Non-Parametric Visibility Modeling

• Non-Parametric Visibility Modelling is a family of methods that compute visibility directly as continuous fields or set-valued maps from occupancy data and free-space geometry.
• It encompasses formulations like grid-based PDE transport, differentiable volumetric models, and exact geometric functionals, each addressing occlusion with either smooth or sharp boundaries.
• This approach enhances optimization and computational efficiency in applications such as robotics, graphics, and vision by bypassing fixed occlusion templates.

Searching arXiv for recent and foundational papers on visibility modelling across geometry, graphics, robotics, and occlusion-aware vision. {"query":"visibility modelling non-parametric visibility differentiable visibility polygonal environments visibility graphs occlusion arXiv","max_results":10} Searching arXiv more specifically for non-parametric or continuous visibility formulations in reconstruction, robotics, and geometry. {"query":"site:arxiv.org non-parametric visibility modeling differentiable visibility pose estimation grid visibility polygonal environments arXiv", "max_results": 10} Non-parametric visibility modelling denotes a family of formulations in which visibility is represented as a field, set-valued map, latent mask, or geometric functional defined directly from occupancy, image evidence, or free-space geometry, rather than solely as a binary z-buffer event or a small family of occlusion templates. In the literature, the term is used in more than one sense. It can refer to visibility as a continuous function on a discretized domain, as in grid-based PDE transport; to visibility as a high-dimensional latent mask with only weak local priors; or to exact geometric visibility metrics derived from visibility polygons in polygonal free space. Closely related work also includes continuous volumetric models and attenuation-weighted visibility graphs that are not fully non-parametric in the strict statistical sense, but clarify how visibility can be made smooth, graded, or context dependent (Rhodin et al., 2016, Ibrahim et al., 2024, Ott et al., 2013, Schwartz et al., 2021, Banzal et al., 29 Jan 2026).

1. Conceptual scope and meanings of the term

The phrase “non-parametric visibility” does not identify a single canonical formalism. Instead, the literature exhibits several technically distinct meanings. In Cartesian occupancy grids, visibility is a scalar field $u:\mathbb{Z}^2 \to [0,1]$ computed directly from per-cell occupancy, with no explicit parameterization of obstacle geometry, no dependence on the number of obstacles, and no visibility graph or triangulation; this is non-parametric in the sense of a grid-defined field over a discretized domain (Ibrahim et al., 2024). In occlusion-aware object detection, visibility is a latent binary flag $v_i$ for each block in a sliding window, yielding a combinatorial space of masks regularized only by an Ising prior; this is non-parametric or very weakly parametric because there is no fixed template family such as “left-half occluded” or “bottom-third occluded” (Ott et al., 2013). In polygonal environments with holes, visibility is the exact set-valued map $S(x)=\{y\in F\mid [x,y]\subset F\}$ and the derived area metric $V(x)=\mu(S(x)\cap D_2)$; here non-parametric means that the model is determined by the geometry of the free space and the visibility polygon itself, not by a low-dimensional surrogate (Banzal et al., 29 Jan 2026).

A parallel strand of work uses smooth or graded visibility representations that are continuous but not fully non-parametric. Rhodin et al. replace opaque surfaces by a translucent Gaussian medium so that visibility becomes Beer–Lambert transmittance through a density field $D(\mathbf{x})$; they explicitly note that the model is still parameterized by Gaussian parameters, but that there is no explicit parametric model of occlusion boundaries such as silhouettes or z-buffer layers (Rhodin et al., 2016). In visibility graph analysis under weather, edge weights are defined by $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$, which is physically motivated and continuous, but explicitly parametric in $\sigma$ and distance (Schwartz et al., 2021).

