# Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling

**Authors:** Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, Zihan Zhou  
**arXiv:** [1908.00222](https://arxiv.org/abs/1908.00222)  
**Date:** 2019-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.00222v3/

## Key points

- Growing interest in learning-based methods for 3D scene modeling and understanding
- Importance of detecting and utilizing salient semi-global or global structures such as junctions, lines, planes, cuboids, smooth surfaces, and symmetries
- Challenges of obtaining ground truth annotations manually due to the large number of 3D structure instances and factors like viewpoints and occlusions
- Introduction of a new synthetic dataset called Structured3D to address this issue
- Structured3D provides large-scale photo-realistic images with rich 3D structure annotations for structured 3D modeling tasks
- Professional interior designs are leveraged to automatically extract 3D structures and high-quality images are generated using an industry-leading rendering engine
- The dataset is used in combination with real images to train deep networks for room layout estimation
- Improved performance on benchmark datasets compared to previous methods by utilizing the synthetic dataset
- Structured3D provides a valuable resource for training and evaluating learning-based methods in structured 3D modeling
- Suitable for various applications in computer vision and graphics research due to its photo-realistic nature and comprehensive annotations.

## Summary

Recently, there has been a growing interest in developing learning-based methods for 3D scene modeling and understanding that can detect and utilize salient semi-global or global structures such as junctions, lines, planes, cuboids, smooth surfaces and various types of symmetries. Obtaining ground truth annotations for these structures through manual labor is challenging and inefficient due to the large number of 3D structure instances (e.g., line segments) and factors like viewpoints and occlusions. To address this issue, the authors of this paper introduce a new synthetic dataset called Structured3D. The primary objective of this dataset is to provide large-scale photo-realistic images with rich 3D structure annotations for a wide range of structured 3D modeling tasks. The authors leverage professional interior designs to automatically extract 3D structures from them. They generate high-quality images using an industry-leading rendering engine. The Structured3D dataset is used in combination with real images to train deep networks for room layout estimation. By utilizing this synthetic dataset, the authors demonstrate improved performance on benchmark datasets compared to previous methods. Overall, the introduction of Structured3D provides researchers with a valuable resource for training and evaluating learning-based methods in the field of structured 3D modeling. The dataset's photo-realistic nature and comprehensive annotations make it suitable for various applications in computer vision and graphics research.