Vision-guided three-dimensional range-gated imaging based on epistemic uncertainty estimation

Abstract. In recent years, vision-guided three-dimensional (3D) range-gated imaging has broken through the hardware limitations of traditional methods and brought new ideas to the field of 3D range-gated imaging. However, the existing approaches do not consider the uncertainty caused by incomplete training data, which make accuracy of the existing methods still possible for further improvement. In our work, we extend the well-known Gated2Depth framework using epistemic uncertainty by introducing Bayesian neural networks to provide uncertainty that does not exist in the input data due to incomplete training data. Finally, in the proof experiments, mean absolute error achieved 8.7% improvement on the night data and 9% improvement on the daytime data. The improvement of 3D range-gated imaging accuracy reduced the holes and blurred problems in the depth map and obtained sharper target edges.