CFF-KDNet: Cross-Scale Feature Fusion Network with Knowledge Distillation for Camouflaged Object Detection

Authors: Bo Cai, Houjie Li, Yanping Yang, Jin Yan

Our work has been accepted by ESWA. The code has already been open sourced.

If you are interested in our work, please do not hesitate to contact us at 765499227@qq.com via email.

🔆 News

[17.9.2025] We have released the code and model checkpoints for CFF-KDNet !

[27.10.2025] CFF-KDNet has been accepted by ESWA !

🕳 Performance

Model	Backbone	Params	CAMO (S_α/Fβω/MAE/Eξ)	COD10K (S_α/Fβω/MAE/Eξ)	NC4K (S_α/Fβω/MAE/Eξ)
CFF-KDNet-E2	EfficientNet-B2	13.06M	0.855/0.812/0.052/0.912	0.866/0.779/0.023/0.922	0.878/0.827/0.034/0.925
CFF-KDNet-R2	Res2Net-50	36.06M	0.832/0.786/0.062/0.886	0.866/0.789/0.023/0.922	0.871/0.824/0.036/0.920
CFF-KDNet-P2	PVTv2-B2	35.62M	0.870/0.833/0.048/0.920	0.889/0.821/0.019/0.937	0.897/0.854/0.030/0.936
CFF-KDNet-S	SMT-T	21.73M	0.880/0.847/0.045/0.929	0.891/0.822/0.019/0.940	0.899/0.861/0.029/0.940
CFF-KDNet-P4	PVTv2-B4	72.81M	0.888/0.859/0.040/0.940	0.900/0.842/0.017/0.949	0.902/0.867/0.027/0.943

🔧 Setup

📖 Prepare Data

COD dataset can be found at google drive.

⛓ Install Requirements

• Python 3.10.16
• torchvision 0.16.2
• Others: pip install -r requirements.txt

🧪 Evaluation

Tercher:
python train.py --config configs/icod_train.py --model-name <MODEL_NAME> --evaluate --load-from <TRAINED_WEIGHT>

Student:
python train_kd.py --config configs/icod_train.py --model-name <MODEL_NAME> --evaluate --load-from <TRAINED_WEIGHT>

🪢 Training

First training teacher model

python train.py --config configs/icod_train.py --pretrained --model-name CFFKDNet_PVTB4

Then training student model

python train_kd.py --config configs/icod_train.py --pretrained --model-name Res2Net50_KD

🖼 Prediction maps

We provide the prediction maps of our CFF-KDNet model COD Dataset here, including: PVT,Res2Net,EfficientNet and SMT.

CFF-KDNet-weights (PVT/SMT/Res2Net) here.

🗂 Citation

If you use CFF-KDNet method in your research or wish to refer to the baseline results published in the Model, please use the following BibTeX entry.

@article{Cai2026CFFKDNet,
  title={CFF-KDNet: Cross-scale feature fusion network with knowledge distillation for camouflaged object detection},
  author={Bo Cai and Houjie Li and Yanping Yang and Jin Yan},
  journal={Expert Systems with Applications},
  volume={299},
  number={C},
  pages={130209},
  year={2026},
  month={Mar},
  doi={10.1016/j.eswa.2025.130209}
}