🥈 NTIRE 2026 CD-FSOD Challenge @ CVPR Workshop

We are the CDiscover team of the NTIRE 2026 Cross-Domain Few-Shot Object Detection (CD-FSOD) Challenge at the CVPR Workshop.

• 🏆 Track: open-source track
• 🎖️ Award: 2nd Place
• 🧰 Method: GiPL-Grounding: Generative augmented iterative Pseudo-Labeling for Grounding

🔗 NTIRE 2026 Official Website
🔗 NTIRE 2026 Challenge Website
🔗 CD-FSOD Challenge Repository

---

📰 News

• [2026.3] 🎉 Win the 2nd Place in the NTIRE 2026 CD-FSOD Challenge, CVPR2026.

🧠 Overview

This repository contains our solution for the open-source track of the NTIRE 2026 CD-FSOD Challenge.
We propose a method that integrates Qwen-based generative augmentation for GroundingDINO and GLIP-driven iterative pseudo-labeling, which achieves strong performance on the challenge.

---

Our method addresses the CDFSOD challenge through a domain-adaptive hybrid strategy.

For Dataset 2 (a multi-object natural domain with vehicles), we identify that only a single instance is labeled per image despite the presence of multiple targets. To prevent the model from treating unlabeled vehicles as background, we employ a GLIP-based pseudo-labeling pipeline to recover missing annotations, followed by iterative self-training to achieve dense and robust object localization. Refer to GLIP-Repository.

For Dataset 1&3, we leverage the Qwen to synthesize generative data, enriching the feature space for GroundingDINO. Implementation details are given as follows.

🛠️ Environment Setup

The experimental environment is based on mmdetection, the installation environment reference mmdetection's installation guide.

conda create --name ets python=3.8 -y
conda activate ets
cd ./mmdetection
pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"

# Develop and run directly mmdet
pip install -v -e .
pip install -r requirements/multimodal.txt
pip install emoji ddd-dataset
pip install git+https://github.com/lvis-dataset/lvis-api.git

Then download the BERT weights bert-base-uncased into the weights directory,

cd CDiscover/
huggingface-cli download --resume-download google-bert/bert-base-uncased --local-dir weights/bert-base-uncased

📂 Dataset Preparation

Please follow the instructions in the official CD-FSOD repo to download and prepare the dataset.

.
├── configs
├── data
├── LICENSE
├── mmdetection
├── pkl2coco.py
├── pkls
├── README.md
├── submit
├── submit_codalab
└── weights

🏋️ Training

Mix Image Augmentation Config

train_pipeline = [
    dict(type='LoadImageFromFile', backend_args=backend_args),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='CachedMosaic', img_scale=(640, 640), pad_val=114.0, prob=0.6),
    # dict(type='CopyPaste', max_num_pasted=5, paste_by_box=True),  # 添加 CopyPaste 数据增强
    dict(type='YOLOXHSVRandomAug'),
    dict(type='RandomFlip', prob=0.5),
    dict(
        type='CachedMixUp',
        img_scale=(640, 640),
        ratio_range=(1.0, 1.0),
        max_cached_images=10,
        pad_val=(114, 114, 114),
        prob = 0.3),
    dict(
        type='RandomChoice',
        transforms=[
            [
                dict(
                    type='RandomChoiceResize',
                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
                            (736, 1333), (768, 1333), (800, 1333)],
                    keep_ratio=True)
            ],
            [
                dict(
                    type='RandomChoiceResize',
                    # The radio of all image in train dataset < 7
                    # follow the original implement
                    scales=[(400, 4200), (500, 4200), (600, 4200)],
                    keep_ratio=True),
                dict(
                    type='RandomCrop',
                    crop_type='absolute_range',
                    crop_size=(384, 600),
                    allow_negative_crop=True),
                dict(
                    type='RandomChoiceResize',
                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
                            (736, 1333), (768, 1333), (800, 1333)],
                    keep_ratio=True)
            ]
        ]),
    # dict(type='RandomErasing', n_patches=(0,2), ratio=0.3, img_border_value=128, bbox_erased_thr=0.9),
    dict(
        type='PackDetInputs',
        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
                   'scale_factor', 'flip', 'flip_direction', 'text',
                   'custom_entities'))
]

To train the model:

Experiments were carried out on on a single NVIDIA A6000 GPU for Datasets 1 and 3.

cd ./mmdetection

./tools/dist_train_muti.sh configs/grounding_dino/CDFSOD/GroudingDINO-few-shot-SwinB.py "0" 1

use sampling4val.py for sampling test set for validation set.

use sata_logs for search to get best model parameter from train logs.

pretrained model:

Download the checkpoint files to dir ./weights.

