View Cone Sampling (VCS)

• VCS is a probabilistic ray sampling approach that replaces single-ray sampling with a Gaussian-distributed bundle of rays within a small angular cone matching the human fovea.
• It uses Gaussian angular sampling and weighted intersections to create smoother, more robust saliency maps that tackle high-frequency textures and complex mesh geometries.
• Quantitative studies show that VCS offers improved consistency and coverage—up to 31× enhancement—over traditional methods in VR eye-tracking pipelines.

View Cone Sampling (VCS) is a probabilistic ray sampling methodology designed to improve 3D mesh saliency ground-truth (GT) acquisition in virtual reality (VR), particularly for eye-tracking pipelines. Unlike classical single-ray gaze sampling, VCS simulates the finite spatial extent of the human foveal receptive field by casting a Gaussian-distributed bundle of rays within a small angular cone centered on the gaze direction. This approach enhances robustness to high-frequency textures, geometric sparsity, and topologically complex mesh regions, mitigating aliasing artifacts and signal discontinuities prevalent in single-ray methods (Zheng et al., 6 Jan 2026).

1. Conceptual Motivation and Definition

VCS arises from the inadequacy of conventional single-ray sampling to account for the spatial spread of human visual attention. Traditional VR eye-tracking maps a user's gaze to a solitary zero-area ray intersecting the 3D mesh, providing limited surface coverage and being highly sensitive to missed or noisy intersections, especially on textured or punctured surfaces. VCS instead forms a circular sampling cone $\mathcal{C}$ of apex $O$ (eye position), axis $d_0$ (gaze vector), and full angle $R_f$ (typically $R_f \approx 5^\circ$), mimicking the approximate angular extent of the human fovea. Sampling $M$ rays per cone, distributed according to a zero-mean angular Gaussian in cone-centric coordinates, each intersection is weighted to reflect its angular proximity to the central axis, producing saliency maps that more faithfully represent perceptual foveation and local attention (Zheng et al., 6 Jan 2026).

2. Mathematical Formalism of Gaussian Angular Sampling

Let $\theta$ be the angular deviation from the central axis $d_0$, $\phi$ the azimuthal angle, and $\sigma_1 = R_f / 6$ to ensure $O$0 covers nearly the entire cone, following the Gaussian $O$1 rule. The angular deviation $O$2 is sampled from a truncated normal $O$3 with $O$4, and $O$5. Using the Box–Muller transform, for uniform $O$6:

$O$7

$O$8

$O$9

$d_0$0

The final world-space direction for the $d_0$1th sample ray, $d_0$2, is

$d_0$3

where $d_0$4 is the alignment matrix mapping the local $d_0$5 to $d_0$6, and $d_0$7, $d_0$8 are standard rotation matrices. The angular joint probability density is

$d_0$9

valid for $R_f$0.

3. Algorithmic Procedure and Implementation Details

Within eye-tracking acquisition, the core pseudocode workflow for VCS comprises:

$M$7

Within real-time VR pipelines, GPU/CPU-resident BVHs or engine-level colliders enable efficient ray intersection. Per-face saliency counts $R_f$1 are incremented for each valid hit, and adjacency lists for mesh faces/vertices are maintained for subsequent geodesic smoothing operations.

4. Comparative Robustness and Quantitative Effectiveness

VCS exhibits substantial robustness and accuracy improvements over single-ray methods. Qualitatively, VCS fills spatial and topological gaps unaddressed by single-ray sampling, yielding smooth, blob-like attention maps even before subsequent diffusion. In saliency alignment and statistical stability, the following representative metrics were obtained (Zheng et al., 6 Jan 2026):

Metric	Single Ray (SR)	VCS (with HCD)
Internal Consistency (IC)	0.0557	0.8137
Correlation Coefficient (CC)	0.1970	0.4829
KL-Divergence	3.2092	1.1278
sAUC	0.7865	0.8288

Additionally, sampling coverage is improved by factors ranging from $R_f$2 to $R_f$3 for mesh sizes up to 1M faces. Ablation studies indicate VCS-derived saliency peaks are more tightly aligned to ground-truth eye-tracking densities.

5. Hyperparameterization and Tuning Strategies

Salient hyperparameters in VCS are:

• Cone apex angle $R_f$4: $R_f$5 (aligned to foveal receptive field estimates).
• Angular standard deviation $R_f$6: $R_f$7, ensuring 99.7% inclusion within the cone.
• Rays per cone $R_f$8: $R_f$9–$R_f \approx 5^\circ$0, trading sampling density with real-time performance ($R_f \approx 5^\circ$1 Hz on GTX-1080 hardware).
• Backface threshold $R_f \approx 5^\circ$2: $R_f \approx 5^\circ$3 (rejects rays at incidence angles $R_f \approx 5^\circ$4).
• Post-diffusion geodesic $R_f \approx 5^\circ$5: $R_f \approx 5^\circ$6 (applied in subsequent HCD geodesic smoothing; not directly part of VCS itself).

If $R_f \approx 5^\circ$7 is too small, under-sampled regions yield noisy or unstable saliency with degraded IC. Excessive $R_f \approx 5^\circ$8 yields diminishing quality improvements but increases computational cost. The $R_f \approx 5^\circ$9 ratio regulates the central weighting of rays; $M$0 emerges as a strong empirical choice.

6. Integration in VR Eye-Tracking Pipelines

VCS is deployed as follows:

• Raw eye-tracker outputs (pupil center, corneal glints) are mapped to a 3D gaze vector $M$1 in eye-camera coordinates, then transformed to world-space $M$2 using the head's 6-DoF pose.
• The alignment matrix $M$3 is computed per frame to map canonical $M$4 to $M$5.
• Efficient mesh intersection is achieved through precomputed BVHs or physics engine infrastructure (e.g., Unity3D colliders).
• Each cone sampling event records valid hits, with optional per-hit Gaussian weighting $M$6.
• Almost all of the pipeline operates at real-time frame rates with 200–500 rays per cone, with mesh adjacency information reserved for smoothing and diffusion.

Handling non-manifold mesh configurations and excluding back-facing or grazing intersections (by dot product threshold) are key for precise and topologically valid hit registration.

7. Visual Interpretation and Figure Annotations

Figure 1(b) (cross-sectional diagram) illustrates the eye at the cone apex with the central gaze (solid line) and a radially fanning bundle of rays. Rays are densest near the axis and sparse towards the cone perimeter. Figure 1(c) (projection diagram) depicts the ray distribution in top view: a Gaussian radial density with central clustering fading towards the boundary, confirming the probabilistic spread of sampling directions.

In synthesis, VCS operationalizes a “foveal-field” model that supersedes single-pixel gaze sampling. The approach's Gaussian angular ray distribution and precise hit filtering establish a robust and perception-aligned foundation for mesh saliency acquisition. When combined downstream with Hybrid Manifold–Euclidean Diffusion, VCS enables perceptual fidelity and statistical robustness, even for large-scale, high-resolution, and topologically intricate meshes (Zheng et al., 6 Jan 2026).