HDPNet: Hourglass Vision Transformer with Dual-Path Feature Pyramid for Camouflaged Object Detection (WACV2025)

Usage

The training and testing experiments are conducted using PyTorch with an NVIDIA A100-SXM of 40 GB Memory.

1. Prerequisites

Note that HDPNet is only tested on Ubuntu OS with the following environments.

• Creating a virtual environment in terminal: conda create -n HDPNet python=3.8.
• Installing necessary packages: pip install -r requirements.txt

2. Downloading Training and Testing Datasets

• Download the training set (COD-TrainDataset) used for training
• Download the testing sets (COD10K-test + CAMO-test + CHAMELEON + NC4K ) used for testing

3. Training Configuration

• The pretrained model(PVT2) is stored in Google Drive and Baidu Drive (g3ea). After downloading, please change the file path in the corresponding code.
• Run train.sh to train.

4. Testing Configuration

Our well-trained model is stored in Google Drive and Baidu Drive (gv9n). After downloading, please change the file path in the corresponding code.

5. Evaluation

• Evaluate HDPNet: After configuring the test dataset path, run hpvt_eval.sh in the run_slurm folder for evaluation.
• PR-Curves: We provide the code for obtaining PR-Curves through detection results. Please refer to 'PR_Curve.py'.
• Super- and Sub-Classes: To evaluate the performance of each method on COD10K superclasses and subclasses through detection results, please refer to 'class_eval.py'.

6. Results download

The prediction results of our HDPNet are stored on Google Drive. Please check.

7. Quantitative Results

Our final results, which perform very well on the COD10K dataset (contains a lot of small objects and detailed labeling of the objects' fine boundaries).

we adopt five kinds of evaluation metrics: S-measure($S_m$), weighted F-measure($F_\beta^\omega$), adaptive F-measure($F_\beta^{adp}$), mean F-measure($F_\beta^{mean}$),max F-measure($F_\beta^{max}$), adaptive E-measure($E_\phi^{apd}$), mean E-measure($E_\phi^{mean}$), max E-measure ($E_\phi^{max}$), and mean absolute error($\mathcal{M}$)

Dataset	$S_m \uparrow$	$F_\beta^\omega \uparrow$	$F_\beta^{adp} \uparrow$	$F_\beta^{mean} \uparrow$	$F_\beta^{max} \uparrow$	$E_\phi^{adp} \uparrow$	$E_\phi^{mean} \uparrow$	$E_\phi^{max} \uparrow$	$\mathcal{M} \downarrow$
CAMO	0.893	0.851	0.848	0.870	0.890	0.932	0.934	0.948	0.040
CHAMELEON	0.921	0.861	0.849	0.874	0.902	0.943	0.947	0.970	0.021
COD10K	0.888	0.794	0.770	0.820	0.852	0.915	0.925	0.951	0.020
NC4K	0.902	0.850	0.845	0.871	0.891	0.931	0.934	0.950	0.029

8. Qualitative Results

Quantitative results in several typical complex situations, including occlusion, small objects, multiple objects, and object boundaries.

Citation

@inproceedings{he2025hdpnet,
  title={HDPNet: Hourglass Vision Transformer with Dual-Path Feature Pyramid for Camouflaged Object Detection},
  author={He, Jinpeng and Liu, Biyuan and Chen, Huaixin},
  booktitle={2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages={8638--8647},
  year={2025},
  organization={IEEE}
}