PaperlayerView exact revision on arXiv

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

arXiv:2604.11998v1cs.CVcs.AI

TL;DR

Few-shot detectors often assume source and target data share a domain, although real-world detection can involve substantial domain shifts. This paper reports the second NTIRE 2026 CD-FSOD Challenge, evaluating methods under limited target annotations across closed- and open-source tracks. Nearly all open-source teams improved over the provided CD-ViTO baseline.

  • Problem

    Most few-shot object detection approaches assume source and target data share a domain, although real-world scenarios can require generalization across substantial domain shifts.

    Source: §1 · p. 1

  • Method

    The challenge evaluates CD-FSOD in closed- and open-source tracks, using six published target domains for validation and three newly constructed datasets for final testing.

    Source: §2.1 · p. 2, §2.2 · p. 2, §2.3 · p. 3

  • Results

    Nearly all open-source teams improved over the provided CD-ViTO baseline, with FDUROILab Lenovo leading that track with a score of 217.21.

    Source: §3 · p. 4

  • Takeaways & Limitations

    The reported findings indicate that relaxing source-data restrictions offers an advantage for CD-FSOD and underscore the significance of the open-source task setting.

    Source: §3 · p. 4

    The described approach may degrade for rare categories with weak semantic alignment.

    Source: §4.7.2 · p. 12

Abstract

from arXiv · show

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

Few-shot detectors often assume training and test data come from the same domain, an assumption that fails under real-world domain shifts. The second NTIRE CD-FSOD Challenge evaluates generalization across additional unseen domains under closed- and open-source settings.

1. Introduction

  • The challenge addresses few-shot object detection’s same-domain assumption, which can fail when detectors trained on natural images encounter substantially different domains such as remote sensing.FSOD recognizes and localizes novel categories from a handful of labeled examples, but real-world domain shifts challenge this capability.Source: §1 · p. 1
  • CD-FSOD remains relatively underexplored; CD-ViTO [27] defined the task and benchmarked COCO against six target datasets, followed by several methods.Subsequent approaches include CDFormer [69], StyleProto, ETS [74], Domain-RAG [58], and LMP [112].Source: §1 · p. 1
  • The second challenge adds RUOD, CARPK [36], and CarDD [114] as unseen targets and evaluates both closed-source and open-source CD-FSOD settings.These domains exhibit discrepancies in visual style, inter-class variance, and category-boundary ambiguity; open-source participants may use additional data, prior knowledge, and foundation models.Source: §1 · p. 2
  • The NTIRE 2026 challenge makes open-source the primary track and closed-source supplementary, using MS-COCO as the exclusive source in the closed-source setting.Validation uses six predefined target domains, while three newly introduced domains are reserved for final evaluation in both tracks; mAP is the official ranking metric.Source: §1 · p. 2

2. NTIRE 2026 CD-FSOD Challenge

The challenge evaluates cross-domain few-shot object detection under domain shift, pairing newly introduced final-test datasets with closed- and open-source tracks, baseline methods, and a weighted few-shot evaluation protocol.

2.1. Challenge Overview

  • The challenge advances CD-FSOD by using six published target domains [27] for validation, introducing three new final-test datasets, and allowing free source-dataset and pretrained-model selection in its open-source setting.The challenge presents both predefined closed-source and open-source settings.Source: §2.1 · p. 2

2.2. Task Formulations

  • The closed-source track requires transfer between disjoint source and target classes, while the open-source track relaxes source-data constraints but retains domain-shift and few-shot detection challenges.The open-source track is the main track, with awards for its top three teams; the closed-source special track awards its top team. Both use an N-way K-shot support set to assist detection on a query set.Source: §2.2 · pp. 2–3

2.3. Challenge Phases and Datasets

  • The challenge uses six existing domains for development and three previously unseen datasets—RUOD, Carpk [36], and CarDD [114]—for final testing.MS-COCO [59] is the closed-source source dataset in both stages, while the open-source track permits unlimited source data; final-test query annotations are held by the organizers.Source: §2.3 · p. 3

2.4. CD-ViTO Baseline Model

  • CD-ViTO [27], built on DE-ViT [125] and fine-tuned on support examples, is the closed-source baseline, combining learnable instance features, instance reweighting, domain prompting, and fine-tuning.Learnable instance features promote inter-class variance; instance reweighting emphasizes higher-quality objects. Domain perturbations simulate styles while prototype consistency and domain diversity losses preserve semantics and encourage varied domains, improving robustness to shifts.Source: §2.4 · p. 3

