| --- |
| language: |
| - en |
| license: cc-by-nc-4.0 |
| task_categories: |
| - image-classification |
| pretty_name: VPD-100K |
| size_categories: |
| - 100K<n<1M |
| tags: |
| - privacy |
| - computer-vision |
| - object-detection |
| - livestream |
| - visual-privacy |
| - icml-2026 |
| - video privacy |
| paperswithcode_id: null |
| --- |
| |
| # VPD-100K: Towards Generalizable and Fine-grained Visual Privacy Protection |
|
|
| > **Official dataset for the ICML 2026 paper** |
| > |
| > **VPD-100K: Towards Generalizable and Fine-grained Visual Privacy Protection** |
|
|
| 🌐 **Project Page:** https://vpd-100k.github.io/ |
|
|
| 📄 **Paper:** https://arxiv.org/abs/2605.10229 |
|
|
| --- |
|
|
| # Overview |
|
|
| Visual privacy protection has become increasingly important as people continuously share images and live-stream videos online. Existing visual privacy datasets are generally limited in scale, annotation granularity, and scene diversity, making it difficult to train models that generalize to real-world privacy-sensitive scenarios. |
|
|
| VPD-100K is a large-scale benchmark specifically designed for **generalizable visual privacy detection**. It contains **100,000 images**, over **190,000 annotated privacy instances**, and **33 fine-grained privacy categories** covering a broad spectrum of real-world privacy leakage scenarios. |
|
|
| The dataset was introduced in the following paper: |
|
|
| > **VPD-100K: Towards Generalizable and Fine-grained Visual Privacy Protection** |
| > ICML 2026 |
|
|
| --- |
|
|
| # Highlights |
|
|
| - 📷 **100,000 images** |
| - 🎯 **190,000+ annotated privacy instances** |
| - 🏷️ **33 fine-grained categories** |
| - 🌍 Covers diverse real-world environments |
| - 📺 Designed for both image understanding and live-stream privacy protection |
| - 🔍 High-resolution images (over half exceed 1080p) for detecting tiny privacy-sensitive objects |
|
|
| --- |
|
|
| # Privacy Taxonomy |
|
|
| VPD-100K organizes privacy-sensitive content into **four primary domains**. |
|
|
| ## 1. Human Presence |
|
|
| Sensitive human identity information, including different types of faces under diverse environments. |
|
|
| Examples include: |
|
|
| - Adult faces |
| - Child faces |
| - Crowd faces |
| - Partial faces |
|
|
| --- |
|
|
| ## 2. On-Screen Personally Identifiable Information (PII) |
|
|
| Digital information displayed on monitors, phones, tablets, or other screens. |
|
|
| Examples include: |
|
|
| - Passwords |
| - Chat messages |
| - Email addresses |
| - User accounts |
| - Verification codes |
| - Banking interfaces |
|
|
| --- |
|
|
| ## 3. Physical Identifiers |
|
|
| Physical documents and objects that contain sensitive personal information. |
|
|
| Examples include: |
|
|
| - Passport |
| - ID card |
| - Bank card |
| - Boarding pass |
| - Ticket |
| - Driver license |
|
|
| --- |
|
|
| ## 4. Location Indicators |
|
|
| Objects revealing the physical location of users. |
|
|
| Examples include: |
|
|
| - Street signs |
| - Shop signs |
| - Community names |
| - Building names |
| - Address plates |
|
|
| --- |
|
|
| # Dataset Statistics |
|
|
| | Property | Value | |
| |-----------|------:| |
| | Images | 100,000 | |
| | Object Instances | 190,000+ | |
| | Categories | 33 | |
| | Primary Domains | 4 | |
| | Resolution | Over 50% >1080p | |
|
|
| The dataset exhibits realistic characteristics including: |
|
|
| - Long-tail class distribution |
| - Small-object dominance |
| - High visual complexity |
| - Diverse indoor and outdoor environments |
|
|
| These properties make VPD-100K particularly suitable for evaluating privacy detection methods in challenging real-world applications. |
|
|
| --- |
|
|
| # Intended Uses |
|
|
| VPD-100K can be used for: |
|
|
| - Visual privacy detection |
| - Object detection |
| - Privacy-aware computer vision |
| - Live-stream privacy protection |
| - Privacy-preserving AI |
| - Benchmarking privacy detection algorithms |
|
|
| --- |
|
|
| # Ethical Considerations |
|
|
| To avoid exposing real users' sensitive information, privacy-critical scenarios involving digital interfaces (such as face and account information) are reconstructed using high-fidelity simulated environments instead of collecting real personal data whenever possible. |
|
|
| Researchers should ensure that models trained on this dataset are used responsibly and comply with applicable privacy regulations. |
|
|
| --- |
|
|
| # Citation |
|
|
| If you use VPD-100K in your research, please cite: |
|
|
| ```bibtex |
| @inproceedings{vpd100k2026, |
| title={VPD-100K: Towards Generalizable and Fine-grained Visual Privacy Protection}, |
| author={Hu, Xiaobin and Zuo, Enpu and Hu, Lanping and Yang, Kaiwen and Liao, Dianshu and Zhang, Tianyi and Yin, Bo and Zhou, Yinsi and Pan, Shidong and Sun, Xiaoyu}, |
| booktitle={Proceedings of the International Conference on Machine Learning (ICML)}, |
| year={2026} |
| } |
| ``` |
|
|
| --- |
|
|
| # License |
|
|
| Please refer to the official project page for the latest licensing information. |
|
|
| --- |
|
|
| # Links |
|
|
| - 🌐 Project Page: https://vpd-100k.github.io/ |
| - 📄 Paper: https://arxiv.org/abs/2605.10229 |
| - 🤗 Hugging Face Dataset: https://huggingface.co/datasets/XiaoyuSunANU/Visual_Privacy_Dataset |
