Lingjie Mei will present his General Exam "Procedurally Generating 3D Indoor Scenes for Computer Vision and Robotics" on Monday, May 6, 2024 at 3:00 PM in Friend 109 and via zoom. Zoom link: https://princeton.zoom.us/j/8415068766 Committee Members: Jia Deng (advisor), Felix Heide, Ellen Zhong Abstract: Synthetic data rendered by conventional computer graphics has seen increasing adoption in computer vision and AI research, especially for 3D vision and embodied AI. Synthetic data can be rendered in unlimited quantities and can automatically provide high-quality 3D ground truth, enabling large-scale training of computer vision models and embodied agents. We propose a synthetic data generator that can produce infinite photorealistic 3D indoor scenes. Our system is entirely procedural: every asset, including furniture, architecture elements, appliances, and other day-to-day objects, is generated from scratch via mathematical rules and randomized parameters. We also introduce a constraint-based arrangement system, which consists of a domain-specific language for expressing diverse constraints and preferences on indoor scene composition, and a solver that generates floorplans and scene layouts that maximally satisfy the constraints. The scenes are photorealistic, generated with pixel-perfect annotations, and can be exported to real-time simulators and run at interactive frame rates. Overall, anyone can directly use and customize our system to generate infinite high-quality indoor scene data for a variety of computer vision tasks, including object detection, semantic segmentation, optical flow, and 3D reconstruction. Robot scientists can also use the data in simulation for manipulation and grasping. We expect this to be a useful tool for researchers beyond. Reading List: https://docs.google.com/document/d/1mmf5W0sKZkLgWVKXpCA7TkSSWUfBoeuKFXWHw8k4... Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.