Corner Centered Layout: Principles & Applications

• Corner-centered layout is a design paradigm that uses sparsely placed corner anchors to efficiently represent and reconstruct global spatial structures.
• In panoramic indoor scene estimation, methods like SSLayout360 and CFL exploit 1D and 2D corner signals to recover 3D room geometry with improved label efficiency and deterministic post-processing.
• Beyond vision, the concept supports controlled generative image editing and algebraic placements in integrable systems by emphasizing explicit geometric conditioning and sparse-anchor formulations.

Searching arXiv for the provided topic and cited works to ground the article in current records. Corner-centered layout is a representational and algorithmic design in which corners, or other corner-adjacent structural anchors, are treated as the primary variables from which a larger spatial configuration is inferred. The term appears in distinct technical senses across computer vision, generative image editing, discrete integrable systems, and layout adjustment. In 360° indoor scene understanding, it denotes room representations centered on wall–wall junctions and boundary curves rather than dense pixel labels (Tran, 2021, Fernandez-Labrador et al., 2019). In identity-aware image editing, it denotes a specific composite canvas that places four references at the corners and the editable region at the center to balance attention geometry for FLUX.1-Fill (Xiong et al., 24 Aug 2025). In the theory of face-centered cubic integrable systems, it distinguishes equations assigned to cube corners from those assigned to face centers (Kels, 2021). A related but more general notion also appears in center-based rectangle adjustment, where orthogonal order is preserved through reference-point coordinates (Bandyapadhyay et al., 2015). Across these uses, the recurring idea is that a sparse set of structurally privileged locations governs global organization.

1. Corner-centered layout in panoramic indoor scene reconstruction

In 360° indoor layout estimation, corner-centered layout refers to a representation in which 3D room geometry is predicted primarily via room corners and the boundary curves that pass through them, rather than via dense pixel-wise semantic labels or full 2D floor-plan masks (Tran, 2021). SSLayout360 defines the task on a single 360° RGB panorama in equirectangular projection and represents the room through floor-wall boundaries, ceiling-wall boundaries, and wall-wall boundaries (“corners”) (Tran, 2021). Under the Manhattan-world assumptions stated in the paper—vertical walls, horizontal floor and ceiling, right angles in the floor plan, and panorama alignment—the room can be reconstructed from the vertical positions of the floor-wall and ceiling-wall curves at each horizontal direction together with the presence or absence of a corner at each column (Tran, 2021).

SSLayout360 adopts HorizonNet’s 1D corner-based parameterization. For an input panorama $x \in \mathbb{R}^{3 \times 512 \times 1024}$, the target layout is encoded as

$$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$

where $y_c(j)$ is the ceiling-wall boundary vertical position at column $j$, $y_f(j)$ is the floor-wall boundary vertical position, and $y_w(j)$ is the corner probability (Tran, 2021). This reduces the prediction problem to three 1D signals over azimuth. Corners are represented as peaks in $y_w$, and walls are the intervals between peaks, whose vertical extents are specified by $y_f$ and $y_c$ (Tran, 2021).

The network architecture remains HorizonNet’s encoder–decoder design: a ResNet-50 encoder pre-trained on ImageNet followed by convolution + ReLU layers producing a $1024 \times 1 \times 256$ feature tensor, and a Bi-RNN decoder that outputs $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$0 (Tran, 2021). The centrality of corners here is not merely descriptive. The 1D representation is the basis for both the supervised and consistency losses, as well as for deterministic geometric post-processing.

A related but distinct corner-centered formulation appears in CFL, which predicts 2D corner and edge probability maps directly from 360° panoramas and then recovers the full 3D layout from the corner predictions without additional optimization or model fitting (Fernandez-Labrador et al., 2019). CFL treats the ordered sequence of spherical corner directions $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$1 as the layout parameterization, permitting arbitrary polygons and relaxing Manhattan and box-shaped assumptions (Fernandez-Labrador et al., 2019). In that sense, SSLayout360 and CFL share a common principle—corners as primary geometric primitives—but implement it differently: SSLayout360 via 1D boundary-plus-corner signals under Manhattan alignment, CFL via 2D corner heatmaps and spherical polygon reconstruction.

2. Representation, inference, and geometry in vision models

The representational difference between SSLayout360 and CFL clarifies two major variants of corner-centered layout in computer vision. SSLayout360 compresses the panorama into azimuth-indexed structural signals, while CFL keeps a dense spatial prediction but uses corners as the decisive primitive for downstream geometry (Tran, 2021, Fernandez-Labrador et al., 2019).

