Releasing the Smoothing: Region-by-Region Sharpening of Unsigned Distance Fields

Many real 3D objects are open sheets rather than closed solids: a jacket with its zipper undone, a car seat, a thin door panel. A common way to represent them is an unsigned distance field (UDF), a neural network that takes a 3D point and returns how far it is from the nearest surface, with the surface being wherever that distance hits zero. UDFs handle open shapes that "inside/outside" representations cannot. They are hard to train, though, because the distance profile across a sheet is a sharp V. The slope flips direction exactly at the surface, and networks fit that kink badly, which leaves holes and shattered meshes. One fix rounds the V into a smooth U so gradients behave. That makes training stable but blurs fine detail. The opposite fix keeps the sharp V to preserve detail and accepts a harder optimization.

We propose to stop treating this as a fixed choice. Training starts with every part of the shape in the heavily smoothed, easy regime. The smoothing is then removed region by region, only where a local test shows the field has already found the surface reliably, and it comes back automatically if that region later goes wrong. The idea is borrowed from the "gradual release of responsibility" in teaching: support is withdrawn as each sub-skill is shown to be mastered, not on a calendar. Here it becomes a concrete control rule on a per-region smoothing parameter. The claim is that this region-wise, competence-gated release beats both fixed choices and a simple global annealing schedule on thin open geometry.

Core Claim

The research file documents a direct conflict. DUDF replaces the raw distance $d(x)$ with a smoothed target $d(x)\tanh(\alpha d(x))$, which is quadratic near the surface and linear far away. This removes the gradient kink, but the surface normal then has to be computed from second derivatives, and the method needs ground-truth distances. DEUDF argues that this rounding over-smooths and erases sharp folds and slats. It keeps the V, relaxes the constraint that the gradient have unit length near the surface, and lets outputs go negative. Both treat the sharpness of the field as one global design decision. Our claim is that the right amount of smoothing varies across space and over time. Flat, well-sampled panels can afford to sharpen early. Creases, close parallel layers, and sparse regions need the smooth regime for longer. The release should therefore be decided locally, by measuring whether the field already places the surface correctly there.

Concretely, the network regresses the target

$$t(x) = d(x)\,\tanh\big(\alpha(x)\,d(x)\big),$$

where $d(x)$ is the target unsigned distance and $\alpha(x)$ is a sharpness value stored per cell of a coarse spatial grid or octree. The smoothing band has width roughly $1/\alpha$. As $\alpha$ grows, $t$ approaches the sharp distance $d$. Every cell starts at a small $\alpha_0$. For each cell $c$ we keep a running average $e_c$ of a projection error: near-surface queries in $c$ are moved onto the predicted surface using the field's value and gradient, and we measure how far each projected point lands from the nearest input point. The update rule is

$$\alpha_c \leftarrow \begin{cases} \min(\rho\,\alpha_c,\ \alpha_{\max}) & \text{if } e_c < \tau_{\text{rel}} \text{ for } K \text{ consecutive checks},\\ \max(\alpha_c/\rho,\ \alpha_0) & \text{if } e_c > \tau_{\text{re}},\\ \alpha_c & \text{otherwise,}\end{cases}$$

with a growth factor $\rho>1$ and two thresholds $\tau_{\text{rel}} < \tau_{\text{re}}$, so that a cell does not flicker between states. The first branch releases support where the field has shown it can find the surface. The second restores support where the field has regressed. Cells that reach $\alpha_{\max}$ behave like a sharp UDF. There, normals come from first-order gradients evaluated at a small offset from the surface, as in NDF, so the Hessian cost of DUDF is paid only in the cells that are still smoothed.

Relation to Prior Work

Several UDF methods already contain some form of progression or guidance. Each differs from this proposal in a specific way.

CAP-UDF trains in discrete, global stages: a coarse consistency-aware field first, then densification using its own well-moved queries. It shows that staging helps (Chamfer-$L_2$ 0.110 vs 0.202 on ShapeNet-Cars). Its stages change what data is used, apply to the whole shape at once, and never change the shape of the distance profile. We stage the target sharpness, decide it per region, and allow it to reverse.

LevelSetUDF uses the smooth gradients of non-zero level sets to regularize the zero set. Its weighting by distance from the surface is static. We reuse its insight that the smooth regime is reliable, but hand responsibility over adaptively.

Metric–Phase Fields and GaussianUDF keep an auxiliary structure throughout training: a phase field that produces a thin double-cover, or 2D Gaussian disks. We considered fading those scaffolds instead. Fading the phase field still leaves the question of how the remaining unsigned part learns the kink. Fading Gaussians belongs to the multi-view setting, which mixes in photometric difficulties. The smoothing radius is the narrowest scaffold that directly targets the V-shaped singularity, so we choose it.

The teaching literature supplies the control logic, not evidence about networks. The adaptive haptic flute moves learners from enforced trajectories, to brief hints, to intervention only on detected errors, and reports faster learning and less forgetting than static guidance. Our release-and-reengage rule is modeled on that design. That study concerns human motor learning, so whether the analogy holds for gradient descent is exactly what our experiments test.

Research Program

We will build the method on a SIREN-style network, a multilayer network with sine activations as used by DUDF, and on an MLP that allows negative outputs as in DEUDF. Testing both checks that the gain does not depend on the architecture. Two supervision settings are used.

In the dense setting, ground-truth distances are computed from meshes. This matches DUDF's protocol and isolates the sharpness question.

