Abstract:
3D imaging is critical in applications requiring precise spatial detail. Among available technologies, LiDAR sensors are particularly prized for their accuracy and reliability. However, in realistic conditions, their performance is usually compromised by photon noise and sparse, low-resolution measurements.
To overcome these limitations, we introduce a deep learning approach with multiscale processing to produce high-quality depth reconstructions despite low-quality input data. Tested on simulated LiDAR datasets, the method has notable improvements in accuracy and robustness.