Baidu Disk: [link] or 通过网盘分享的文件：weights 链接: https://pan.baidu.com/s/1r5H8TsV1qgwuWPuvc1EJZQ?pwd=hg8e 提取码: hg8e --来自百度网盘超级会员v5的分享

🔍 Inference & Evaluation

Run evaluation:

cd ./mmdetection

bash tools/dist_test.sh configs/grounding_dino/CDFSOD/GroudingDINO-few-shot-SwinB.py /path/to/model/ 4

Run inference:

Save to *.pkl file and convert to submit .json format.

cd ./mmdetection

## 1-shot-dataset1
bash tools/dist_test_out.sh ../configs/1-shot-dataset1.py ../weights/1-shot-dataset1-db4c5ebf.pth 1 ../pkls/dataset1_1shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset1/annotations/test.json --pkl_file ../pkls/dataset1_1shot.pkl --output_json ../pkls/dataset1_1shot_coco.json --annotations_json ../submit/dataset1_1shot.json

## 1-shot-dataset2
bash tools/dist_test_out.sh ../configs/1-shot-dataset2.py ../weights/1-shot-dataset2-0bd5d280.pth 1 ../pkls/dataset2_1shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset2/annotations/test.json --pkl_file ../pkls/dataset2_1shot.pkl --output_json ../pkls/dataset2_1shot_coco.json --annotations_json ../submit/dataset2_1shot.json

## 1-shot-dataset3
bash tools/dist_test_out.sh ../configs/1-shot-dataset3.py ../weights/1-shot-dataset3-433149f8.pth 1 ../pkls/dataset3_1shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset3/annotations/test.json --pkl_file ../pkls/dataset3_1shot.pkl --output_json ../pkls/dataset3_1shot_coco.json --annotations_json ../submit/dataset3_1shot.json

## 5-shot-dataset1
bash tools/dist_test_out.sh ../configs/5-shot-dataset1.py ../weights/5-shot-dataset1-ad2ac5f0.pth 1 ../pkls/dataset1_5shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset1/annotations/test.json --pkl_file ../pkls/dataset1_5shot.pkl --output_json ../pkls/dataset1_5shot_coco.json --annotations_json ../submit/dataset1_5shot.json

## 5-shot-dataset2
bash tools/dist_test_out.sh ../configs/5-shot-dataset2.py ../weights/5-shot-dataset2-0bfccba8.pth 1 ../pkls/dataset2_5shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset2/annotations/test.json --pkl_file ../pkls/dataset2_5shot.pkl --output_json ../pkls/dataset2_5shot_coco.json --annotations_json ../submit/dataset2_5shot.json

## 5-shot-dataset3
bash tools/dist_test_out.sh ../configs/5-shot-dataset3.py ../weights/5-shot-dataset3-0011f4b1.pth 1 ../pkls/dataset3_5shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset3/annotations/test.json --pkl_file ../pkls/dataset3_5shot.pkl --output_json ../pkls/dataset3_5shot_coco.json --annotations_json ../submit/dataset3_5shot.json

## 10-shot-dataset1
bash tools/dist_test_out.sh ../configs/10-shot-dataset1.py ../weights/10-shot-dataset1-33caf03b.pth 1 ../pkls/dataset1_10shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset1/annotations/test.json --pkl_file ../pkls/dataset1_10shot.pkl --output_json ../pkls/dataset1_10shot_coco.json --annotations_json ../submit/dataset1_10shot.json

## 10-shot-dataset2
bash tools/dist_test_out.sh ../configs/10-shot-dataset2.py ../weights/10-shot-dataset2-46b5584c.pth 1 ../pkls/dataset2_10shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset2/annotations/test.json --pkl_file ../pkls/dataset2_10shot.pkl --output_json ../pkls/dataset2_10shot_coco.json --annotations_json ../submit/dataset2_10shot.json

## 10-shot-dataset3
bash tools/dist_test_out.sh ../configs/10-shot-dataset3.py ../weights/10-shot-dataset3-7325994e.pth 1 ../pkls/dataset3_10shot.pkl

python ../pkl2coco.py --coco_file ../data/dataset3/annotations/test.json --pkl_file ../pkls/dataset3_10shot.pkl --output_json ../pkls/dataset3_10shot_coco.json --annotations_json ../submit/dataset3_10shot.json

Acknowledgements

Provided codes were adapted from:

• jaychempan/ETS
• microsoft/GLIP
• IDEA-Research/DINO
• IDEA-Research/GroundingDINO
• longzw1997/Open-GroundingDino