Triangular-Masked Convolution: Domain-Specific

• Triangular-masked convolution is a family of operators that constrains the local support to a triangular pattern, adapted to different domain geometries.
• It is applied in three distinct settings: vertex-domain filters on semi-regular meshes, face-domain filters on arbitrary meshes, and upper-triangular masks for blind-spot denoising.
• The approach offers computational efficiency and practical invariance, reducing the need for re-meshing and enabling improved performance in classification and denoising tasks.

Searching arXiv for the cited papers to ground the article in the published record. Triangular-masked convolution denotes a family of convolution operators in which the admissible local support is constrained by triangular structure rather than by a fully isotropic stencil. In the recent literature, the term is used in at least three technically distinct settings: a vertex-domain filter on semi-regular triangulated meshes, a face-domain operator on arbitrary triangular meshes, and an upper-triangular masking of a $3\times 3$ image kernel for blind-spot denoising. The common element is not a single canonical formula, but the imposition of a triangular support pattern adapted to a particular domain geometry or statistical constraint (Liu et al., 2019, Bregeon et al., 7 Jan 2025, Park et al., 6 Apr 2026).

1. Terminological scope in current literature

The phrase “triangular-masked convolution” is not tied to a single architectural primitive. In Liu et al., it refers to convolution on a semi-regular triangulated mesh, where the local support is the center vertex plus six ordered one-ring neighbors. In MeshConv3D, the triangular mask is a support region over adjacent triangular faces. In TM-BSN, it denotes a fixed binary mask that preserves only the upper-triangular half of a $3\times 3$ kernel (Liu et al., 2019, Bregeon et al., 7 Jan 2025, Park et al., 6 Apr 2026).

Formulation	Support definition	Reported purpose
Semi-regular mesh filter (Liu et al., 2019)	Center vertex plus six ordered one-ring neighbors	Direct spatial CNN on a vertex domain
MeshConv3D face operator (Bregeon et al., 7 Jan 2025)	Central face plus $K$ neighboring faces grown by breadth-first expansion	Efficient convolution on meshes of arbitrary topology
TM-BSN image operator (Park et al., 6 Apr 2026)	Upper-triangular binary mask on a $3\times 3$ kernel, then shift and four-way rotation	Diamond-shaped blind spot for real-world self-supervised denoising

A recurrent misconception is to treat the term as synonymous with blind-spot convolution. The published record does not support that restriction: blind-spot construction is specific to TM-BSN, whereas the mesh papers use triangular masking to define local support on triangulated domains rather than to exclude a target sample.

2. Vertex-domain triangular masks on semi-regular triangulated meshes

On a semi-regular triangulated mesh $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$, each vertex $x_i$ has approximately six neighbors. Liu et al. define the one-ring neighborhood as

$N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$

together with the center $P[i,0]\equiv x_i$. The six neighbors are ordered clockwise on the tangent plane at $x_i$, obtained by projecting the neighbors onto the plane tangent to the approximating sphere $\mathcal S_i$ at $3\times 3$0, then sorting by angle from the local $3\times 3$1-axis (Liu et al., 2019).

The resulting 1-ring triangular-masked filter is

$3\times 3$2

and convolution of a vertex signal $3\times 3$3 is defined by

$3\times 3$4

Equivalently, with the data matrix $3\times 3$5 given by $3\times 3$6, one has

$3\times 3$7

No extra normalization is imposed on $3\times 3$8, and boundary vertices of valence $3\times 3$9 are handled via zero-padding of the undefined $K$0. The same framework extends to a $K$1-ring triangular-masked filter with support size

$K$2

which the paper states exactly parallels the support size of a $K$3 finite difference filter on a regular grid.

The accompanying multiresolution scheme is defined by midpoint subdivision,

$K$4

where each triangle at level $K$5 is split into 4 sub-triangles at level $K$6, hence $K$7. Mean-pooling and max-pooling from level $K$8 back to level $K$9 are taken over $3\times 3$0, consisting of a vertex and its six new one-ring children. By repeated application, this yields down-sampling factors $3\times 3$1 in vertex count.

Within the reported CNN, four feature-extraction blocks $3\times 3$2 apply convolution, ReLU, and mean-pool with stride $3\times 3$3, using filter counts $3\times 3$4, $3\times 3$5, $3\times 3$6, and $3\times 3$7. A final block $3\times 3$8 uses a fully connected layer of 512 units followed by softmax or 2-way logistic classification.

