3D building fabrication with geometry and texture coordination via hybrid GAN
- PDF / 7,631,003 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 61 Downloads / 517 Views
ORIGINAL RESEARCH
3D building fabrication with geometry and texture coordination via hybrid GAN Zhenlong Du1 · Haiyang Shen1 · Xiaoli Li1 · Meng Wang1 Received: 31 December 2019 / Accepted: 17 August 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract 3D building plays the essential role in digital city construction, city augmented reality and smart urban planning & design. Conventional building construction is accomplished by modeling software which requires significant human intervention. In this paper, a method of 3D building fabrication via Hybrid generative adversarial network (GAN) is proposed, in which a loss function with the introduction of cycle consistency loss and perceptual loss is given, a multi-properties GAN chain is built to create the building with complex architectures. Additionally, a mixed GAN network to generate the geometry and texture coordination is put forward. The discussed method can refine rough architectural models for outputting realistic buildings. Experiments show that generated 3D buildings utilizing the presented method are realistic, with geometry and textural consistency, which improves performance by 20% over traditional methods. Keywords GAN chain · Hybrid GAN · 3D building generation · multi-properties generation
1 Introduction Currently, 3D building models are one of the most important components in digital city construction. They play an indispensable role in urban construction & planning and comprehensive city management. 3D modelling construction has been a hot topic in the fields of industrial vision, product prototype design, 3D printing, and building information modeling (BIM). Although CAD software, such as, ProE, 3DSMax, Maya, Unity, Unreal Engine 4 or AutoCAD, has been utilized to make models for several decades, building generation is still labor intensive and time consuming. Deep learning (Wen et al. 2019; LeCun et al. 2015; Lu et al. 2018, 2019; Uemura et al. 2019; Zhao et al. 2018; Xu et al. 2019b, a; Singhal et al. 2019; Kim and Chung 2020)
* Zhenlong Du duzhl‑[email protected] Haiyang Shen [email protected] Xiaoli Li [email protected] Meng Wang [email protected] 1
Nanjing Tech University, Nanjing, Jiangsu, China
provides a solution to the issue. Generative adversarial networks (GAN) (Arora et al. 2017; Goodfellow et al. 2014) can learn the distribute from given data and generate new data. It can be applied to generate 3D building models. In this paper, a novel 3D building model generation approach based on hybrid GAN is further investigated. GAN is based on the principle of a two-player zero-sum game. It involves a generating model G (generator) and a discriminant model D (discriminator). Generator G captures the distribution of sample data, and generates the new data similar to the training data with noise z following a certain distribution (uniform distribution, Gaussian distribution, etc.). Discriminator D is a two-classifier, which estimates the probability of data derived from the training data. In this paper, the loss function
Data Loading...
