A Soap-Film Test for Depth Edges: Scale-Free Plateau Problems for Monocular Depth

When a network estimates depth from a single photo, the hardest pixels lie at occlusion edges, where a near object sits in front of a far background. A model trained to output one depth value per pixel hedges at these edges and paints a smooth ramp between foreground and background, producing "flying pixels" that belong to no real surface. PointDiT names this over-smoothing as the basic failure of regressors faced with an ambiguous input. Whether two adjacent image regions are joined by a real surface or separated by empty space is a discrete question about topology. Current systems answer it implicitly, pixel by pixel, and the single-image setting makes it worse because absolute scale is unknown: a small near scene and a large far scene give the same image.

This note proposes answering that question with a soap-film comparison in a geometry built so that the unknown scale does not matter. Give the space in front of the camera the metric $(dX^2+dY^2+dZ^2)/Z^2$, where $(X,Y,Z)$ are camera coordinates and $Z$ is depth. This space is exactly hyperbolic 3-space in its upper half-space form, with the camera sitting on the boundary at infinity, and uniform rescaling of the scene about the camera becomes a rigid motion. The monocular scale ambiguity is therefore built into the geometry. At a candidate depth edge, the "split" answer and the "connected" answer correspond to the two classical competitors of the Plateau problem: two separate minimal disks versus one minimal annulus, the catenoid. Comparing their hyperbolic areas gives a training-free, exactly scale-invariant decision about whether to cut. The size of the area gap measures how ambiguous the edge is.

Core Claim

In this metric, the area of a surface patch seen from the camera is

$$dA_{\mathrm{hyp}} = \frac{d\Omega}{\cos\theta\,\cos^2\varphi},$$

where $d\Omega$ is the solid angle the patch subtends at the camera, $\theta$ is the angle between the patch normal and the viewing ray, and $\varphi$ is the angle between the ray and the optical axis. The identity follows from the Euclidean relation $d\Omega = dA\cos\theta / r^2$ together with $Z = r\cos\varphi$.

The formula says two things. Hyperbolic area charges each surface by its image footprint plus a penalty for slanting away from the viewing ray, and it is unchanged by global rescaling. The minimal disk spanning a circular rim centred on the optical axis is a cap of a sphere centred on the camera. Such a cap has constant range, so it is exactly the "sharp discontinuity, fronto-facing layer" that a cut depth map produces. The competing minimal annulus is the sloped bridge that flying pixels approximate.

The claim is that choosing, at every candidate edge, whichever topology has smaller hyperbolic area yields sharper and better-calibrated depth discontinuities than ratio thresholds, total-variation smoothing, or diffusion sampling. For coaxial circular rims the choice reduces to comparing the log range ratio $\log(r_2/r_1)$ against a critical value that depends on the rims' angular size. For general contours the choice is computed numerically, and it depends on contour shape as well as depth ratio. That shape dependence is the part no simple threshold reproduces.

What the Plateau Literature Supplies

The asymptotic Plateau literature studies area-minimizing surfaces in hyperbolic space, and three of its results shape the design.

The first concerns which solutions to trust. Coşkunuzer's existence result for Gromov hyperbolic spaces (Properly embedded least area planes in Gromov hyperbolic $3$-spaces) and the Meeks–Yau cut-and-paste argument show that global least-area surfaces cannot fold over or pile up on themselves. In contrast, merely stationary minimal surfaces can fail proper embeddedness (NON-PROPERLY EMBEDDED MINIMAL PLANES IN HYPERBOLIC 3-SPACE). This constrains the method. Running a gradient flow from one starting surface would find a stationary surface whose topology is fixed by the starting guess. We instead minimize separately within each topological class and compare the minima.

The second concerns ties. Number of least area planes in Gromov hyperbolic 3-spaces proves that boundary curves bounding more than one least-area surface form a meager set, so ties are non-generic. Minimal Planes in Hyperbolic Space links non-uniqueness in hyperbolic 3-space to changes in topology. This motivates reading the area gap as an ambiguity margin: a near-tie should be rare and should mark genuinely ambiguous edges.

The third is a limit on what to borrow. These theorems concern curves at the sphere at infinity, whereas our rims lie in the interior. We use the geometry and the reasoning about which topology wins, not the theorems themselves.

