---
license: mit
task_categories:
- image-segmentation
tags:
- wsss
- boundary-refinement
- priors
---

# MS COCO 2014 NAMLab Priors for HGA

This dataset repository hosts the pre-computed hierarchical invariant region partition priors for weakly supervised semantic segmentation (WSSS) on MS COCO 2014, generated using the NAMLab framework.

Due to Hugging Face's individual file size limits, the total ~63.69 GB dataset is uploaded as a **split-volume zip archive** (`priors_coco.zip.part1` to `priors_coco.zip.part3`). You must concatenate them locally before extraction.

## 🔗 Main Codebase
The complete implementation, environment setup, and training guidelines are available in our official GitHub repository:
👉 **[Uncertainty-42/HGA](https://github.com/Uncertainty-42/HGA)**

## 📂 Expected Directory Layout
After concatenation and extraction, the folder structure should match the following layout:

```text
MSCOCO/
├── JPEGImages/                
│   ├── train/                 
│   └── val/                   
└── priors/                    
    └── namlab_pt/             # <- Extract here (will automatically create train/ and val/)
```

## 📥 Concatenation & Extraction Guide
Navigate to your local `priors/` directory where the downloaded split parts reside, and run the following commands to reconstruct and extract the dataset:

```bash
# Step 1: Concatenate the split parts into a single valid zip archive
cat priors_coco.zip.part* > priors_coco.zip

# Step 2: Extract the contents into namlab_pt/
unzip priors_coco.zip -d namlab_pt/

# Step 3: Clean up the temporary archive and parts (Optional)
rm priors_coco.zip.part* priors_coco.zip
```

## 📝 Citation & Attribution
If you use these priors in your research, please cite both our work and the foundational NAMLab paper:

```bibtex
@article{zheng2021hierarchical,
  title={Hierarchical Image Segmentation Based on Nonsymmetry and Anti-Packing Pattern Representation Model},
  author={Zheng, Yunping and Yang, Bowen and Sarem, Mudar},
  journal={IEEE Transactions on Image Processing},
  volume={30},
  pages={2408--2421},
  year={2021},
  publisher={IEEE}
}
```