In SSLayout360, layout reconstruction begins by detecting corner columns from $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$2, then using $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$3 and $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$4 as floor and ceiling angles for each column, and finally intersecting the corresponding rays with horizontal floor and ceiling planes (Tran, 2021). The paper gives the yaw angle example $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$5 and the ray direction

$$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$6

which is intersected with the floor plane to obtain 3D floor points (Tran, 2021). Walls are then vertical planes passing through corresponding floor and ceiling points at corner columns, and Manhattan alignment constrains wall normals to axis-aligned directions (Tran, 2021). The full 3D layout thus emerges from a sparse corner-centered representation plus deterministic geometry.

CFL uses a different path. It predicts two binary probability maps: an edge map $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$7 and a corner map $$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$8, with each map defined as

$$y = \begin{bmatrix} y_c \ y_f \ y_w \end{bmatrix} \in \mathbb{R}^{3 \times 1 \times 1024},$$9

for each pixel $y_c(j)$0 (Fernandez-Labrador et al., 2019). Corner locations are extracted as maxima in the corner map, mapped from equirectangular coordinates $y_c(j)$1 to spherical angles

$y_c(j)$2

and then joined in angular order to form a closed polygon on the viewing sphere (Fernandez-Labrador et al., 2019). Floor and ceiling are assumed parallel, and a unit camera height fixes scale up to an arbitrary factor (Fernandez-Labrador et al., 2019). Unlike SSLayout360, CFL does not enforce Manhattan perpendicularity; the corner sequence can therefore encode non-Manhattan rooms and rooms with more than four walls (Fernandez-Labrador et al., 2019).

CFL’s architectural contribution is also tied to corner-centered layout. It proposes EquiConvs, which define kernels as spherical surface patches with angular field of view $y_c(j)$3 and resolution $y_c(j)$4, making convolution invariant to equirectangular distortions (Fernandez-Labrador et al., 2019). The offsets are not learned but deterministically computed from projection geometry, and the result is that corner detection remains more robust under rotations and translations of the panorama than with standard planar convolutions (Fernandez-Labrador et al., 2019). This suggests that when corners are the main latent variables, projection-aware feature extraction becomes especially consequential, because distortions can directly perturb the corner map.

3. Semi-supervision, consistency, and label efficiency

SSLayout360 extends the corner-based HorizonNet parameterization into a semi-supervised student–teacher framework (Tran, 2021). For labeled panoramas $y_c(j)$5 with ground-truth $y_c(j)$6, the supervised loss is

$y_c(j)$7

using $y_c(j)$8 for the continuous boundary channels and squared $y_c(j)$9 for the corner-probability channel (Tran, 2021). The paper explicitly notes that this differs from HorizonNet’s original binary cross-entropy for $j$0, choosing the Brier score because it is bounded and works better with SSL consistency (Tran, 2021).

For unlabeled panoramas, SSLayout360 uses Mean Teacher consistency regularization with teacher parameters updated by exponential moving average: $j$1 The consistency loss mirrors the supervised one: $j$2 and the total objective is

$j$3

with ramp-up

$j$4

The paper stresses that the structural identity of $j$5 and $j$6 makes the scales compatible, so no delicate balancing of $j$7 is needed (Tran, 2021).

The reported results make the label-efficiency implications concrete. On PanoContext with only 20 labels, corner error decreases from $j$8 for HorizonNet to $j$9 for SSLayout360; with 100 labels, it decreases from $y_f(j)$0 to $y_f(j)$1 (Tran, 2021). On MatterportLayout with ST3D pre-training, SSLayout360 with only 200 labels plus unlabeled panoramas matches the performance of a fully supervised HorizonNet trained on all 1,650 real labels, which the paper describes as about 12% of the real labels (Tran, 2021). The synthetic pre-training stage itself uses Structured3D with 21,835 synthetic panoramas, of which 18,362 are used for supervised pre-training (Tran, 2021).

These results are presented in the paper as consequences of consistency-supervised corner prediction: unlabeled panoramas contribute more examples of corner configurations, boundary shapes, and clutter or occlusion patterns, allowing the model to refine the corner-centered representation (Tran, 2021). A plausible implication is that low-dimensional structural targets can be especially amenable to semi-supervision when their regression geometry is shared exactly between labeled and unlabeled branches.

4. Corner-centered layout in identity-aware image editing

In PosBridge, the phrase “Corner Centered Layout” has a different meaning. It names a specific composite input arrangement for FLUX.1-Fill in which the editable target region is centrally located and four reference views are placed symmetrically at the four corners of a single canvas (Xiong et al., 24 Aug 2025). The purpose is to make attention distances from target to references approximately balanced and to provide a geometry on which positional embedding transplant can act (Xiong et al., 24 Aug 2025).

