Neural representation for surface reconstruction from ultra-sparse depth measurements

Low-cost, low-power sensors are essential for embedded and edge applications such as miniature Unmanned Aerial Vehicles (UAV) and robotics, but their sparse, noisy, and low-fidelity outputs limit their effectiveness in perception tasks. We propose an AI-driven framework that enhances these sensors by transforming low-resolution measurements into dense surface representations. Our approach learns the unsigned distance field anchored by noisy depth measurements to extract predicted truth surface. To enable robust testing, we use simulation and digital twins to generate synthetic data across diverse environments. Experiments show that dense reconstructions from our approach can robustly improve downstream tasks such as navigation, and extends the utility and resilience of low-power sensor systems in resource-constrained environments.