Depth-dependent noise interference reveals scene-structure constraints on contour completion.

Contour completion allows perception of whole objects despite fragmented input, yet the mechanisms that support this process remain debated. Purely local, bottom-up models posit that oriented edge signals propagate across gaps according to two-dimensional (2-D) principles of good continuation, predicting that interference from luminance noise should depend only on local image geometry. In contrast, scene-based frameworks - including accounts that incorporate 3-D surface layout - propose that completion depends on a global analysis of depth, occlusion, and surface assignment, predicting stronger disruption when noise lies on the same depth surface as an interpolated contour. We tested these competing predictions using stereoscopic Kanizsa figures embedded in dense luminance noise while manipulating whether the noise was coplanar with or displaced in depth from the figure. Thin-fat shape-discrimination thresholds were 20% higher for coplanar noise than for depth-separated noise (1.01∘ vs. 0.84∘), independent of whether the figure appeared in front of or behind the inducer plane. A control experiment ruled out the possibility that this difference arose from between-eye noise sampling. These results demonstrated that noise interfered with - rather than augmented - contour completion, and that such interference depended on 3-D surface assignment. The findings place constraints on purely monocular, feedforward accounts and suggest that illusory contours and luminance noise are not represented in a common format. Instead, the visual system appears to evaluate scene structure - including depth relationships - before determining whether and how contour interpolation proceeds.