Dataset Viewer
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
TripVVT-10K: A Large-Scale Triplet Dataset for In-the-Wild Video Virtual Try-On
TripVVT-10K 是目前规模最大、多样性最丰富的“野外”(In-the-wild)视频虚拟试衣(VVT)三元组数据集。它提供了显式的视频级跨服装监督(Cross-garment supervision),旨在解决现有视频试衣模型在真实复杂环境下背景抖动、人体畸变及服装还原度差的问题。
🔥 核心亮点
- 大规模三元组 (Triplet): 包含 10,000 组高质量三元组数据:
<合成原始视频, 参考服装, 真实试衣视频>。 - 真实野外环境 (In-the-Wild): 覆盖弱光、久坐、地铁拥挤人群等挑战性场景,视频分辨率高达 720x1280。
- 逆向训练范式 (Reverse-training): 将真实采集的视频作为 Ground-truth,将合成的换装视频作为模型输入,确保监督信号来自真实物理世界,极大提升了生成的真实感。
- 全方位标注: 提供配套的人体掩码(Human Mask)、姿态序列(DWPose/ViTPose)以及由 LLM 生成的服装描述。
📊 数据集统计与对比
| 数据集 | 模态 | 分辨率 | 场景 | 配对方式 | 数量 |
|---|---|---|---|---|---|
| VVT | Video | 192x256 | Studio | Pair | 791 |
| ViViD | Video | 624x832 | Studio | Pair | 9,700 |
| TripVVT-10K (Ours) | Video | 720x1280 | Outdoor/Real | Triplet | 10,000 |
数据集包含 30 种 服装细分类别,涵盖上装(t-shirt, shirt, jacket等)、下装(pants, skirt, jeans等)和全身装(dress, long dress等)。
license: apache-2.0
- Downloads last month
- 20