Physically-Based Rendering for Indoor Scene Understanding
Using Convolutional Neural Networks


Abstract

Indoor scene understanding is central to applications such as robot navigation and human companion assistance. Over the last years, data-driven deep neural networks have outperformed many traditional approaches thanks to their representation learning capabilities. One of the bottlenecks in training for better representations is the amount of available per-pixel ground truth data that is required for core scene understanding tasks such as semantic segmentation, normal prediction, and object edge detection. To address this problem, a number of works proposed using synthetic data. However, a systematic study of how such synthetic data is generated is missing. In this work, we introduce a large-scale synthetic dataset with 400K physically-based rendered images from 45K realistic 3D indoor scenes. We study the effects of rendering methods and scene lighting on training for three computer vision tasks: surface normal prediction, semantic segmentation, and object boundary detection. This study provides insights into the best practices for training with synthetic data (more realistic rendering is worth it) and shows that pretraining with our new synthetic dataset can improve results beyond the current state of the art on all three tasks.

Paper

Dataset

  • Our dataset consists synthetic image rendered from virtual scene. For each camera view, we provide physically based color rendering. Compared to the rendering from OpenGL pipeline, our physically based rendering is more natural looking containing proper indoor/outdoor illumination and soft shadow. Surface normal, semantic segmentation, and instance boundary that are perfectly aligned with the color rendering are also provided. We also plan to release depth image and kinect simulation. The whole dataset will be available soon.

  • Snapshot:
  • Physically Based Rendering Surface Normal Semantic Segmentation Instance Boundary

Source Code and Pre-trained Models

  • Source code for the whole system will be available.