3. Face-domain triangular masks on arbitrary triangular meshes

MeshConv3D shifts the locus of convolution from vertices to faces. For an input triangular mesh $3\times 3$9, each face $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$0 is associated with a descriptor $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$1, and the face-adjacency matrix $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$2 lists up to three neighbors per face, zero-padded at boundaries. The support region around a central face $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$3 is

$T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$4

where $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$5 is the user-defined kernel size in number of neighbors (Bregeon et al., 7 Jan 2025).

The convolutional response is

$T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$6

with learnable matrices $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$7. Because both neighbor terms are sums, the operator is invariant to permutations of the neighbors. The paper therefore avoids the need to define a canonical ordering of the $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$8 faces around $T=(\{x_i\}_{i=1}^N,\{\Sigma_{ijk}\})$9.

The support region is constructed by breadth-first expansion. It is initialized as the central face plus its up to three 1-ring neighbors, expanded by repeatedly adding neighbors of newly added faces until $x_i$0 faces are collected, and truncated if necessary when more than $x_i$1 faces are reached. In indicator form,

$x_i$2

A central design claim of MeshConv3D is that it does not rely on local barycentric coordinates, charting, or alignment of a continuous kernel over each face. Instead, each face is treated as a discrete “pixel” in a graph-like structure. The use of summed neighbor descriptors and summed absolute differences is described as enforcing rotation/translation invariance and removing the need for consistent face orientation. This is the basis for the paper’s statement that the operator is valid on arbitrary topology without pre-remeshing or canonical parameterization.

The abstract places this convolution inside a broader mesh-dedicated methodology that also includes face collapse-based pooling operators and states that the method operates directly on meshes of arbitrary topology, without any need of prior re-meshing or conversion techniques.

4. Upper-triangular kernel masking in blind-spot denoising

In TM-BSN, triangular-masked convolution is an image-space operator defined by a fixed binary mask on a standard convolutional kernel. Let $x_i$3 with $x_i$4, let $x_i$5, let $x_i$6 denote element-wise multiplication, and let $x_i$7 denote convolution. A single layer computes

$x_i$8

with

$x_i$9

For the $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$0 case,

$N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$1

so all lower-left weights are permanently zero (Park et al., 6 Apr 2026).

A single $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$2 TMC at pixel $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$3 aggregates only from neighbors $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$4 for which $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$5, described in the paper as the upper-triangle above the main diagonal. Stacking $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$6 such layers yields the $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$7-fold Minkowski sum of the $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$8 triangle; for $N(x_i)=\{P[i,1],P[i,2],\dots,P[i,6]\},$9, the receptive field becomes a $P[i,0]\equiv x_i$0 triangular region extending further in the north-east direction.

Blind-spot formation is not achieved by the triangular mask alone. TM-BSN adds a feature-map shift by $P[i,0]\equiv x_i$1 pixels upward or rightward so that the output at a location has no direct connection to the target pixel, and then applies four-way rotation at $P[i,0]\equiv x_i$2, $P[i,0]\equiv x_i$3, $P[i,0]\equiv x_i$4, and $P[i,0]\equiv x_i$5. After inverse rotation and merging, the intersection of the four shifted triangular supports forms a centered diamond-shaped blind spot. The paper states that all pixels within radius $P[i,0]\equiv x_i$6 in the $P[i,0]\equiv x_i$7 sense are excluded, while the remaining four corners are fully used.

The reported motivation is specific to real sRGB denoising. TM-BSN attributes spatially correlated noise to the camera ISP pipeline, particularly demosaicing, and states that this correlation follows a diamond-shaped pattern because each pixel is reconstructed from neighboring samples with spatially decaying weights. On that account, the triangular-masked convolution is introduced to align the receptive field with the geometry of the correlation pattern, without downsampling or post-processing.

5. Ordering, invariance, and exclusion mechanisms

The three formulations differ sharply in how triangular support is coupled to ordering and alignment. On semi-regular meshes, the operator depends on an explicit clockwise ordering of six neighbors on the tangent plane, and the analogy to classic spatial convolution is direct: the local stencil is fixed once this ordering has been established (Liu et al., 2019).

MeshConv3D takes the opposite position on ordering. Its face-centered operator is built from permutation-invariant sums and absolute differences, precisely to avoid the hard problem of defining a canonical ordering of neighboring faces. The paper further states that it does not rely on local barycentric coordinates, charting, or consistent face orientation (Bregeon et al., 7 Jan 2025).