Formulation	Visibility representation	Status in source
Differentiable scene model	Beer–Lambert transmittance through Gaussian density	Continuous volumetric; no explicit parametric occlusion boundaries
Cartesian grid solver	Scalar field $u(x,y)$ on an occupancy grid	Non-parametric field on a discretized domain
Partial-visibility detector	Latent block mask $v$ with MRF prior	Non-parametric or very weakly parametric
Weather-weighted VGA	Edge weight $w(e_{ij})=\exp(-\sigma d_{ij})$	Parametric attenuation model
Polygonal hide/surveillance model	Visibility polygon $v_i$0 and area metric $v_i$1	Non-parametric, non-smooth geometric model
NeuralPVS	CNN-predicted binary visibility volume	Parametric learned alternative

A common misconception is that “non-parametric” means “parameter-free.” The surveyed work does not support that equivalence. Several models remain parameterized while abandoning explicit parametric boundary descriptions; others are genuinely non-parametric in the stronger sense that visibility is computed directly as a field or geometric set from occupancy or free-space structure (Rhodin et al., 2016, Schwartz et al., 2021).

2. Continuous volumetric visibility and differentiability

A major development in visibility modelling is the replacement of hard, opaque surfaces by a smooth volumetric density field so that visibility becomes differentiable. Rhodin et al. define a scalar density field

$v_i$2

where each isotropic 3D Gaussian has magnitude $v_i$3, center $v_i$4, standard deviation $v_i$5, and albedo $v_i$6 (Rhodin et al., 2016). Visibility is then interpreted as transmittance along a ray $v_i$7,

$v_i$8

so that classical binary visibility is replaced by a smooth attenuation law. Along a given ray, each 3D Gaussian reduces to a 1D Gaussian in depth, yielding a closed-form transmittance involving the error function. The resulting radiance model is

$v_i$9

which is approximated by Gaussian-centered depth samples to define per-Gaussian visibility weights $S(x)=\{y\in F\mid [x,y]\subset F\}$0 and the rendered image $S(x)=\{y\in F\mid [x,y]\subset F\}$1 (Rhodin et al., 2016).

The significance of this formulation is not merely smooth rendering. Because Gaussians, exponentials, and error functions are analytic, $S(x)=\{y\in F\mid [x,y]\subset F\}$2, $S(x)=\{y\in F\mid [x,y]\subset F\}$3, and $S(x)=\{y\in F\mid [x,y]\subset F\}$4 are differentiable with respect to Gaussian positions $S(x)=\{y\in F\mid [x,y]\subset F\}$5, scales $S(x)=\{y\in F\mid [x,y]\subset F\}$6, magnitudes $S(x)=\{y\in F\mid [x,y]\subset F\}$7, albedos $S(x)=\{y\in F\mid [x,y]\subset F\}$8, camera parameters, and pose parameters that move the Gaussians through rigid or articulated kinematics. The paper emphasizes that there are no binary decisions at occlusion boundaries; even double occlusion events become smooth transitions. This yields smooth, analytically differentiable, and efficient to optimize pose similarity energies with rigorous occlusion handling, fewer local minima, and experimentally verified improved convergence of numerical optimization (Rhodin et al., 2016).

The formulation is “non-parametric” only in a restricted sense. The scene is parameterized by $S(x)=\{y\in F\mid [x,y]\subset F\}$9, but visibility is not represented by explicit visibility lists, silhouette templates, or z-buffer layers. Occlusion emerges implicitly from the integral of density along rays. This suggests a useful distinction between boundary-parametric visibility, which makes occlusion a discrete combinatorial object, and field-based visibility, which makes occlusion an emergent property of a continuous medium.

Rhodin et al. use the model in marker-less multi-object pose estimation, marker-less human motion capture with few cameras, and image-based 3D geometry estimation. In the “Marker” sequence with 2 cameras, 44 pose parameters, and 72 Gaussians, the reported average 3D joint error is $V(x)=\mu(S(x)\cap D_2)$0 cm versus $V(x)=\mu(S(x)\cap D_2)$1 cm for Stoll et al., described as a 45% improvement; in the “Soccer” sequence, joint tracking of both actors yields $V(x)=\mu(S(x)\cap D_2)$2 improved accuracy over tracking each actor independently (Rhodin et al., 2016). The computational cost is

$V(x)=\mu(S(x)\cap D_2)$3

with 5 samples per Gaussian, occlusion pruning for $V(x)=\mu(S(x)\cap D_2)$4, and about $V(x)=\mu(S(x)\cap D_2)$5 gradient iterations per second on a $V(x)=\mu(S(x)\cap D_2)$6 GHz quad-core CPU for the “Marker” sequence (Rhodin et al., 2016).

