weixiny0408/vtonqa

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

# VTONQA: A Multi-Dimensional Quality Assessment Dataset for Virtual Try-On

## 📌 Overview

VTONQA is a benchmark dataset for evaluating the quality of virtual try-on (VTON) results.
It provides multi-dimensional human annotations to assess realism, alignment, and perceptual quality of synthesized try-on images.

Compared with existing VTON datasets, VTONQA focuses on **quality assessment** rather than generation, enabling standardized evaluation across different methods.

---

## 🎯 Key Features

- Multi-dimensional quality scores (e.g., realism, alignment, overall)
- Human-annotated evaluation labels
- Designed for benchmarking VTON models
- Supports both regression and ranking tasks

---

## 🧾 Data Description

### Input Data (`dataset/`)

The input data follows a standard virtual try-on (VTON) pipeline and consists of multiple modalities describing human appearance, clothing, and structural information.

#### Directory Structure

```
dataset/
├── image/ # original person images
├── cloth/ # clothing images
├── cloth_mask/ # binary masks for clothing
├── agnostic/ # person images with clothing removed
├── agnostic_mask/ # masks for removed clothing regions
├── image-parse/ # human parsing (semantic segmentation)
├── densepose_gray/ # DensePose representation
├── openpose_img/ # visualized human pose
├── openpose_json/ # pose keypoints (JSON format)
├── test_pairs_unpaired.txt
```

---

#### Components Description

- `image/`:
Original images of persons wearing their initial clothing.
These provide identity, pose, and background information.

- `cloth/`:
Standalone images of garments to be transferred onto the person.

- `cloth_mask/`:
Binary masks for clothing images, where foreground indicates the clothing region.
Used to improve garment alignment and shape modeling.

- `agnostic/`:
Person images with original clothing removed while preserving identity-related regions such as face and hair.
Serves as a clean input for try-on generation.

- `agnostic_mask/`:
Masks indicating the regions where clothing has been removed.
Helps guide the generation of new garments.

- `image-parse/`:
Human parsing maps that segment the body into semantic regions (e.g., upper body, arms, background).
Provides structural guidance for clothing placement.

- `densepose_gray/`:
DensePose representations encoding fine-grained human body geometry.
Enables better alignment between clothing and body surface.

- `openpose_img/`:
Visualized human skeletons representing pose structure.

- `openpose_json/`:
2D keypoint coordinates of human pose in JSON format.
Provides precise numerical pose information.

- `test_pairs_unpaired.txt`:
Defines the pairing between person images and clothing items.

Example:
```
167_whole.jpg d1.jpg
283_whole.jpg d1.jpg
85_whole.jpg d1.jpg
195_whole.jpg d1.jpg
```

---

### Generated Results (`result/`)

The `result/` directory contains outputs generated by multiple state-of-the-art virtual try-on (VTON) methods, as well as predefined evaluation splits for different quality dimensions.

---
#### 📂 Directory Structure
```
result/
├── CAT-DM/
├── CatV2TON/
├── DS-VITON/
├── FS-VTON/
├── keling/
├── ladivton/
├── LINKFOX/
├── OOTDiffusion/
├── StableVITON/
├── TPD/
├── VITON-HD/
│
├── body_compatibility_test.txt
├── body_compatibility_train.txt
├── clothing_fit_test.txt
├── clothing_fit_train.txt
├── overall_quality_test.txt
├── overall_quality_train.txt
```
---

#### 🧾 Method Outputs

Each subfolder corresponds to a specific VTON method:

- `CAT-DM/`, `CatV2TON/`, `DS-VITON/`, `FS-VTON/`, `VITON-HD/`, etc.

Each folder contains generated try-on images for the same set of input pairs.

---

#### 📊 Evaluation Splits

The dataset provides predefined splits for different evaluation dimensions:

- `body_compatibility_*.txt`
→ Evaluates how well the clothing matches the human body structure

- `clothing_fit_*.txt`
→ Evaluates how naturally the clothing fits the person

- `overall_quality_*.txt`
→ Evaluates overall perceptual quality

Each split includes:

- `*_train.txt`: training subset (for learning-based evaluators)
- `*_test.txt`: testing subset (for evaluation)

---

#### 📄 Split File Format

Each `.txt` file contains a list of sample identifiers.