2.5. Domain-RAG Baseline Model

  • Domain-RAG [58], the open-source baseline, augments training images with domain-consistent synthetic backgrounds without extra supervision or additional model training.It retrieves relevant backgrounds, generates target-aligned backgrounds conditioned on retrieved context, and composes them with original foregrounds; this preserves foreground semantics while adapting background appearance and can be applied to different base methods.Source: §2.5 · p. 4

2.6. Evaluation Protocol

  • The final score aggregates nine mAP scores across three test datasets and 1-, 5-, and 10-shot settings, weighting 1-shot performance twice as heavily as the other settings.The protocol emphasizes 1-shot performance because it is generally lower with fewer labeled examples and encourages robustness under extremely limited data.Source: §2.6 · p. 4

3. Challenge Results

The challenge results show strong baseline improvements in both tracks, with higher performance in the open-source track and a leading closed-source result from FewShotEverything.

Challenge Results

  • 15 open-source and 4 closed-source teams submitted results, codes, and factsheets in the final testing stage.Source: §3 · p. 4

Open-Source Track Results

  • Nearly all open-source teams improved clearly over the CD-ViTO baseline, and relaxing source-data restrictions offered a substantial advantage for CD-FSOD.Source: §3 · p. 4
  • 217.21, 192.79, and 191.38 were the scores of open-source track leaders FDUROILab Lenovo, CDiscover, and NJUST-KMG, respectively, significantly outperforming the baseline and other teams.Table 1 summarizes open- and closed-source results on CD-FSOD.Source: §3 · p. 4, Table 1 · p. 5

Closed-Source Track Results

  • Closed-source teams generally scored below open-source teams under stricter constraints, but still clearly improved the baseline.Source: §3 · p. 4
  • 134.31 was FewShotEverything’s final score, significantly ahead of competitors in the Special Closed-Source Track.Its pipeline generated underwater images from text prompts, then used a vision-language model to produce pseudo labels; other teams also improved results through closed-source large-model data augmentation.Source: §3 · p. 4

4. Main Open-Source Track Methods

This section surveys open-source teams’ approaches to cross-domain few-shot detection, spanning detector adaptation and training-free foundation-model pipelines. It covers strategies including pseudo-labeling and self-training, data augmentation and enrichment, multimodal fusion or prompting, fine-tuning and distillation, and inference-time refinement tailored to target domains.

4.1.1. Proposed Method

  • The team adapts an open-vocabulary detector by cropping, augmenting, rescaling, and pasting k-shot objects, then fine-tunes it and uses Qwen3-VL [4] for label prompts and classification correction.Qwen3-VL supports filtering or reclassification, selected by dataset; training is omitted for Dataset 2, where oversized boxes are filtered, and NMS is applied across datasets.Source: §4.1.1 · pp. 5–6, §4.1.2 · p. 6

4.2.1. Proposed Method

  • A domain-adaptive hybrid uses Qwen-generated training data for GroundingDINO on Datasets 1 and 3, and GLIP pseudo-labeling with iterative self-training to recover missing vehicle annotations on Dataset 2.The team selects GroundingDINO or GLIP according to domain, and also augments the limited support set with semantically consistent Qwen-image samples.Source: §4.2.1 · pp. 6–7, §4.2.2 · p. 7

4.3.1. Proposed Method

  • A near-60 mAP training-free FSOD-VFM baseline on Dataset 2 is combined with GroundingDINO-based ETS, pseudo-label continuation, and Domain-RAG; the best branch is selected per dataset-shot setting.The pipeline uses high-confidence FSOD-VFM predictions as pseudo annotations for ETS and adds compositional synthetic samples, particularly for low-shot settings.Source: §4.3.1 · pp. 7–8, Inference Details · p. 8

4.4.1. Proposed Method

  • SAM3 and Grounding DINO generate support labels that are fused with WBF; object-centric mosaics and dataset-specific adaptation address scarce samples and variable zero-shot performance.SAM3 zero-shot mAP exceeds 50% on Carpk but falls below 10% on CarDD; only the text encoder is updated during SAM3 fine-tuning.Source: §4.4.1 · pp. 8–9, §4.4.2 · p. 9

4.5.1. Proposed Method

  • Qwen3.5-35B-A3B and SAM3 [10] generate pseudo-labels, which FSOD-mAP selects by dataset before two HED [121]-equipped MMGroundingDINO-L [127] models are fine-tuned and ensembled.SAM3 is selected for Datasets 1 and 2, Qwen3.5 for Dataset 3; inference uses horizontal-flip TTA and Soft-NMS.Source: §4.5.1 · p. 9