In the raw point cloud setting, $d(x)$ is replaced by the distance to the nearest input point, and the projection error becomes the natural competence signal. The test here is whether the method survives without ground-truth distances, which is one of DUDF's stated limitations.

The analysis has three parts.

First, we will record maps of $\alpha$ over training to check whether release order follows geometric difficulty, measured by local curvature, sampling density, and the gap between parallel layers. If release order is random with respect to these, the "competence" story is wrong even if accuracy improves.

Second, we will measure dependency, the analog of the flute study's forgetting rate. We freeze $\alpha$ at $\alpha_{\max}$ everywhere for a final block of steps and measure how much accuracy drops. A field that genuinely learned the sharp surface should not degrade.

Third, we will count re-engagement events to see whether the hysteresis branch is ever needed. If it never triggers, the method reduces to monotone local annealing, and we will report that simplification honestly.

Baselines

The naive reference is a plain sharp UDF trained on raw distances with a non-negative output layer, as in NDF. It shows the size of the problem we are addressing.

The two strongest fixed-regime competitors are DUDF with its global $\alpha$ and DEUDF. Together they define the trade-off we claim to dissolve. DUDF gets the smoothing grid search over $\alpha$ that we also run on held-out shapes. DEUDF uses its published relaxation settings plus an equal tuning budget.

The strongest numbers in the file come from the progressive and projection-based methods. On ShapeNet-Cars, LevelSetUDF reports Chamfer-$L_2$ of 0.098, and reports CAP-UDF at 0.119 under its protocol. On thin synthetic shells, Metric–Phase Fields reports Chamfer distance 0.672 against 4.977 for UDF baselines. For DUDF and DEUDF on ShapeNet-Cars there is no reported numeric result on Lacuna. DEUDF reports Chamfer-L1 of 3.09 on the Stanford 3D Scene data.

Published numbers come from differing sampling and scaling protocols, so we will rerun LevelSetUDF, CAP-UDF, DUDF, DEUDF, and Metric–Phase Fields under one protocol. All methods get the same input points, the same iteration budget and wall-clock cap on one GPU, and the same hyperparameter search size. All meshes are extracted with one extractor, Neural Surface Detection at $128^3$, with MeshUDF as a check. This keeps meshing differences from contaminating the comparison. Beating LevelSetUDF means lower Chamfer distance on ShapeNet-Cars under this shared protocol. It also means higher F-score and normal consistency in near-crease regions, without losing on the flat regions where smoothing is harmless.

The ablation that isolates the key ingredient is a global annealing schedule. It raises a single $\alpha$ over training, using the best of linear, cosine, and step schedules, each tuned with the same budget. A second ablation keeps per-cell release but drives it on a fixed timetable proportional to curvature instead of measured competence. This separates "spatially varying" from "competence-gated." A third removes the re-engagement branch.

Why not just…?

Why not just anneal one global smoothing parameter on a schedule? This is the most serious threat. Coarse-to-fine annealing is a standard trick and may capture most of the benefit. Our argument is that a global schedule must either sharpen hard regions too early, which causes holes at creases and close layers, or sharpen easy regions too late, which blurs them for no reason. Nothing in the file tests this, so it is the central experiment. If tuned global annealing matches the region-wise method within seed variance on the near-crease metrics, we will drop the competence-gating claim and report global annealing as the contribution.

Why not just use DEUDF and skip smoothing entirely? DEUDF already preserves detail by keeping the V and relaxing constraints near the surface. It is the right comparison for sharpness. The file reports that DEUDF remains sensitive to point sparsity, and that is where a temporary smooth phase should help. We test this by subsampling input points to 25% and 10% density. If DEUDF matches us at all densities, the smoothing phase is unnecessary.

Why not just use DUDF with a well-chosen $\alpha$? A fixed $\alpha$ makes one trade-off for the whole shape. DEUDF reports that the rounded profile erases thin edges and folds. DUDF also needs Hessians for normals everywhere. The grid-searched DUDF baseline, together with per-region error broken down by curvature, will show whether any single $\alpha$ comes close.

Why not just close the shape and use a signed distance field? For open interiors this has already failed. NDF reports that signed and occupancy models produced hollow shells on ShapeNet-Cars interiors. Metric–Phase Fields reports signed-distance baselines at 5.591 Chamfer distance on thin shells against its 0.672. We include one signed-distance baseline only as a sanity check.

Proposed Experiments

The benchmarks are ShapeNet-Cars for interiors and thin panels, the Multi-Garment dataset for open hems and cuffs, the Thingi10K and Sketchfab thin-shell set used by Metric–Phase Fields, and the Stanford 3D Scene and KITTI scans for real noise at scale. Metrics are Chamfer-$L_2$ and Chamfer-$L_1$, F-score at 0.0025, and normal consistency. Two diagnostics target the claim. The first is the count of spurious holes and webbing faces along known open boundaries. The second is Chamfer distance restricted to points within a few voxels of high-curvature creases or of a second parallel surface. We also report training time and peak memory, because tracking per-cell state and partial Hessians has a cost.

The decisive checks are threefold. First, the region-wise method must beat tuned global annealing on the crease-restricted and double-layer metrics on ShapeNet-Cars and the garment data, across five seeds. Second, the $\alpha$ maps must show release order correlating with geometric difficulty. Third, freezing all cells at full sharpness after training must cost little accuracy. If the first check fails, the idea reduces to a schedule. If the third fails, the field has learned to depend on the smoothing rather than outgrow it. Either way, the result settles whether competence-gated release adds anything beyond annealing.