TM-BSN uses a third strategy. The upper-triangular mask is fixed in kernel coordinates, and asymmetry is intentional rather than incidental. The goal is not orientation invariance but architectural exclusion: the masked kernel, shift, and four-way rotation jointly carve out the correlated $P[i,0]\equiv x_i$8-diamond around the target pixel. In the paper’s comparison, this differs from zero-center blind-spot kernels, which remove only the target pixel while still including strongly correlated neighbors, and from pixel-shuffle downsampling, which breaks correlations but changes image statistics, reduces resolution, and demands post-processing (Park et al., 6 Apr 2026).

A plausible implication is that “triangular masking” should be understood as a structural constraint on support, not as a single invariance principle. In one setting it relies on ordered local coordinates, in another it is explicitly ordering-invariant, and in the blind-spot setting it is a mechanism for removing access to a statistically undesirable neighborhood.

6. Computational characteristics and reported empirical behavior

The computational profile of triangular-masked convolution depends on the domain and the implementation. On semi-regular meshes, each 1-ring filter has 7 weights plus 1 bias, and the reported model trains in approximately $P[i,0]\equiv x_i$9 s/epoch on an NVIDIA Tesla M40 (24 GB), versus $x_i$0 s/epoch for the Chebyshev-polynomial spectral G-CNN of Defferrard et al. (2016), which the paper reports as a $x_i$1 speedup (Liu et al., 2019). On the ADNI-2 cohort of 3 169 scans with 10-fold cross-validation and no subject overlap between train and test, the vertex-based CNN attains higher accuracy, sensitivity, specificity, and G-mean in five out of six binary tasks; for CON vs AD classification, the paper reports $x_i$2 for the vertex CNN and $x_i$3 for the graph CNN.

For MeshConv3D, the reported forward complexity per layer is $x_i$4, ignoring the minor overhead of breadth-first expansion, with adjacency storage $x_i$5, feature storage $x_i$6, and weights $x_i$7 (Bregeon et al., 7 Jan 2025). On a SHREC11 mesh-batch of 50 faces $x_i$8 50 meshes, the paper states that MeshConv3D ran in $x_i$9 the baseline time, MeshCNN took approximately $\mathcal S_i$0 longer and used more GPU memory, MeshNet++ took approximately $\mathcal S_i$1 longer, and SubdivNet ran at $\mathcal S_i$2 the time but required $\mathcal S_i$3 more GPU memory because of remeshing to a Loop subdivision structure with $\mathcal S_i$4–$\mathcal S_i$5 more faces. The abstract summarizes the resulting classification behavior on three benchmark datasets as equivalent or superior classification results with minimized memory footprint and computational load.

For TM-BSN, a TMC layer is described as having essentially the same FLOPs as a standard $\mathcal S_i$6 convolution because the extra element-wise multiplication by the mask is negligible at runtime (Park et al., 6 Apr 2026). The full teacher model, including one forward pass, four rotations, shift concatenation, and a $\mathcal S_i$7 convolution, costs about $\mathcal S_i$8 GFLOPs, has $\mathcal S_i$9 M parameters, and runs in approximately $3\times 3$00 ms on a $3\times 3$01 patch. The distilled student, TM-BSN (D), is a $3\times 3$02 M-parameter U-Net with approximately $3\times 3$03 GFLOPs and $3\times 3$04 ms runtime. The paper reports $3\times 3$05 dB PSNR and $3\times 3$06 SSIM on SIDD validation, exceeding APR(RD) by $3\times 3$07 dB, and $3\times 3$08 dB PSNR and $3\times 3$09 SSIM on DND in the fully self-supervised setting, $3\times 3$10 dB above TBSN.

The TM-BSN ablations also make the masking geometry operationally specific. With shift offset $3\times 3$11, the paper reports identity mapping, described as the network learning to copy noise; with $3\times 3$12, useful near-neighbor context is discarded and PSNR decreases; $3\times 3$13 is identified as optimal. For distillation, the offset set $3\times 3$14 yields the best student performance at $3\times 3$15 dB, whereas including offset $3\times 3$16 or offsets $3\times 3$17 degrades texture restoration.

Taken together, these results indicate that triangular-masked convolution is best understood as a domain-adapted support design. On semi-regular meshes it produces compact spatial filters; on arbitrary triangular meshes it yields ordering-invariant aggregation without re-meshing; and in self-supervised denoising it constructs a diamond-shaped blind spot matched to the reported correlation structure of demosaiced sRGB noise.