How to see depth in the dark without training a neural network

Depth sensing in darkness is hard because pixel noise drowns out fine details. This work uses Field of Junctions—a classical technique that extracts only coarse, noise-stable edges—then feeds those into boundary-aware Semi-Global Matching that protects true depth discontinuities. The result: sparser but more accurate disparity maps than recent learning-based methods on benchmark datasets, without any training.