CADSim: Robust and Scalable in-the-wild 3D Reconstruction for Controllable Sensor Simulation

: In this supplementary material, we provide additional details on our method and experiments, and then show additional application results using CADSim. For our method, we describe how we create a shared low-dimensional representation space for optimizing over a set of CAD models (Sec A.1), provide details on our selected appearance representation (Sec A.2), and the exact inference optimization procedure performed and hyperparameters used CADSim (Sec A.3). For our experiments, we first provide implementation details of the baselines compared against (Sec B), as well as dataset and metric details (Sec C). We then report thorough ablations on our choice of geometry and appearance (Sec D.1), demonstrate the robustness of CADSim to data noise (Sec D.2), show our approach applied to non-vehicle objects (Sec D.3), and show more experiment results of our model improving perception evaluation (Sec D.5). Finally, we show additional applications of CADSim. We show using CADSim for multi-sensor simulation examples for scenario replay (Sec E.1) and mixed reality (Sec E.2), as well as showing CADSim naturally supporting texture transfer for creating diverse assets (Sec E.3). Additionally, we include a supplementary video, supplementary 56.mp4 providing an overview of our methodology, as well as video results on novel-view synthesis, and realistic multi-sensor simulation.