Why depth cameras see ghosts at object edges—and how to stop them

Depth cameras struggle at object boundaries, producing spurious 3D points (flying points) in the empty space between foreground and background. The problem: each pixel gets assigned one depth value, which at boundaries forces the model to predict an intermediate depth that exists nowhere real. MDA swaps this for a mixture-density representation, letting pixels predict multiple depth hypotheses with probabilities attached. At boundaries, different hypotheses can align with different surfaces; the final depth picks one hypothesis rather than averaging into empty space. Tested across backbones, MDA removes flying-point artifacts, handles transparent objects as depth layers, and generates clean skylines—all with negligible overhead.