MDA (Mixture-Density Ambiguity)

MDA: Mixture-Density Representation for Flying-Point-Free Depth Estimation

This paper introduces MDA (Mixture-Density Ambiguity), a depth estimation framework that replaces single-depth-per-pixel prediction with a mixture-density representation allowing multiple depth hypotheses and associated probabilities per pixel. The approach directly addresses 'flying point' artifacts at object boundaries, where conventional models predict spurious intermediate depths between foreground and background surfaces. MDA extends naturally to transparent objects and sky regions, substantially improving boundary reconstruction with negligible runtime overhead across multiple backbone architectures.