3. Visibility as a transported field on Cartesian grids

In grid-based robotics, non-parametric visibility modelling appears in a stricter form: visibility is a scalar field defined directly on the occupancy grid, without any explicit geometric parameterization of obstacles. The 2024 PDE-based method defines visibility $V(x)=\mu(S(x)\cap D_2)$7 on a 2D Cartesian grid and recovers the binary visibility region by thresholding,

$V(x)=\mu(S(x)\cap D_2)$8

where visibility means the existence of at least one uninterrupted line segment from the source point to the cell that does not intersect occupied cells (Ibrahim et al., 2024).

The central equation is a linear first-order hyperbolic PDE,

$V(x)=\mu(S(x)\cap D_2)$9

with $D(\mathbf{x})$0 for straight-line visibility from the source. The coefficients are chosen so that the characteristics coincide with radial rays in the continuous limit (Ibrahim et al., 2024). Occlusion is introduced algebraically through the occupancy complement $D(\mathbf{x})$1 and a decay factor $D(\mathbf{x})$2: $D(\mathbf{x})$3 A fully occupied cell becomes completely dark, and the upwind stencil propagates that occlusion causally to downstream cells (Ibrahim et al., 2024).

The numerical scheme is a first-order entropy-satisfying upwind method with two update forms, chosen according to local slope so that the effective Courant number stays below one. The algorithm performs a single forward sweep over the grid, one update per cell per source, giving compute and memory complexity $D(\mathbf{x})$4 with $D(\mathbf{x})$5, at most ten arithmetic operations per grid cell, and no preprocessing (Ibrahim et al., 2024). The paper explicitly contrasts this with ray-casting and geometric visibility methods whose cost depends on repeated voxel traversals or obstacle geometry. Because the scheme is monotone and entropy-satisfying, it converges to the true visibility polygon as the step size goes to zero (Ibrahim et al., 2024).

This grid-based formulation is non-parametric in a direct sense. Obstacles are encoded only by a per-cell occupancy value $D(\mathbf{x})$6, and visibility emerges from repeated local PDE updates rather than from polygon edges, vertices, line-segment intersection tests, or visibility graphs. The authors also interpret the algorithm as a dynamic-programming recurrence on a field, analogous in structure to distance transforms and fast marching but solving a transport equation rather than an Eikonal equation (Ibrahim et al., 2024).

The reported implementation computes a $D(\mathbf{x})$7 grid visibility field in about $D(\mathbf{x})$8, shows linear scaling from $D(\mathbf{x})$9 to $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$0, is $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$1 faster than ray-casting on an empty $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$2 grid, and up to $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$3 faster on a $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$4 grid (Ibrahim et al., 2024). The same visibility field becomes a heuristic for a deterministic, local-minima-free path planner whose cumulative visibility field is

$w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$5

and whose waypoint heuristic is

$w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$6

A $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$7 maze is reported as solved in about $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$8 ms (Ibrahim et al., 2024).

4. Latent visibility masks and weighted visibility graphs

A different non-parametric tradition arises in computer vision, where visibility is not a geometric field over physical space but a latent field over image support. In sliding-window object detection, partial visibility is modeled by a binary flag $w(e_{ij})=\exp(-\sigma\|\vec v_i-\vec v_j\|)$9 for each HOG block, with the window score

$\sigma$0

and

$\sigma$1

Visible blocks contribute appearance evidence $\sigma$2; occluded blocks ignore the observed features and contribute a learned constant bias $\sigma$3. The pairwise Ising prior penalizes label discontinuities, encouraging spatially contiguous occlusions without imposing a fixed template family (Ott et al., 2013).