FLUX.1-Fill takes a background image $y_f(j)$2 and a binary mask $y_f(j)$3, from which it constructs three spatial token streams and a text token (Xiong et al., 24 Aug 2025). The paper gives

$y_f(j)$4

$y_f(j)$5

where $y_f(j)$6,

$y_f(j)$7

and

$y_f(j)$8

After linear projection, these are fed to the MMDiT backbone together with the T5 text token (Xiong et al., 24 Aug 2025).

The corner-centered canvas is built from a background $y_f(j)$9, four references $y_w(j)$0, and their segmentation masks $y_w(j)$1. Each reference is cropped around the object as $y_w(j)$2, resized and padded to a corner slot, and then inserted into one of the four corner regions $y_w(j)$3, while the background occupies the center region $y_w(j)$4 (Xiong et al., 24 Aug 2025). The corresponding mask uses the desired insertion region only inside the center and sets the reference regions to all zero. The paper states that this all-black mask for references improves integration of the inserted object with the background (Xiong et al., 24 Aug 2025).

PosBridge then crops the token grid into the four corner patches and the center patch, concatenates them into a compact sequence, and restructures positional embeddings in parallel so that each token retains the rotary positional embedding it had on the original corner-centered canvas (Xiong et al., 24 Aug 2025). During early denoising steps $y_w(j)$5, positional embeddings are transplanted between reference and target tokens so that the target region follows the reference geometry; after $y_w(j)$6, native positional embeddings are restored to allow context-aware blending with the background (Xiong et al., 24 Aug 2025). The paper reports that $y_w(j)$7 gives a good trade-off and that $y_w(j)$8 is used in all experiments, while larger $y_w(j)$9 causes desaturated backgrounds and “ghosting” from references (Xiong et al., 24 Aug 2025).

The role of the corner-centered layout here is therefore not geometric reconstruction but conditioning geometry. By placing references at the four corners and the editable region at the center, the method seeks balanced self-attention, reduced boundary leakage, and more controlled structural transfer than side-by-side concatenation (Xiong et al., 24 Aug 2025). This suggests a broader definition of corner-centered layout: not only a corner-based representation of a scene, but also a corner-based spatial arrangement designed to regulate model behavior.

5. Corner-centered versus face-centered equations on the FCC unit cell

In the theory of discrete integrable equations, the phrase appears in a third and highly specialized sense. The paper “Discrete integrable equations on face-centered cubics: consistency and Lax pairs of corner equations” distinguishes equations assigned to corners of the face-centered cubic unit cell from equations assigned to face centers (Kels, 2021). Here the issue is not visual layout but the lattice placement of five-point equations within the consistency-around-a-face-centered-cube (CAFCC) framework.

The FCC cube carries eight corner variables

$y_w$0

and six face-center variables

$y_w$1

together with rapidity-type parameters

$y_w$2

assigned to the three orthogonal directions (Kels, 2021). A face-centered quad equation is a five-point relation on a square with four corner variables and one face-center variable, multilinear in the corner variables (Kels, 2021).

The central contrast is between the original CAFCC formulation and the new one. In the original formulation, type-A and type-B equations are face-centered, while type-C equations appear only as corner-centered corner equations attached to cube corners (Kels, 2021). In the new formulation, there are two distinct type-C equations,

$y_w$3

with $y_w$4 used as a face-centered equation and $y_w$5 as a corner-centered equation (Kels, 2021). This reassignment allows type-C equations to be treated as independent multidimensionally consistent systems.

CAFCC itself involves fourteen equations on the FCC cube: six face-centered equations and eight corner-centered equations (Kels, 2021). The paper states that in the new formulation the six face-centered equations are given in Eq. (2.15), while the eight corner-centered equations are given in Eq. (2.16), mixing type-A, type-C, and $y_w$6 placements (Kels, 2021). The algorithm then fixes initial data

$y_w$7

and checks consistency by solving successive equations for the remaining variables and requiring agreement along multiple paths (Kels, 2021).

This relocation of type-C from a purely corner-centered role to a dual face-centered and corner-centered role is what enables the paper’s Lax pair construction (Kels, 2021). Since the Lax interpretation requires face-centered geometry, the new arrangement makes it possible to derive Lax matrices directly from type-C equations. In this context, “corner-centered layout” denotes a specific placement of local relations on an FCC cube and is conceptually opposed to face-centered layout rather than to pixel-wise or dense representations.

