Screened-Poisson Geometry Regularization Against Baked-In Shadows in Inverse Rendering

When a 3D reconstruction system tries to explain photographs of an object, it must decide what part of each pixel's color comes from the object's shape, what part comes from its paint, and what part comes from the light falling on it. Optimizers tend to take the cheapest explanation. Cast shadows get carved into the shape as dents, specular glints get pushed out as bumps, and whatever the shape cannot absorb gets painted into the color map. These are "baked-in lighting" artifacts. This note targets the geometric half of the problem. We propose that much of the shape damage comes from how the shape is regularized. The usual constraint, the eikonal loss, asks a signed distance function (a field whose value at each point is the distance to the surface, with sign marking inside and outside) to have unit-length gradient everywhere. That constraint lets a local mistake travel along straight lines through the whole field, so one misread shadow can distort geometry far from where it occurred.

Electrostatics offers a different template. In a salt solution, mobile ions crowd around any charge and cancel its field beyond a short distance, the Debye length. The governing equation, the screened Poisson equation, has solutions whose disturbances die off exponentially with distance. HotSpot showed that this same equation, used as a loss on distance fields fit to point clouds, is an asymptotically sufficient condition for a true distance function, and that its errors decay exponentially away from their source. HotSpot names inverse rendering as future work and stops there. Our claim is that putting this screened operator inside a physics-based inverse renderer confines shadow- and highlight-induced shape errors to within roughly one screening length of where they arise. The screening length should then be set by the electrostatic rule of thumb that screening layers must not overlap across thin gaps. Together, these should reduce baked-in geometry without the over-smoothing that generic regularizers cause.

Core Claim

Let $u(\mathbf{x})$ be the learned signed distance and $h(\mathbf{x}) = \exp(-\lambda |u(\mathbf{x})|)$ its "screened" transform. Here $\lambda$ is the inverse screening length, the analogue of the inverse Debye length. HotSpot's condition requires

$$\nabla^2 h(\mathbf{x}) - \lambda^2 h(\mathbf{x}) = 0 \quad \text{away from the surface}, \qquad h = 1 \text{ on the surface},$$

and the distance is recovered as $u \approx -\lambda^{-1}\ln h$ as $\lambda$ grows. We add the squared residual of this equation, evaluated at sampled off-surface points, to the photometric loss of a physically based inverse renderer. The renderer uses a microfacet reflectance model, environment lighting, and secondary visibility traced on the same signed distance field. The surface boundary condition is the current zero level set, which the photometric loss moves.

We make two testable predictions. First, a photometric error injected at one surface location should produce shape error that decays roughly as $e^{-\lambda r}$ with distance $r$ under the screened loss, and much more slowly under the eikonal loss. The exponential form comes from the Green's function of the screened operator. Second, the colloid study Renormalized charge and dielectric effects in colloidal interactions solved the nonlinear screened electrostatics of nearby particles and found that cross-boundary effects matter only when screening layers overlap. By analogy, the regularizer should couple opposite faces of thin structures, such as leaves, studs, and thin walls, only when $1/\lambda$ exceeds about half the gap. That gives a principled cap on the screening length rather than a free hyperparameter.

The charge-screening literature plays two roles here. It supplies the operator and its locality property, and it supplies the overlap criterion. We do not model ions or electrolytes in the scene.

Relation to Prior Work

The inverse-rendering work most relevant here attacks baking by improving the light-transport model rather than the shape regularizer. TensoIR uses a compact tensor radiance field with online ray marching for secondary shadows and inter-reflections. Radiometrically Consistent Gaussian Surfels for Inverse Rendering (RadioGS) ties learned radiance to a physically rendered radiance on flat Gaussian disks and traces rays between them for indirect light. Both motivate this proposal, because they show that better transport reduces albedo baking. Neither uses a distance-field regularizer with local error support, so their geometry relies on density or primitive placement. We keep transport comparable and change only the geometry prior.

ExMesh++ extracts relightable meshes and likewise relies on standard geometric regularization. HotSpot supplies the method but was tested only on point clouds, where no lighting ambiguity exists. Our contribution is the inverse-rendering setting, the locality analysis under photometric error, and the overlap-based choice of $\lambda$.

The research file raises a tension that could contradict us. Global illumination needs long-range information, so a strongly localized operator might stop the system from explaining a distant shadow caster. Our answer is that the screening acts on the shape field, while light transport still traces rays across the whole scene. This proposal tests that answer directly instead of assuming it.

Two works constrain the evaluation settings. EndoPBR models endoscopy with a spotlight attached to the camera. In that regime, glare on wet tissue is the dominant artifact and the light cannot be moved independently, which makes it a natural stress test for highlight-induced bumps. CryoGEM shows in a different imaging domain that unmodeled physical effects get baked into reconstructed 3D structure. It is the motivating evidence that the failure mode is general.

Research Program

We will build one signed-distance inverse renderer and hold everything fixed except the geometry regularizer. The fixed parts are reflectance (albedo and roughness with the dielectric Fresnel assumption TensoIR uses), environment lighting, and secondary visibility by sphere tracing. The regularizer comes in four variants: eikonal only, screened only, both, and screened with $\lambda$ annealed from small to large. Annealing follows HotSpot's asymptotic argument, since large $\lambda$ gives accurate distance but weak far-field gradients.