This formulation is described as non-parametric or very weakly parametric because the space of admissible visibility patterns is all subsets of the block lattice, constrained only by locality and contiguity. Inference is exact by graph cuts, because the energy is a binary submodular MRF. The detector also records the argmax mask $\sigma$4, allowing visible part identification, and uses an upper bound $\sigma$5 to prune most windows before running the full MRF. On PASCAL VOC 2010, the mean AP reported for the baseline detector is $\sigma$6, rising to $\sigma$7 with partial visibility modelling; the method improves 16 out of 20 classes, and the visibility-aware NMS yields more detailed scene interpretations even though its AP change is modest (Ott et al., 2013).

Weighted visibility graph analysis provides a contrasting case. Classical VGA is binary: an edge exists if there is line of sight. The weather-aware extension retains the geometric line-of-sight test but assigns continuous edge weights

$\sigma$8

together with node summaries

$\sigma$9

Here $u(x,y)$0 is an attenuation coefficient derived from rain, snow, or fog models, and the resulting visibility is continuous rather than binary (Schwartz et al., 2021).

The important interpretive point is that this model is not non-parametric in the strict sense. The source explicitly characterizes it as parametric: the decay law is prescribed as an exponential, and $u(x,y)$1 follows specific physical formulas. Yet the graph formulation creates a template in which visibility becomes a real-valued weight rather than an adjacency bit, and the authors explicitly note that this structure could support arbitrary empirically estimated functions or lookup tables (Schwartz et al., 2021). The case studies show that attenuation can re-rank locations: in a city example under clear weather, $u(x,y)$2 has the highest $u(x,y)$3, but under dry snow $u(x,y)$4 has the highest total visibility (Schwartz et al., 2021). This suggests that graded visibility can expose context dependence invisible to binary line-of-sight analysis even when the underlying weighting function remains parametric.

5. Exact geometric visibility as a non-smooth functional

In polygonal environments with holes, visibility can be modeled exactly as a non-smooth geometric functional. Let $u(x,y)$5 be the polygonal free space, obtained from a polygonal environment $u(x,y)$6 by removing the interiors of polygonal obstacles. The visibility region of $u(x,y)$7 is

$u(x,y)$8

and for a target domain $u(x,y)$9, the visibility metric is

$v$0

where $v$1 is Lebesgue measure on $v$2 (Banzal et al., 29 Jan 2026). Minimizing $v$3 corresponds to hiding, while maximizing it corresponds to surveillance.

The technical contribution is an explicit non-smooth analysis of $v$4 and $v$5. Visibility changes are organized by reflex vertices, anchors, and inflection segments. A reflex vertex $v$6 is an anchor for observer $v$7 if $v$8 and the projected ray $v$9; the vertices of the visibility polygon $w(e_{ij})=\exp(-\sigma d_{ij})$0 are exactly the visible vertices of $w(e_{ij})=\exp(-\sigma d_{ij})$1 together with the projected intersections generated by the anchors (Banzal et al., 29 Jan 2026). Critical points are positions where the anchor set changes under arbitrarily small perturbations. The paper proves that the set of critical points $w(e_{ij})=\exp(-\sigma d_{ij})$2 coincides with the union $w(e_{ij})=\exp(-\sigma d_{ij})$3 of a finite family of inflection segments generated by single reflex vertices and mutually visible pairs of reflex vertices: $w(e_{ij})=\exp(-\sigma d_{ij})$4 Removing $w(e_{ij})=\exp(-\sigma d_{ij})$5 partitions the reduced free space $w(e_{ij})=\exp(-\sigma d_{ij})$6 into convex components on each of which the anchor set is constant (Banzal et al., 29 Jan 2026).

