[Database: MIQD-2.5M] Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models

Xiaoqi Wang, Yun Zhang, Weisi Lin
Sun Yat-sen University | Nanyang Technological University | 📖 Paper | 👨‍💻 Code

📊 Overview

• 2.5M degraded images from 10,000 originals
• 3 vision tasks:
  • Image Classification (ImageNet)
  • Object Detection (COCO)
  • Instance Segmentation (COCO)
• 10 distortion types × 5 levels × 3 region patterns
• 75 models used for generating quality labels
• MMOS labels: consistency + accuracy + composite score

📂 Dataset Structure

MIQD_2.5M/
├── miqa_cls/
│   ├── images/
│   ├── labels/
│   ├── src_images/
│   └── additional_info/
├── miqa_det/
│   ├── images/
│   ├── labels/
│   ├── src_images/
│   └── additional_info/
├── miqa_ins/
│   ├── labels/

📥 Download & Extraction

The dataset is split into .tar.part_* files for efficient transfer.

Step 1: Merge parts

cat classification_images.tar.part_* > classification_images.tar
cat classification_labels.tar.part_* > classification_labels.tar
cat classification_src.tar.part_* > classification_src.tar
cat classification_info.tar.part_* > classification_info.tar

cat detection_images.tar.part_* > detection_images.tar
cat detection_labels.tar.part_* > detection_labels.tar
cat detection_src.tar.part_* > detection_src.tar
cat detection_info.tar.part_* > detection_info.tar

cat segmentation_labels.tar.part_* > segmentation_labels.tar

Step 2: Extract

# classification
mkdir -p miqa_cls
tar -xf classification_images.tar -C miqa_cls/
tar -xf classification_labels.tar -C miqa_cls/
tar -xf classification_src.tar -C miqa_cls/
tar -xf classification_info.tar -C miqa_cls/

# detection
mkdir -p miqa_det
tar -xf detection_images.tar -C miqa_det/
tar -xf detection_labels.tar -C miqa_det/
tar -xf detection_src.tar -C miqa_det/
tar -xf detection_info.tar -C miqa_det/

# segmentation
mkdir -p miqa_ins
tar -xf segmentation_labels.tar -C miqa_ins/

🧪 Annotation (MMOS)

Each image is labeled with:

• Accuracy (task performance)
• Consistency (prediction stability)
• Composite score (MMOS)

📸 Examples

Database Examples

Sample illustration showcasing the corresponding PSNR, SSIM, and LPIPS values, along with the predicted Consistency, Accuracy, and Composite scores, and their respective ground-truth Consistency, Accuracy, and Composite scores. Panels (a)–(d), (e)–(g), and (h)–(j) present example images along with MIQA-related scores for image classification under UD, object detection under ROI-DD, and instance segmentation under BG-DD, respectively. Numbers in parentheses indicate distortion severity in ROI and background regions. Note: Lower LPIPS indicates higher perceptual quality, whereas higher values are preferred for other metrics.

⚠️ Notes

• Instance segmentation shares images with detection
• Labels are task-specific
• Dataset size: ~984 GB

📚 Citation

@article{wang2025miqa,
  title={Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models},
  author={Wang, Xiaoqi and Zhang, Yun and Lin, Weisi},
  journal={arXiv preprint arXiv:2508.19850},
  year={2025}
}

📄 License

MIT License

📮 Contact

wangxq79@mail2.sysu.edu.cn