Example:
```
523_lower_tr10_DS-VITON.png, 54.447846865424935
85_lower_tr6_OOTDiffusion.png, 37.64992522777412
510_upper_shirt10_ladivton.png, 56.459661598182684
```
---

---
license: CC BY 4.0
---

# VTONQA：面向虚拟试穿（Virtual Try-On，以下简称VTON）的多维度质量评估数据集

## 📌 概述

VTONQA是一款用于评估虚拟试穿结果质量的基准数据集，其提供多维度人工标注以评估合成试穿图像的真实感、对齐度与感知质量。

相较于现有VTON数据集，VTONQA聚焦于**质量评估**而非图像生成，可实现不同模型方法间的标准化评测。

---

## 🎯 核心特性

- 多维度质量评分（如真实感、对齐度、整体质量）
- 人工标注的评估标签
- 专为VTON模型基准测试设计
- 支持回归与排序两类评测任务

---

## 🧾 数据说明

### 输入数据（`dataset/`）

输入数据遵循标准虚拟试穿流水线，包含描述人体外观、服装及结构信息的多模态数据。

#### 目录结构

dataset/
├── image/ # 原始人物图像
├── cloth/ # 服装图像
├── cloth_mask/ # 服装二值掩码
├── agnostic/ # 无服装人物图像
├── agnostic_mask/ # 无服装区域掩码
├── image-parse/ # 人体解析（语义分割）结果
├── densepose_gray/ # DensePose 表征
├── openpose_img/ # 可视化人体姿态
├── openpose_json/ # 姿态关键点（JSON格式）
├── test_pairs_unpaired.txt

---

#### 组件说明

- `image/`：
人物身着初始服装的原始图像，包含身份、姿态与背景信息。

- `cloth/`：
待迁移至人物身上的独立服装图像。

- `cloth_mask/`：
服装图像的二值掩码，前景区域表示服装范围，用于提升服装对齐与形状建模效果。

- `agnostic/`：
移除原有服装但保留面部、毛发等身份相关区域的人物图像，作为试穿生成的干净输入。

- `agnostic_mask/`：
标记已移除服装区域的掩码，用于引导新服装的生成。

- `image-parse/`：
将人体分割为语义区域（如上半身、手臂、背景）的人体解析图，为服装摆放提供结构引导。

- `densepose_gray/`：
编码细粒度人体几何信息的DensePose表征，可实现服装与体表的更精准对齐。

- `openpose_img/`：
可视化的人体骨骼图，用于表示姿态结构。

- `openpose_json/`：
JSON格式的二维人体姿态关键点坐标，提供精准的数值化姿态信息。

- `test_pairs_unpaired.txt`：
定义人物图像与服装配对关系的文件。
示例：

167_whole.jpg d1.jpg
283_whole.jpg d1.jpg
85_whole.jpg d1.jpg
195_whole.jpg d1.jpg

---

### 生成结果（`result/`）

`result/` 目录包含多个前沿虚拟试穿方法生成的试穿结果，以及针对不同质量维度的预定义评测划分集。

---

#### 📂 目录结构

result/
├── CAT-DM/
├── CatV2TON/
├── DS-VITON/
├── FS-VTON/
├── keling/
├── ladivton/
├── LINKFOX/
├── OOTDiffusion/
├── StableVITON/
├── TPD/
├── VITON-HD/
│
├── body_compatibility_test.txt
├── body_compatibility_train.txt
├── clothing_fit_test.txt
├── clothing_fit_train.txt
├── overall_quality_test.txt
├── overall_quality_train.txt

---

#### 🧾 模型输出

每个子文件夹对应一种特定的VTON方法：
- `CAT-DM/`、`CatV2TON/`、`DS-VITON/`、`FS-VTON/`、`VITON-HD/` 等。
每个文件夹包含针对同一组输入配对生成的试穿图像。

---

#### 📊 评测划分集

该数据集提供针对不同评测维度的预定义划分集：
- `body_compatibility_*.txt`：评估服装与人体结构的匹配程度
- `clothing_fit_*.txt`：评估服装在人物身上的贴合自然度
- `overall_quality_*.txt`：评估整体感知质量

每个划分集包含：
- `*_train.txt`：训练子集（用于基于学习的评估器）
- `*_test.txt`：测试子集（用于正式评测）

---

#### 📄 划分文件格式

每个`.txt`文件包含样本标识符列表。
示例：

523_lower_tr10_DS-VITON.png, 54.447846865424935
85_lower_tr6_OOTDiffusion.png, 37.64992522777412
510_upper_shirt10_ladivton.png, 56.459661598182684