This decomposition yields a piecewise-smooth theory. The paper introduces $w(e_{ij})=\exp(-\sigma d_{ij})$7-local Lipschitzness for set-valued maps under the symmetric-difference metric and shows that both $w(e_{ij})=\exp(-\sigma d_{ij})$8 and $w(e_{ij})=\exp(-\sigma d_{ij})$9 are $v_i$00-locally Lipschitz on $v_i$01, implying local Lipschitzness of the scalar metrics $v_i$02 and $v_i$03 there (Banzal et al., 29 Jan 2026). On each convex component with fixed anchor set, $v_i$04 is real analytic and $v_i$05 is $v_i$06. The directional derivative of visibility area is expressed as a sum of anchor contributions,

$v_i$07

and the visibility metric adds weights $v_i$08 that account for the fraction of the infinitesimal area change lying in $v_i$09 (Banzal et al., 29 Jan 2026). The Clarke generalized gradient is then the convex hull of gradient limits taken from neighboring smooth cells.

The optimization problem is non-convex and non-smooth, so off-the-shelf methods can only guarantee convergence to Clarke critical points. To address this, the paper introduces the Normalized Descent algorithm, or Norcent, which computes numerical generalized gradients, normalizes step directions, and injects random perturbations when the generalized gradient norm is below a threshold. The analysis allows non-monotonic decrease in the objective and proves almost sure convergence to the set of local minima of the true geometric metric (Banzal et al., 29 Jan 2026). This is a strong form of non-parametric visibility modelling: the optimized quantity is the exact area of the true visibility polygon intersected with a domain, not a smoothed surrogate.

6. Relation to parametric and learned visibility models

The surveyed literature makes it clear that non-parametric visibility modelling is best understood relative to parametric alternatives. Continuous visibility does not imply non-parametric status, and dense output does not imply geometric exactness. Rhodin et al.’s differentiable Gaussian scene model is continuous, analytic, and boundary-free, but still parameterized by $v_i$10; the source therefore treats it as “continuous, volumetric, Gaussian-basis visibility” rather than fully non-parametric visibility (Rhodin et al., 2016). The weather-weighted visibility graph is likewise continuous and context dependent, but the weighting law $v_i$11 is explicitly parametric (Schwartz et al., 2021).

This distinction becomes sharper in learned visibility computation. NeuralPVS is explicitly introduced as a parametric, learned alternative to classical non-parametric or geometric PVS computation. It maps a froxelized occupancy volume $v_i$12 to a binary visibility volume $v_i$13 by a sparse 3D CNN with volume-preserving interleaving, learning the mapping $v_i$14 from sampled from-region PVS data (Wang et al., 29 Sep 2025). The method reports approximately $v_i$15 Hz processing with less than $v_i$16 missing geometry, average false negative rates typically $v_i$17–$v_i$18, false positive rates typically $v_i$19–$v_i$20, and SSIM $v_i$21 across test scenes; it is also reported as faster than Trim Regions, at about $v_i$22 ms per PVS versus about $v_i$23–$v_i$24 ms (Wang et al., 29 Sep 2025). Yet its visibility model is stored in network weights $v_i$25, so it is parametric even though its input and output are dense 3D fields.

A plausible synthesis is that visibility modelling can be organized along two largely orthogonal axes. One axis concerns representation: binary adjacency, graded edge weights, latent masks, scalar fields, volumetric densities, or exact visibility polygons. The other concerns epistemic status: explicit physical formulas, exact geometric computation, or learned parametric approximation. Non-parametric visibility modelling occupies the part of this space where visibility is computed directly as a field or geometric set from occupancy or free-space geometry, or inferred as a high-dimensional latent mask without a small prescribed template family. Parametric models remain important in this history because they often expose which aspects of visibility can be smoothed, weighted, or amortized without changing the underlying problem statement (Ibrahim et al., 2024, Ott et al., 2013, Banzal et al., 29 Jan 2026, Wang et al., 29 Sep 2025).

From that perspective, the field supports three stable conclusions. First, many hard visibility problems become substantially better behaved when recast as continuous fields or set-valued functionals rather than binary events. Second, the strongest non-parametric formulations tie complexity to discretization or geometric structure rather than to pre-specified boundary templates. Third, the main technical trade-off is persistent across domains: exactness and sharp boundaries tend to produce non-smooth, combinatorial optimization landscapes, whereas smooth surrogates and learned approximations improve optimization or runtime at the cost of blur, approximation error, or dependence on a chosen parameterization.