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

This repository contains the implementation of HDPNet, a state-of-the-art model for camouflaged object detection. The model uses a combination of Hourglass Vision Transformer and Dual-Path Feature Pyramid architecture to achieve high accuracy in detecting camouflaged objects.

Features

• Hourglass Vision Transformer architecture
• Dual-Path Feature Pyramid for multi-scale feature extraction
• High accuracy on various camouflaged object detection datasets
• Easy-to-use inference script with visualization capabilities

Installation

1. Create a conda environment:

conda create -n HDPNet python=3.8
conda activate HDPNet

2. Install required packages:

pip install -r requirements.txt

Model Weights

Download the pretrained model weights from the following link:

• HDPNet Weights

After downloading:

1. Create a weights directory in the project root:

mkdir weights

2. Place the downloaded weights file in the weights directory:

mv /path/to/downloaded/weights.pth weights/

Usage

Inference

To run inference on test images:

1. Place your test images in the testdata directory:

mkdir -p testdata
# Copy your images to testdata/

2. Run the inference script:

python inference.py

This will:

• Process all images in the testdata directory
• Save predictions in results/predictions
• Create side-by-side visualizations in results/visualizations
• Generate a video visualization in results/visualization.mp4

Example Results

Here are some example results showing the original images and their predictions:

Left: Original Image, Right: Prediction

Left: Original Image, Right: Prediction

Input/Output Format

• Input: RGB images (any size, will be resized to 384x384)
• Output:
  • Binary mask predictions (0-255 grayscale)
  • Side-by-side visualizations
  • Video compilation of results

Directory Structure

HDPNet/
├── weights/              # Model weights
├── testdata/            # Input test images
├── results/
│   ├── predictions/     # Binary mask predictions
│   ├── visualizations/  # Side-by-side visualizations
│   └── visualization.mp4  # Video compilation
├── inference.py         # Inference script
└── requirements.txt     # Dependencies

Results

The model achieves state-of-the-art performance on various camouflaged object detection datasets. For detailed quantitative results, please refer to the paper.

Citation

If you use this code in your research, please cite:

@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}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.