---
license: cc-by-4.0
---

# 🧵 3DFDReal: 3D Fashion Data from the Real World

**ETRI Media Intellectualization Research Section, ETRI**

![Teaser](figures/teaser.png)

**3DFDReal** is a real-world fashion dataset tailored for 3D vision tasks such as **segmentation**, **reconstruction**, **rigging**, and **deployment in metaverse platforms**. Captured from high-resolution multi-view 2D videos (4K @ 60fps), the dataset includes both **individual fashion items** and **combined outfits** worn by mannequins.

---

## 🔍 Overview

**3DFDReal** bridges the gap between high-quality 3D fashion modeling and practical deployment in virtual environments, such as **ZEPETO**. It features over **1,000 3D point clouds**, each enriched with detailed metadata including:

- Class labels  
- Gender and pose type  
- Texture and semantic attributes  
- Structured segmentations

This dataset provides a foundation for advancing research in **pose-aware 3D understanding**, **avatar modeling**, and **digital twin applications**.

---

## 🎥 Data Collection Pipeline

The dataset is built through a structured four-stage pipeline:

1. **Asset Selection:** Fashion items (e.g., shoes, tops, accessories) are selected and tagged individually or in sets.
2. **Recording Setup:** Items or mannequins are filmed using an iPhone 13 Pro from multi-view angles for 3D reconstruction.
3. **3D Ground Truth Generation:** Videos are converted into colored point clouds and manually segmented using professional 3D labeling tools.
4. **Application & Validation:** Assets are rigged and tested in avatar environments like ZEPETO for deployment readiness.

---

## 📊 Dataset Statistics

### 📈 Class Distribution

![Used Fasion Item Count for 3D Dataset](figures/fashion_class_distribution.png)
**Pants** and **sweatshirts** are used more than other fashion items.


![Fashion Item Count in Mannequin-wear Combinations](figures/Count appears in Combination.png)
**Sneakers** and **Pants** are the most frequent fashion items in Mannequin-wear combinations.

### 👚 Combination Metadata

![Combination Overview](figures/combination_overview_stats.png)

Key observations:
- Most mannequin outfits contain **four distinct fashion items**.
- Gender distribution is balanced across combinations.
- **T-poses** are selectively used for rigging, while **upright poses** dominate standard recordings.

---

## 📁 Dataset Structure

```
dataset/
├── PointCloud_Asset/
├── Video_Asset/
├── Label_Asset/
├── PointCloud_Combine/
├── Video_Combine/
├── Label_Combine/
└── meta/
    ├── asset_meta.json
    ├── combination_meta.json
    ├── train_combination_meta.json
    ├── val_combination_meta.json
    ├── test_combination_meta.json
    └── label_map.csv
```

---

## 📦 Data Description

### 🔹 Individual Asset Files

- **PointCloud_Asset/**  
  Contains raw point clouds of individual clothing or body parts in `.ply` format.

- **Video_Asset/**  
  Rendered 3D videos of individual assets showing different rotations or views.

- **Label_Asset/**  
  Label information (e.g., category, class ID) for each individual asset.

---

### 🔹 Combined Assets (Mannequin Representations)

- **PointCloud_Combine/**  
  Combined point clouds representing mannequins wearing multiple assets. Split into `train`, `val`, and `test` sets.

- **Video_Combine/**  
  Rendered 3D videos of mannequins with asset combinations. Also split into `train`, `val`, and `test`.

- **Label_Combine/**  
  Label files corresponding to the combined point clouds and videos.

---

## 🗂️ Metadata Files (`meta/`)
Each mata contains this detailed information:
  - `label_str`: class name  
  - `gender`, `pose`, `type`  
  - `wnlemmas`: fine-grained semantic tags

- **asset_meta.json**:  
  Metadata for individual assets
  
- **combination_meta.json**:  
  Metadata for all combinations

  
- **train_combination_meta.json**, **val_combination_meta.json**, **test_combination_meta.json**:  
  Define which combinations belong to each data split.

- **label_map.csv**:  
  Maps label for the first data acquisition fullid and the second acquisition label fullid.

---

## 🧪 Benchmarks

### 3D object segmentation
The Baseline model using [**SAMPart3D**](https://yhyang-myron.github.io/SAMPart3D-website/) demonstrates high segmentation quality (mIoU: 0.9930) but shows varying average precision (AP) across classes.

![3D object segmentation with SAMPart3D](figures/seg_tuning.png)


### 3D data reconstruction
The baseline models are [**DDPM-A**](https://github.com/lucidrains/denoising-diffusion-pytorch) diffusion-based probabilistic model for the generation task, and [**SVD-SVDFormer**](https://github.com/czvvd/SVDFormer_PointSea) for the completion task.
Performance is measured using Chamfer Distance (CD), Density-aware Chamfer Distance (DCD), and F1-Score (F1). For DDPM, sampled point clouds are shuffled without considering the sampling ratio 𝑛, and the performance of DDPM is measured with CD. DDPM shows 0.628±0.887 of the average CD. 

![3D data re onstruction example](figures/sampledPC.png)

---

## 💻 Use Cases

- **Virtual try-on**
- **Metaverse asset creation**
- **Pose-aware segmentation**
- **Avatar rigging & deformation simulation**


## 📃 License

CC-BY 4.0

---

## 📚 Citation

```bibtex
@misc{3DFDReal,
  title={3DFDReal: 3D Fashion Data from the Real World},
  author={Jiyoun Lim, Jungwoo Son, Alex Lee, Sun-Joong Kim, Nam Kyung Lee, Won-Joo Park},,
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/kusses/3DFDReal}},
}
```

---

## 💬 Contact

For questions, please reach out via [kusses@etri.re.kr] or use the Discussions tab on Hugging Face.