frozen DINOv2 ViT backbone for feature extraction. It features a Siamese-style encoder-decoder structure integrated with Object-Background Discrimination (OBD) and Object-Object Discrimination (OOD) modules to mitigate cross-domain feature confusion. The final detection head employs a multi-strategy ensemble (Prototype Cosine, Graph Diffusion, and MLP) to enhance classification robustness in few-shot scenarios.

Few-Shot Object Detection

Few-shot object detection adapts detectors to recognize and localize new object classes from only a handful of labeled examples.

See all 27 papers

A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps
Mar 30, 2026Xuanlong Yu, Youyang Sha, Longfei Liu +2Few-Shot Object Detection Object Detection
Intellindust AI Lab · Suzhou Institute for Advanced Research, USTC
Learning Multi-Modal Prototypes for Cross-Domain Few-Shot Object Detection
Feb 21, 2026Wanqi Wang, Jingcai Guo, Yuxiang Cai +1Few-Shot Object Detection Object Detection
University of Chinese Academy of Sciences · The Hong Kong Polytechnic University · Zhejiang University +1
Remedying Target-Domain Astigmatism for Cross-Domain Few-Shot Object Detection
Mar 19, 2026Yongwei Jiang, Yixiong Zou, Yuhua Li +1Few-Shot Object Detection Object Detection
School of Computer Science and Technology, Huazhong University of Science and Technology

Mar 30, 2026J/K move · Enter open · S save

A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps

Xuanlong Yu, Youyang Sha, Longfei Liu, Xi Shen+1

Intellindust AI Lab · Suzhou Institute for Advanced Research, USTC

Few-Shot Object Detection Object Detection Fine-Tuning OOD Generalization

Open paper AskPDF arXiv

Few-shot object detection (FSOD) is challenging due to unstable optimization and limited generalization arising from the scarcity of training samples. To address these issues, we propose a hybrid ensemble decoder that enhances generalization during fine-tuning. Inspired by ensemble learning, the decoder comprises a shared hierarchical layer followed by multiple parallel decoder branches, where each branch employs denoising queries either inherited from the shared layer or newly initialized to encourage prediction diversity. This design fully exploits pretrained weights without introducing additional parameters, and the resulting diverse predictions can be effectively ensembled to improve generalization. We further leverage a unified progressive fine-tuning framework with a plateau-aware learning rate schedule, which stabilizes optimization and achieves strong few-shot adaptation without complex data augmentations or extensive hyperparameter tuning. Extensive experiments on CD-FSOD, ODinW-13, and RF100-VL validate the effectiveness of our approach. Notably, on RF100-VL, which includes 100 datasets across diverse domains, our method achieves an average performance of 41.9 in the 10-shot setting, significantly outperforming the recent approach SAM3, which obtains 35.7. We further construct a mixed-domain test set from CD-FSOD to evaluate robustness to out-of-distribution (OOD) samples, showing that our proposed modules lead to clear improvement gains. These results highlight the effectiveness, generalization, and robustness of the proposed method. Code is available at: https://github.com/Intellindust-AI-Lab/FT-FSOD.

Abstract

Few-Shot Object Detection

Explore similar work

A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps

Learning Multi-Modal Prototypes for Cross-Domain Few-Shot Object Detection

Remedying Target-Domain Astigmatism for Cross-Domain Few-Shot Object Detection