4.6.1. Proposed Method

  • SAIDA targets false negatives from sparse labels with a four-phase pipeline progressing from shot-agnostic adaptation to iterative pseudo-labeling and generative data expansion.Its initial phase uses ZERO [14] with open-vocabulary detectors and Qwen3-VL [3] to form noun-phrase pairs, then applies CLIP [82] and CapPa [104] losses and LoRA [37] prompt-encoder tuning.Source: §4.6.1 · p. 10, §4.6.2 · p. 10

4.6.2. Module Details

  • SAIDA iteratively optimizes class-wise pseudo-label thresholds for precision then recall, merges labels with ground truth using class-agnostic NMS, and trains ZERO and Co-DETR [132].It expands training data tenfold with diffusion-generated samples, then uses color augmentation and multi-scale, horizontal-flip TTA; the method emphasizes parameter-efficient adaptation and reduced human involvement.Source: §4.6.2 · pp. 10–11

4.7.1. Proposed Method

  • KLETech-CEVI presents a GLIP-based vision-language grounding pipeline comprising zero-shot detection and pseudo-label generation to address localization limitations under domain shift.The supplied passage introduces a third stage but does not specify it.Source: §4.7.1 · p. 11
  • More than 2× improvement in the 1-shot scenario is reported for KLETech-CEVI’s iterative pseudo-label adaptation, which combines GLIP grounding with confidence, NMS, and class-consistency filtering.The team also applies multi-scale test-time augmentation; its approach depends on pseudo-label quality and adds iterative fine-tuning overhead.Source: §4.7.1 · p. 11, §4.7.2 · p. 12

4.7.2. Training Details

  • KLETech-CEVI’s ablations attribute the largest gain to pseudo-labeling, with Focal Loss helping hard examples, GIoU Loss improving box quality, and multi-scale inference aiding size robustness.The training uses large-scale vision-language datasets and progressively adds high-confidence predictions as supervision without further manual annotations.Source: §4.7.2 · p. 12

4.8.1. Proposed Method

  • The MPDA team uses Qwen3-VL background prompts and Flux.1 Fill inpainting to expand few-shot samples while preserving foreground boxes, then fine-tunes Grounding DINO on synthesized data.The method is source-free, uses support sets across 1-, 5-, and 10-shot settings on three datasets, and evaluates nine configurations.Source: §4.8.1 · p. 12, §4.8.2 · p. 12

4.9.1. Proposed Method

  • 11.23 → 23.42 mAP on D1 followed descriptive multi-synonym prompting, while D3 rose from 9.12 to 36.08 mAP (+26.96) after fine-tuning on 60 images.The team uses zero-shot prompting for D1 and D2 but fine-tunes GroundingDINO for specialized D3; adding synonyms hurt D2 (57.06 → 41.04), and support-image pseudo-label self-training degraded performance.Source: §4.9.1 · p. 13, Figure 14 · p. 14, Ablation and analysis · p. 14

4.10.1. Proposed Method

  • AIRCAS MILab trains GroundingDINO separately for nine dataset-shot tasks, using semantically filtered Domain-RAG samples and task-specific configuration and inference post-processing.The team abandoned second-stage pseudo-label retraining because of overfitting risks in extremely low-shot conditions.Source: §4.10.1 · p. 14, §4.10.2 · p. 15

4.11.1. Proposed Method

  • J G adapts FSOD-VFM with shifted-and-scaled positive boxes as negative context and query-adaptive cross-attention prototypes; the method is training-free and uses no task-specific fine-tuning.Its pipeline also uses UPN for proposals, SAM2 for support masks, and C-RADIOv4-H for visual features.Source: §4.11.1 · p. 15, §4.11.2 · p. 15

4.12.1. Proposed Method

  • The FSOD-VFM approach combines UPN proposals, DINOv2 support-query feature matching, graph-based confidence diffusion, and SAM2 mask refinement without target-domain fine-tuning.It uses only the provided K-shot support set and no additional training data.Source: §4.12.1 · p. 16

4.12.2. Training Details

  • Training-free vision foundation models outperformed a fine-tuned GroundingDINO baseline on test domains, although fine-tuned GroundingDINO performed better on validation domains.The challenge emphasizes 1-shot performance with 2× weight, and its authors suggest that over-parameterized few-shot adaptation can harm cross-domain generalization.Source: §4.12.2 · p. 16

4.13.1. Proposed Method

  • WRC combines a multi-scale visual query library, cascaded feature refinement, and gated fusion to align visual queries with text semantics in a GroundingDINO-based detector.Its lightweight fine-tuning freezes the visual backbone and shallow text encoder, uses learnable prompts and LoRA, and tunes only the fusion module to improve few-shot detection and alleviate overfitting.Source: §4.13.1 · pp. 16–17
  • WRC trains with joint image-text fine-tuning and EMA Mean Teacher pseudo-labels to address partially unlabeled instances in few-shot datasets.The visual backbone and shallow text-encoder layers remain frozen; learnable prompts and LoRA adapt the text branch and cross-attention fusion.Source: §4.13.1 · p. 17

