24 |
25 |
27 | 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. 28 |
29 | 30 |
31 |
32 |
34 | Figure 2. Sample images from the 24 composition categories in PICD. Category abbreviations appear in blue parentheses. 35 |
36 | 37 | --- 38 | 39 | ## 📊 Dataset Information 40 | 41 | PICD is actively maintained and will continue to be expanded. The current release includes: 42 | 43 | - ✅ **49,123 images** 44 | - ✅ Verified composition category annotations 45 | - ⏳ Negative samples (images not conforming to any predefined category) — *coming soon* 46 | - ⏳ Composition quality scores — *coming soon* 47 | - ⏳ Textual composition descriptions — *coming soon* 48 | 49 | --- 50 | 51 | ## 🔗 Download 52 | 53 | PICD consists of both image files and annotations. 54 | 55 | ### 1. Images 56 | 57 | **1) Direct Access:** 58 | 59 | Image download is divided into two parts based on licensing: 60 | 61 | **Part 1: Images with redistribution permission** 62 | - This includes 44,577 images from open platforms and redistributable open-source datasets (e.g., Unsplash, Pexels). 63 | - These images can be downloaded directly via the following link: 64 | 👉 **[Baidu Netdisk download link (with code 1517)](https://pan.baidu.com/s/17dWynHJzCTi3fe5dy0v8Fw?pwd=1517)** 65 | 👉 **[Google Drive download link](https://drive.google.com/drive/folders/10MyHEtoOj61n7cIC6jAMZOXK7K4hgv4e?usp=drive_link)** 66 | 67 | **Part 2: Images requiring user-side access** 68 | - This includes 4546 images from public datasets that do not permit redistribution (e.g., AVA). 69 | - 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: 70 | 👉 **[Image ID Mapping File](https://github.com/CV-xueba/PICD_ImageComposition/blob/main/image_link_public.csv)** 71 | 72 | **2) Alternative Access:** 73 | 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](picd2025@outlook.com) with your affiliation and intended use. We will respond with the download links after reviewing your request. 74 | 75 | > Accessing or using the dataset in any way implies agreement to the 76 | > 📄 [PICD Dataset Terms of Use (PDF)](https://github.com/CV-xueba/PICD_ImageComposition/blob/main/PICD_Dataset_Terms_of_Use.pdf) 77 | 78 | ### 2. Annotations 79 | 80 | - ✏️ **Image Annotation File** 81 | 👉 **[Download Annotations](https://github.com/CV-xueba/PICD_ImageComposition/blob/main/labels_PICD.csv)** 82 | This CSV file contains the following fields: 83 | - `img_id`: The PICD image ID 84 | - `category_id`: Index of the composition category (1–24) 85 | - `category_abbre`: Abbreviated category label (as shown in Figure 2) 86 | - `category_full_name`: Full name of the composition category 87 | 88 | The mapping among `category_id`, `category_abbre`, and `category_full_name` follows the structure shown in **Figure 1**. 89 | 90 | --- 91 | 92 | ## 📄 License and Terms 93 | 94 | PICD is released under the 95 | **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)**. 96 | 97 | All users must also agree to the dataset-specific terms of use: 98 | 📄 [PICD Dataset Terms of Use (PDF)](https://github.com/CV-xueba/PICD_ImageComposition/blob/main/PICD_Dataset_Terms_of_Use.pdf) 99 | 100 | --- 101 | 102 | ## 🔧 Citation 103 | 104 | If you use PICD in your research, please cite: 105 | 106 | ```bibtex 107 | @inproceedings{zhao2025can, 108 | title={Can Machines Understand Composition? Dataset and Benchmark for Photographic Image Composition Embedding and Understanding}, 109 | 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}, 110 | booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference}, 111 | pages={14411--14421}, 112 | year={2025} 113 | } 114 | --------------------------------------------------------------------------------