The $\lambda$ cap from the overlap criterion is computed from the current estimate of the thinnest feature gap. We measure that gap by casting short probe rays inward from surface samples.

Two analyses accompany the benchmarks. The first is a controlled locality experiment. On scenes with ground-truth geometry, we perturb the target images by painting a false shadow patch or a false highlight in one region. We then measure normal error as a function of geodesic distance from the patch and fit a decay rate. The second is a thin-structure sweep: on scenes with known gap widths, we vary $1/\lambda$ across the predicted overlap threshold to check whether shape quality breaks where the electrostatic analogy says it should.

Baselines

The naive reference point is the same renderer with the eikonal loss alone. This is the standard choice for distance-field reconstruction and the condition HotSpot calls necessary but insufficient. It shares our code, data, iteration count, and a hyperparameter search of the same size (twelve configurations, chosen by validation-view error). Beating it means lower normal mean angular error on the TensoIR synthetic scenes, and a measurably faster spatial decay of error in the locality experiment.

The strongest existing method on relighting is RadioGS. On the TensoIR synthetic benchmark it reports a Relight PSNR of 31.41 against 29.91 for IRGS, and on its own Blender Cycles indirect benchmark an indirect-illumination PSNR of 32.88 against 24.22 for IRGS. We will run its released configuration on identical scenes and a matched GPU-hour budget. Because RadioGS represents surfaces with disks rather than a distance field, we do not expect to beat it on relighting by changing only the shape prior. We count success as matching its relighting within about half a decibel while improving normal error and albedo error near cast-shadow boundaries. The research file gives no normal-error or albedo-PSNR numbers for RadioGS or TensoIR on this benchmark, so we will measure both ourselves.

TensoIR is a second strong competitor. It is run the same way, since its fast training makes matched budgets easy.

The ablation that isolates the key ingredient keeps the screened loss but fixes $\lambda$ at a value that violates the overlap cap. If the cap does not matter, the electrostatic reasoning adds nothing beyond HotSpot. A second ablation swaps the screened loss for the eikonal loss plus a normal-smoothness term tuned on the same budget. It separates "local error support" from "more regularization."

Why not just…?

Why not just add a stronger smoothness or curvature penalty to the eikonal loss? A smoothness term suppresses the dents and bumps that baked lighting creates, but it suppresses real fine detail just as much. It also does nothing to stop errors from traveling along characteristics. The screened loss constrains what a distance field is, not how smooth it is. The matched-budget smoothness ablation will report normal error near shadows together with Chamfer distance on fine-detail regions. Our proposal wins only if it moves that tradeoff curve outward, rather than landing at a different point on it.

Why not just use a pretrained single-image normal predictor as a prior? This is the most serious threat. A good learned prior could remove shadow dents with no change to the regularizer at all. We will run it as a baseline and in combination with our loss. If the prior alone matches us on the synthetic scenes, the claim narrows to settings where such priors are out of domain. The endoscopy data built on C3VD is the natural test there, with glare on wet tissue and a co-located light. We will report that narrowing honestly rather than argue around it.

Why not just capture images under several lightings? Varying the light breaks the albedo-shading ambiguity directly. But in endoscopy the light is mounted on the camera (EndoPBR), and in outdoor captures the sun cannot be overridden (R3GW). The method is aimed at those single-lighting regimes. On the synthetic benchmarks we report results only for the single-environment training split, so that multi-light information cannot leak in.

Why not just rely on better light transport, as RadioGS does? Transport and shape priors address different leaks. RadioGS's large gain on indirect light, from 24.22 to 32.88 PSNR, says nothing about whether residual unexplained shading is absorbed by the shape. We will graft the screened loss onto a distance-field variant with RadioGS-style radiometric consistency. If the two effects are additive, the gains should stack; if better transport removes the problem, the gain from the screened loss should vanish.

Proposed Experiments

The primary benchmark is the TensoIR synthetic set. We report novel-view PSNR, normal mean angular error, albedo PSNR, and Relight PSNR, with normal error also broken out inside masks of ground-truth cast shadows. The RadioGS Blender Cycles set is used for indirect-light PSNR and for its explicit split between direct and indirect light, which lets us check whether reduced shape baking shifts unexplained light into albedo instead. Stanford-ORB tests real objects; no reported result was found on Lacuna for any method on it, so all numbers there will be our own runs. We also report geometry and synthetic depth on the C3VD-based endoscopy setting for the highlight-bump regime.

Three checks decide whether the idea holds. The locality experiment must show an error decay rate under the screened loss that grows with $\lambda$, and is clearly faster than under the eikonal loss. Without this, the mechanism is wrong even if the averages improve. The thin-structure sweep must show quality degrading once $1/\lambda$ crosses roughly half the gap; otherwise the overlap criterion should be dropped. Finally, relighting error on scenes with long cast shadows (Lego, Bulldozer) must not get worse than with the eikonal loss. A drop there would mean the screened loss blocks the global explanations that shadows need, which is the tension raised above.