4.13.2. Training Details

  • WRC uses Swin-T, BERT-base-uncased, and Deformable DETR, with a Vision Query Bank, gated cross-attention, CoOp prompts, and LoRA modules.Experiments use 1-, 5-, and 10-shot settings on COCO17 with fixed support sets, train for 20 epochs, and use a fixed-scale inference setup without test-time augmentation or ensembling.Source: §4.13.1 · p. 17, §4.13.2 · p. 17

4.14.1. Method overview

  • DDR combines Domain-RAG support-set diversification, DINOv3 pseudo-labeling and distillation, and two-stage fine-tuning to bridge domain gaps with limited annotations.It first adapts on Drone Vehicle and URPC2020, then fine-tunes for the target domain; its architecture uses DINOv3-ConvNeXt-Large with an RT-DETR head.Source: §4.14.1 · pp. 17–18, §4.14.2 · p. 18

4.15.1. Proposed Method

  • Triple-Tower decouples localization from classification, using a vision-only auxiliary model to improve classification beyond language-aligned features.It fine-tunes WeDetect with pseudo-annotations for unannotated objects, then boosts box-class predictions with DINOv3; only LayerNorm layers, 0.03% of weights, are trained.Source: §4.15.1 · p. 18

4.15.2. Training details

  • Triple-Tower adds high-confidence support-set detections as pseudo-annotations from epoch 15 every 5 epochs, then trains DINOv3 with a Nearest Centroid Classifier loss.During inference, it adds vision-based and language-based classifier logits; images are resized to 640×640, with letterbox padding.Source: §4.15.2 · pp. 18–19

5. Special Closed-Source Track Methods

The closed-source track describes methods that address domain shift and limited support data through augmentation, prototype refinement, pseudo-labeling, and feature-discrimination strategies. It also details the training settings and architectures used by the participating teams.

5.1.1. Proposed Method

  • The CD-ViTO-based method combines synthetic-data augmentation, prototype refinement, and iterative pseudo-labeling to address domain shift and incomplete support annotations.It uses a DINOv2 ViT-L/14 backbone with Faster R-CNN; iterative pseudo-labeling is intended to mine unlabeled target instances, including vehicles in Dataset2.Source: §5.1.1 · p. 19

5.1.2. Training details

  • Training uses COCO for source-domain base training and evaluates 1-, 5-, and 10-shot target settings, applying iterative pseudo-labeling only to Dataset2.Qwen-Image2.0 generates target-like images and Qwen-VL produces pseudo-labels; optimization uses SGD at learning rate 0.001 with batch size 16.Source: §5.1.2 · p. 20

5.2.1. Proposed Method

  • FusionFormer combines object-background and object-object discrimination with an offline ensemble of linear, graph-diffusion, and feature-cache predictions.Its workflow also uses source-domain augmentation, test-time pseudo-labeling, and test-time augmentation; the detector uses a frozen DINOv2 ViT-L/14 backbone with a Deformable-DETR encoder-decoder.Source: §5.2.1 · p. 20, Figure 22 · p. 20

5.3.1. Method overview

  • The CD-ViTO-based method adds multi-scale prototype fusion and an Enhanced Training Strategy to improve scale robustness and augment limited support sets.Prototype extraction fuses features from scales 0.9x, 1.0x, 1.1x, and 1.2x using temperature 0.1; the method adds 0.8M trainable parameters.Source: §5.3.1 · p. 21

5.3.2. Model architecture

  • The method retains CD-ViTO components and adds multi-scale prototype extraction plus random-flip, resize, and crop augmentation to its DE-ViT detector.It uses a frozen DINOv2 ViT-L/14 backbone and fuses prototypes from scales 0.9, 1.0, 1.1, and 1.2 with temperature-scaled softmax weighting at temperature 0.1.Source: §5.3.2 · p. 21

A. Teams and affiliations … KLETech-CEVI

KLETech-CEVI’s submission is titled “Zero-Shot Transfer with GLIP and Multi-Scale Test-Time Augmentation.”

KLETech-CEVI

  • KLETech-CEVI submitted “Zero-Shot Transfer with GLIP and Multi-Scale Test-Time Augmentation.”Source: KLETech-CEVI · p. 23
  • The team’s listed affiliations are KLE Technological University’s School of Computer Science and Engineering and Department of Electronics and Communication Engineering, and the Center for Visual Intelligence (CEVI), all in Hubballi, India.Source: KLETech-CEVI · p. 23

Related papers

Loading 2604.11998v1…