The paper discusses the development of CADSim, a tool designed to simulate sensors for self-driving vehicles. Instead of manually created assets, it uses real-world sensor data to automatically build a diverse and comprehensive library for sensor simulation. CADSim combines part-aware object-class priors via a small set of CAD models with differentiable rendering to reconstruct vehicle geometry, including articulated wheels, with high-quality appearance. It offers more accurate shapes from sparse data compared to existing methods and efficiently trains and renders. It can be used in various applications, including accurate testing of autonomy perception systems.
Publication date: 2 Nov 2023
Project Page: https://waabi.ai/cadsim/
Paper: https://arxiv.org/pdf/2311.01447
