Update README.md

README.md

@@ -1,3 +1,90 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+---
+# 3DFDReal: 3D Fashion Data from Real-world
+**Author(s):** ETRI Media Intellectualization Research Section  
+**Contact:** kusses@etri.re.kr  
+
+![Teaser](figure/teaser.png)
+
+**3DFDReal** is a real-world fashion dataset designed for 3D vision tasks such as segmentation, reconstruction, rigging, and metaverse deployment. Captured from multi-view 2D video with 4K resolution and 60fps, it includes both single fashion items and composite combinations worn by mannequins.
+
+## 🔍 Overview
+
+This dataset aims to bridge the gap between high-quality 3D fashion modeling and practical deployment in virtual platforms such as **ZEPETO**. It contains over 1,000 3D point clouds with rich metadata including class labels, gender, texture, pose type, and structured segmentations.
+
+---
+
+## 📸 Data Collection Pipeline
+
+Our data collection follows a structured 4-stage pipeline:
+
+1. **Item Preparation:** Single items or grouped sets of fashion assets (e.g., shoes, tops, accessories) are selected and labeled.
+2. **Shooting Setup:** Mannequins or suspended items are filmed using an iPhone 13 Pro from multi-view angles.
+3. **3D Ground Truth Generation:** The videos are converted to point clouds and manually segmented using professional tools.
+4. **Application Output:** Final 3D assets are rigged and validated on avatar platforms (e.g., ZEPETO), ready for deployment.
+
+---
+
+## 📊 Dataset Statistics
+
+### Class Frequency
+
+![Class Frequency](figure/fashion_class_distribution.png)
+
+Pants, Sweatshirts, and Jeans are among the most frequent classes in the dataset.
+
+### Combination Metadata
+
+![Combination Overview](figure/combination_overview_stats.png)
+
+As shown:
+- Most mannequin-wear combinations contain **4 fashion items**.
+- Gender distribution is relatively balanced.
+- **T-pose** is selectively used for rigging compatibility, while **upright** poses dominate general recording.
+
+---
+
+## 📐 Data Format
+
+- `.ply` format point clouds with vertex color
+- JSON metadata per asset:
+  - `label_str`: class label
+  - `gender`, `pose`, `type`
+  - `wnlemmas`: fine-grained attributes
+- Segmentation masks (optional per task)
+- Textured mesh reconstructions (via Blender + LCM)
+
+---
+
+## 🧪 Benchmarks
+
+Baseline models using [**SAMPart3D**](https://yhyang-myron.github.io/SAMPart3D-website/) demonstrate high segmentation quality (mIoU: 0.9930) but show varying average precision (AP) across classes.
+
+![Result Visulization](figure/seg_tuning.png)
+
+---
+
+
+## 💻 Use Cases
+
+- **Virtual try-on**
+- **Metaverse asset creation**
+- **Pose-aware segmentation**
+- **Avatar rigging & deformation simulation**
+
+---
+
+## 📄 Citation
+
+If you use this dataset, please cite:
+
+```bibtex
+@inproceedings{lim2025volme,
+  title={3DFDReal: 3D Fashion Data from Real-world},
+  author={Jiyoun Lim, Jungwoo Son, Alex Lee, Sun-Joong Kim, Nam Kyung Lee, Won-Joo Park},
+  booktitle={TBD},
+  year={2025}
+}
+```
+