6. Related notions: robust corner detection and center-based layout adjustment

Two additional papers in the source set do not use the exact phrase as a named method but supply adjacent concepts relevant to corner-centered reasoning.

The “Robust fuzzy corner detector” presents a local, template-based fuzzy rule system that scans an image pixel-by-pixel, computes a $y_w$8 matrix of gray-level differences around each center pixel, and outputs a scalar cornerness value followed by thresholding and non-max suppression (Cuevas et al., 2014). The difference matrix is

$y_w$9

and the detector uses 12 corner templates, a cornerness threshold $y_f$0, and non-max suppression window $y_f$1 (Cuevas et al., 2014). The reported averages are stability $y_f$2, noise immunity $y_f$3, and runtime $y_f$4 s, compared with $y_f$5, $y_f$6, and $y_f$7 s for Harris (Cuevas et al., 2014). Although this is a detector rather than a layout representation, it exemplifies a corner-first pipeline in which higher-level shape analysis is intended to build on stable corner anchors.

“A Constant Factor Approximation for Orthogonal Order Preserving Layout Adjustment” addresses a different notion of layout, namely center-based placement of axis-parallel rectangles subject to non-overlap and orthogonal-order preservation (Bandyapadhyay et al., 2015). Given rectangles $y_f$8 and an injective initial layout $y_f$9, the goal is to find a disjoint layout $y_c$0 preserving the orthogonal ordering of rectangle centers while minimizing the bounding-box area (Bandyapadhyay et al., 2015). The paper proves APX-hardness for constrained variants and gives a $y_c$1-approximation for area via a linear programming relaxation plus a factor-2 scaling step (Bandyapadhyay et al., 2015).

This is not a corner-centered formulation, but it is a reference-point-centered formulation. The contrast is instructive: the indoor-layout and editing papers privilege corners because they are sparse geometric or conditioning anchors; the rectangle-adjustment paper privileges centers because order preservation and non-overlap are naturally expressed in center coordinates (Bandyapadhyay et al., 2015). A plausible implication is that “corner-centered” and “center-based” are best understood as members of a broader family of sparse-anchor layout abstractions, with the choice of anchor determined by the geometry or invariants the task must preserve.

7. Comparative interpretation across fields

Despite the sharp domain differences, the source papers reveal three recurring properties of corner-centered layout.

First, it is a compact structural representation. SSLayout360 replaces dense 2D layout annotation with three 1D signals over panorama columns (Tran, 2021). CFL replaces room-type fitting with corner and edge maps whose corner sequence directly yields a polygonal room model (Fernandez-Labrador et al., 2019). PosBridge reduces a multi-image conditioning problem to a carefully arranged corner-plus-center canvas (Xiong et al., 24 Aug 2025). In each case, structure is compressed into a small number of privileged locations or signals.

Second, it supports explicit geometry or explicit control. In panoramic reconstruction, corners delimit wall segments and enable deterministic projection to 3D planes and polygons (Tran, 2021, Fernandez-Labrador et al., 2019). In PosBridge, corner placement of references is used to control RoPE-mediated structural transfer during denoising (Xiong et al., 24 Aug 2025). In the FCC setting, corner-centered placement specifies where five-point relations live in the cube and therefore what consistency and Lax constructions are possible (Kels, 2021).

Third, the term is domain-specific rather than universal. In computer vision, it refers to corners as geometric primitives of room structure (Tran, 2021, Fernandez-Labrador et al., 2019). In generative editing, it refers to a specific spatial arrangement of inputs on a shared canvas (Xiong et al., 24 Aug 2025). In integrable systems, it refers to the assignment of equations to vertices rather than faces of an FCC unit cell (Kels, 2021). Treating these as interchangeable would be misleading; they share a family resemblance, not a single formal definition.

A common misconception would be to assume that corner-centered layout always implies Manhattan geometry or box-like rooms. CFL explicitly rejects both constraints, using only parallel floor and ceiling planes and allowing arbitrary polygonal rooms (Fernandez-Labrador et al., 2019). Another misconception would be to treat corner-centered methods as necessarily dependent on post hoc optimization. CFL emphasizes that no additional optimization or model fitting is required after corner prediction (Fernandez-Labrador et al., 2019), whereas SSLayout360 relies on deterministic geometric post-processing after prediction rather than an external search procedure (Tran, 2021). Conversely, in the FCC literature, corner-centeredness is not about sparse sensing or prediction at all but about geometric placement of algebraic relations (Kels, 2021).

Taken together, these works define corner-centered layout as a cross-disciplinary pattern: the global object of interest is organized around privileged corner-like entities, and the rest of the structure is recovered, constrained, or controlled relative to them.