On the vision side, the closest relatives impose geometric structure without treating topology. NDDepth enforces piecewise planarity and admits it fails on curved geometry. The Plane Parameter MRF encodes connectedness between superpixels as a soft cost. DiffusionDepth and PointDiT sample or regress depth and recover edges through learning. None makes connect-versus-cut an explicit, scale-invariant comparison between topologies.

The section of the research file meant to list prior combinations of these two areas was truncated. Nothing in what I could see joins hyperbolic Plateau problems to depth estimation.

Research Program

The work has three parts.

Theory. We derive the connect-versus-cut criterion for coaxial circular rims in the camera-centred hyperbolic metric. This is the analogue of the classical catenoid-versus-two-disks transition, and it gives a critical log-range-ratio curve as a function of angular radius. We then characterize how non-circular and non-coaxial contours shift the criterion. That characterization is the only place where the method can beat a lookup table.

Algorithm. The algorithm runs as a post-processing layer on any base depth model. Candidate discontinuity contours come from depth-gradient and image edges. For each contour, the two boundary curves are its foreground and background sides. We compute a minimal-annulus and a two-disk solution on triangle meshes, each by energy minimization under the hyperbolic area weight. The lower-area topology is kept. The cut regions are then filled with the disk solutions, which are constant-range caps in the circular case and their curved generalizations otherwise.

Uncertainty. The normalized area gap is exported as a per-edge uncertainty score. Its calibration is tested against true edge errors.

Baselines

The naive reference point is the raw output of each base model with no post-processing. This shows how many flying pixels and how much edge blur exist to remove.

The strongest existing method on edge sharpness, by its own account, is PointDiT, which predicts pixel-space point maps in one step. Our layer will be applied to PointDiT's own outputs and also to DiffusionDepth and NDDepth outputs, on identical images and resolutions, using released checkpoints and no retraining of any base model. No reported result found on Lacuna for PointDiT, DiffusionDepth or NDDepth on NYU or KITTI. The relevant numbers were not in the file, so we will reproduce each baseline's standard metrics ourselves before comparing.

For the uncertainty claim, the competitor is DiffusionDepth's sample variance at edges, using the same number of samples it would need to match our runtime.

The ablation that isolates the key ingredient keeps every component of the pipeline but replaces hyperbolic area with Euclidean area. Euclidean area loses scale invariance about the camera. If it performs as well, the geometry is not doing the work.

Beating the baselines means two things. First, a statistically significant reduction in flying-pixel fraction and in boundary-band depth error, with no loss in global accuracy (absolute relative error, $\delta<1.25$). Second, area-gap uncertainty that ranks erroneous edges better than diffusion variance.

Why not just…?

Why not just threshold the depth ratio across each edge? For coaxial circles, our criterion is exactly a ratio threshold that depends on angular size. A tuned ratio threshold, scale-invariant and fit on validation data, is therefore a real threat. The method earns its keep only if contour shape matters. We test this directly: after fitting the best ratio-plus-angular-size threshold, we measure the residual gain on non-circular, elongated and partly occluded contours. If the gain is negligible, we report that the soap-film view reduces to a two-parameter rule and stop.

Why not just use total-variation smoothing on log-depth? Total variation does cost jumps in a scale-invariant way. However, it is a convex single-solution method on depth graphs. It decides where a jump sits but never compares two topologies, and it gives no ambiguity margin. It runs in the same post-processing slot with a tuned weight. If it matches us on boundary metrics, our remaining claim reduces to the uncertainty score alone.

Why not just trust diffusion models to produce sharp edges and sample their ambiguity? PointDiT's pixel-space design targets exactly this weakness. If its edges are already crisp, our layer may change little, and we would say so. Diffusion variance is uncalibrated and may collapse onto one mode. The calibration comparison above settles whether the area gap adds information beyond sampling.

Proposed Experiments

The data are NYU-style indoor scenes and KITTI. On KITTI, LiDAR covers only about 4% of pixels, as DiffusionDepth notes, so edge metrics are computed only where ground-truth depth exists on both sides of a contour. We add synthetic rendered scenes with known geometry and controlled near-tie configurations.

The decisive checks are three. First, scale invariance: rescale synthetic scenes and confirm that decisions are unchanged under the hyperbolic weight but drift under the Euclidean ablation. Second, the shape-dependence test against the fitted ratio threshold. Third, whether near-tie edges, as flagged by a small area gap, concentrate the remaining errors.

A negative result on the second check is still informative. It would establish that depth-edge topology from a single image is governed by a scale-free two-parameter law.