Deep Fashion3D：从单个图像重建3D服装的数据集和基准

Deep Fashion3D数据集是3D服装模型迄今为止最大的集合。提议的数据集具有服务器吸引力的特征： 首先，Deep Fashion3D包含从真实服装重构的2078个模型，涵盖10个不同类别和563个服装实例。据我们所知，Deep Fashion3D涵盖的服装类别比其他专门针对3D服装的公共可用数据集要多得多。 其次，Deep Fashion3D包含丰富的注释，其中包括3D特征线，3D身体姿势和相应的经过校准的多视图真实图像。值得一提的是，Deep Fashion3D是第一个提供针对3D量身定制的特征线注释的数据集，可以为服装推理相关任务（例如3D服装重构，分类，检索等）提供强大的先验。 第三，Deep Fashion3D中的每件衣服都是随机摆放的，以增加真实衣服的多样性。 Table 1：显示Deep Fashion3D每种服装类别的统计信息。 Deep Fashion3D包含从真实服装重构的2078个模型，涵盖10个不同类别。 Table 1: Deep Fashion3D的每个服装类别的统计信息. 资料范例 ： 图2显示了Deep Fashion3D数据集中每种服装的数据示例。 Deep Fashion3D数据集中的所有模型都是从真实服装中重建的 图3展示了Deep Fashion3D数据集注释的快速浏览。为了促进对3D服装推理和重构任务的未来研究，我们提供了Deep Deep 3D数据集以及丰富的注释，该数据集由以下组件组成： 特征线注释：在Deep Fashion3D数据集中提出了一种针对3D游戏量身定制的新颖特征线注释。类似于面部地标，特征线表示感兴趣的最突出特征，例如开放的边界，领口，袖口，腰部等，可以为忠实地重建3D服装提供强大的先验。 校准的多视图真实图像：对于每种3D服装模型，我们提供相应的校准的多视图真实图像，这对于将训练后的模型推广到野外图像至关重要。 3D姿势：对于每个3D服装模型，我们将其标记为由SMPL系数表示的3D姿势。由于人体和衣服之间的高度耦合特性，我们认为标记的3D姿势可能有助于推断衣服模型的整体形状和姿势相关的变形。 图3：快速浏览Deep Fashion3D数据集的注释

The Deep Fashion3D dataset is the largest collection of 3D clothing models to date. The proposed dataset has several highly appealing characteristics:
First, Deep Fashion3D contains 2078 models reconstructed from real garments, covering 10 distinct categories and 563 clothing instances. To the best of our knowledge, Deep Fashion3D encompasses far more clothing categories than other publicly available datasets specifically targeting 3D clothing.
Second, Deep Fashion3D includes rich annotations, such as 3D feature lines, 3D body poses, and corresponding calibrated multi-view real-world images. Notably, Deep Fashion3D is the first dataset to provide feature line annotations tailored for 3D clothing, which can serve as strong priors for tasks related to clothing reasoning, such as 3D clothing reconstruction, classification, retrieval, and so on.
Third, each garment in Deep Fashion3D is randomly posed to increase the diversity of real-world clothing.
Table 1: Statistics for each clothing category in Deep Fashion3D.
Deep Fashion3D contains 2078 models reconstructed from real garments, covering 10 distinct categories.
Table 1: Statistics for each clothing category in Deep Fashion3D.
Data Examples: Figure 2 shows data examples for each type of clothing in the Deep Fashion3D dataset. All models in the Deep Fashion3D dataset are reconstructed from real garments. Figure 3 presents a quick overview of the annotations in the Deep Fashion3D dataset.
To facilitate future research on 3D clothing reasoning and reconstruction tasks, we release the Deep Fashion3D dataset along with rich annotations, which consists of the following components:
1. Feature line annotations: A novel feature line annotation tailored for 3D clothing is proposed in the Deep Fashion3D dataset. Similar to facial landmarks, feature lines represent the most prominent features of interest, such as open edges, collars, cuffs, waistlines, etc., which can provide strong priors for the faithful reconstruction of 3D clothing.
2. Calibrated multi-view real-world images: For each 3D clothing model, we provide corresponding calibrated multi-view real-world images, which are critical for generalizing trained models to in-the-wild images.
3. 3D poses: For each 3D clothing model, we annotate its 3D pose represented by SMPL coefficients. Due to the highly coupled nature between the human body and clothing, we believe that the annotated 3D poses can help infer the overall shape of the clothing model and pose-related deformations.
Figure 3: Quick overview of the annotations in the Deep Fashion3D dataset

搜集汇总

背景与挑战

背景概述

Deep Fashion3D是目前最大的3D服装模型数据集，包含2078个真实服装重构的模型，涵盖10种服装类别，并提供丰富的3D注释信息，支持3D服装重建和研究。