Robust Depth Completion with Uncertainty-Driven Loss Functions

Recovering a dense depth image from sparse LiDAR scans is challenging task. Despite the popularity of color-guided methods for sparse-to-dense completion, they treated pixels equally during optimization, ignoring uneven distribution characteristics in map and accumulated outliers synthesized ground truth. In this work, we introduce uncertainty-driven loss functions to improve robustness completion handle uncertainty completion. Specifically, propose an explicit formulation robust with Jeffrey's prior. A parametric uncertain-driven introduced translated new that are noisy or missing data. Meanwhile, multiscale joint prediction model can simultaneously predict maps. The estimated also used perform adaptive on high uncertainty, leading residual refining results. Our method has been tested KITTI Depth Completion Benchmark achieved state-of-the-art performance terms MAE, IMAE, IRMSE metrics.