基于图片的自由物体3D打印模型生成数据

通过构建一个包含大量自由类别物体、且均为水密性的3D打印模型及其对应的单张或多张渲染图片的大规模配对数据集，可以为深度学习模型提供训练基础，使其学习从二维物体图像生成完整的三维几何。这一数据集主要适用于桌面游戏棋子和微缩模型的定制、个性化的物体模型快速制作、虚拟化身的实体化打印以及内容创作等领域。利用该数据训练出的模型，能够让用户通过上传一张照片来生成一个可直接用于3D打印的对应模型，解决了从零开始制作三维模型技术难度极高、周期漫长且成本昂贵的问题。基于单张图片生成自由类别的可3D打印模型，旨在让三维人形创作大众化。具体过程包括：（1）数据收集：用户提供一张包含清晰物体主体的RGB图片（I_rgb）（2）数据处理：将输入的图片送入一个在物体图像上经过优化的预训练图像编码器。特征向量通过公式 F_image = Encoder_image(I_rgb) 提取，其中 F_image 为图像特征向量，Encoder_image 为图像编码器。（3）模型构建：使用提取的图像特征向量作为输入，设计并搭建一个专注于物体三维模型生成的深度解码模型。该模型从特征中推断并生成物体模型的隐式三维表示。根据公式 SDF = Decoder_3D(F_image) 从图像特征中解码出三维物体模型，其中 SDF 为三维模型，Decoder_3D 为三维形状解码器。关键的评估指标包括交并比（Intersection over Union, IoU）和倒角距离（Chamfer Distance, CD）。此方法适用于从单张照片快速生成个性化的物体模型，极大地降低了三维物体模型的3D打印门槛。

By constructing a large-scale paired dataset comprising numerous freely categorized, watertight 3D printing models and their corresponding single or multiple rendered images, a solid training foundation is provided for deep learning models to learn to generate complete 3D geometry from 2D object images. This dataset is primarily applicable to scenarios including customization of tabletop game pieces and miniatures, rapid production of personalized object models, physical printing of virtual avatars, and content creation. Models trained on this dataset allow users to generate a corresponding model directly usable for 3D printing by uploading a single photo, solving the problems of extremely high technical threshold, long production cycles, and high costs associated with creating 3D models from scratch. The goal of generating freely categorized, 3D-printable models from a single image is to make 3D humanoid creation accessible to the general public. The specific process includes: (1) Data Collection: Users provide an RGB image (I_rgb) containing a clear main subject of the object. (2) Data Processing: The input image is fed into a pre-trained image encoder optimized on object images. The image feature vector is extracted via the formula $F_{ ext{image}} = ext{Encoder}_{ ext{image}}(I_{ ext{rgb}})$, where $F_{ ext{image}}$ is the image feature vector and $ ext{Encoder}_{ ext{image}}$ is the image encoder. (3) Model Construction: Using the extracted image feature vector as input, a deep decoding model dedicated to generating 3D object models is designed and built. This model infers and generates the implicit 3D representation of the object model from the features. The 3D object model is decoded from the image features via the formula $SDF = ext{Decoder}_{3D}(F_{ ext{image}})$, where $SDF$ represents the 3D model and $ ext{Decoder}_{3D}$ is the 3D shape decoder. Key evaluation metrics include Intersection over Union (IoU) and Chamfer Distance (CD). This method is applicable to rapidly generating personalized object models from a single photo, greatly lowering the barrier to 3D printing of 3D object models.

搜集汇总

背景与挑战

背景概述

该数据集包含5815条CSV格式数据，每条数据由RGB图片、图像特征向量和三维模型等字段组成，用于训练深度学习模型从二维图像生成可直接3D打印的三维物体模型，主要应用于个性化物体定制和快速模型制作，降低了三维建模的技术门槛。