PICD: Photographic Image Composition Dataset

Can Machines Understand Composition? A Dataset and Benchmark for Photographic Image Composition Embedding and Understanding
📌 CVPR 2025 Highlight

📄 CVPR 2025 Paper
📑 Supplementary Appendix

---

📌 Overview

PICD is a large-scale dataset for photographic image composition analysis, currently containing 49,123 high-quality images annotated with 24 composition categories.

This dataset is intended to support the evaluation and advancement of composition learning in AI models. It is applicable to a wide range of tasks, including aesthetic quality assessment, composition-aware image cropping, and more. We encourage researchers and practitioners to explore creative uses of PICD.

The composition label system is structured along two axes:

• Element Types: Points, Lines, and Shapes (inspired by Kandinsky’s principles)
• Arrangement Patterns: Rule of Thirds, Centered, Diagonal, Vertical, Horizontal, Triangle, C-curve, O-curve, S-curve, Radial, Dense, Scatter, etc.

📖 For detailed category definitions and label design, please refer to the Appendix.

Figure 1. The PICD label system is structured along two axes: element types and arrangement patterns. Column 1 (green) shows arrangement types; Columns 2–4 show compositional element types. Categories are numbered 1–24 with abbreviations in blue. Red boxes indicate merged categories; blue strikethroughs mark excluded ones due to low frequency. Column 5 highlights dominant compositional factors.

Figure 2. Sample images from the 24 composition categories in PICD. Category abbreviations appear in blue parentheses.

---

📊 Dataset Information

PICD is actively maintained and will continue to be expanded. The current release includes:

• ✅ 49,123 images
• ✅ Verified composition category annotations
• ⏳ Negative samples (images not conforming to any predefined category) — coming soon
• ⏳ Composition quality scores — coming soon
• ⏳ Textual composition descriptions — coming soon

---

🔗 Download

PICD consists of both image files and annotations.

1. Images

1) Direct Access:

Image download is divided into two parts based on licensing:

Part 1: Images with redistribution permission

• This includes 44,577 images from open platforms and redistributable open-source datasets (e.g., Unsplash, Pexels).
• These images can be downloaded directly via the following link:
  👉 Baidu Netdisk download link (with code 1517) 👉 Google Drive download link

Part 2: Images requiring user-side access

• This includes 4546 images from public datasets that do not permit redistribution (e.g., AVA).
• For these, we provide a mapping file that links each PICD-assigned image ID to the original dataset image ID or URL. You may download the original images from their respective sources using this mapping:
  👉 Image ID Mapping File

2) Alternative Access:
If you prefer to request both parts directly via email (especially for convenience or if you encounter access issues), please send a message to picd2025@outlook.com with your affiliation and intended use. We will respond with the download links after reviewing your request.

Accessing or using the dataset in any way implies agreement to the
📄 PICD Dataset Terms of Use (PDF)

2. Annotations

• ✏️ Image Annotation File
  👉 Download Annotations
  This CSV file contains the following fields:
  • img_id: The PICD image ID
  • category_id: Index of the composition category (1–24)
  • category_abbre: Abbreviated category label (as shown in Figure 2)
  • category_full_name: Full name of the composition category

The mapping among category_id, category_abbre, and category_full_name follows the structure shown in Figure 1.

---

📄 License and Terms

PICD is released under the
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

All users must also agree to the dataset-specific terms of use:
📄 PICD Dataset Terms of Use (PDF)

---

🔧 Citation

If you use PICD in your research, please cite:

@inproceedings{zhao2025can,
  title={Can Machines Understand Composition? Dataset and Benchmark for Photographic Image Composition Embedding and Understanding},
  author={Zhao, Zhaoran and Lu, Peng and Zhang, Anran and Li, Peipei and Li, Xia and Liu, Xuannan and Hu, Yang and Chen, Shiyi and Wang, Liwei and Guo, Wenhao},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={14411--14421},
  year={2025}
}