PGCM: Cross-View Geometric Consistency

• The paper introduces PGCM as a metric that quantifies cross-view geometric consistency by measuring inverse-depth reprojection errors between overlapping views.
• It utilizes dense warping, a perspective camera model, and discrete occlusion masks to accurately evaluate geometric agreement in static scenes.
• By decoupling single-view accuracy from multi-view validation, PGCM offers practical insights for improving multi-view stereo, video, and satellite imagery methods.

Perspective-Geometry Consistency Metric (PGCM) can be understood as a cross-view geometric consistency criterion that evaluates whether predictions from overlapping viewpoints describe the same scene after projection into a common camera frame. In the formulation most directly associated with the term, the consistency quantity is a cross-view reprojection error in inverse-depth, evaluated in a common frame and masked by visibility. Although the 2019 mesh-correction paper did not name this quantity a “Perspective-Geometry Consistency Metric,” it provided nearly all of the ingredients for one through its geometric consistency loss (Săftescu et al., 2019). Later work preserved the same basic concern—agreement under viewpoint change—while instantiating it through forward-backward reprojection, RGB-depth verification on perspective graphs, ego-centric cross-view consistency, pose-divergence diagnostics for video, failure-aware multiview verification, and RPC-constrained satellite evaluation (Vats et al., 2023, Cheng et al., 13 Oct 2025, Yuan et al., 15 Jun 2026, Dou et al., 19 Mar 2026, Paul et al., 18 May 2026, Luo et al., 16 Jun 2026).

1. Conceptual basis

The core premise is static-scene coherence across overlapping views. The foundational statement is explicit:

“Intuitively, since the reconstructions we wish to correct are static, the predictions made from overlapping views should be geometrically consistent. In other words, surfaces that appear in a certain location according to a prediction should appear in the same location in all predictions where they are in view.” (Săftescu et al., 2019)

In this view, perspective inconsistency is not merely a local depth error. It is a cross-view failure: a surface position inferred from one viewpoint does not agree with what a nearby viewpoint predicts for the same 3D surface once both are expressed in the same camera frame. The 2019 formulation ties this failure mode directly to overly smooth corrections, stating that “smooth predictions result in geometrical inconsistencies” and addressing it with “a loss function which penalises re-projection differences that are not due to occlusions” (Săftescu et al., 2019).

This suggests a precise definition of PGCM: a metric should measure the discrepancy between two viewpoint-conditioned geometric predictions only where those predictions are jointly visible and therefore capable of agreement. In the original formulation, the relevant signal is not RGB photometric error and not generic depth error, but agreement of corrected inverse-depth across nearby views after perspective reprojection into a common camera frame (Săftescu et al., 2019).

2. Canonical inverse-depth formulation

The original construction operates on inverse-depth corrections. The ground-truth correction is

$\Delta^{*}(p) = d^{*}(p) - d(p),$

where $p$ is a pixel index, $d^{*}$ is the inverse-depth image from the high-quality reconstruction, and $d$ is the inverse-depth image from the low-quality reconstruction (Săftescu et al., 2019). The network predicts $\Delta$, and the corrected inverse-depth is formed by adding the predicted correction to the low-quality inverse-depth.

The key consistency term is the geometric consistency loss:

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$

Here $t$ is the target view, $n \in N$ is a nearby view, $d_t$ and $d_n$ are corrected inverse-depth images, $p$0 is the target-view prediction expressed in the frame of view $p$1, $p$2 is the nearby-view inverse-depth warped to correspond to target-view geometry and also expressed in view $p$3’s frame, and $p$4 is the set of pixels unoccluded in nearby view $p$5 (Săftescu et al., 2019).

In the 2019 paper this quantity is an auxiliary training loss, weighted in the full objective by $p$6, rather than a standalone benchmark metric (Săftescu et al., 2019). A paper-faithful metric interpretation was stated explicitly in the supplied synthesis: a metric version would simply remove dependence on training and report the normalized masked average reprojection inconsistency,

$p$7

Lower values indicate better perspective-geometry consistency (Săftescu et al., 2019).

The significance of this formulation is that it separates single-view accuracy from multi-view agreement. A corrected depth map may appear locally plausible, yet still be inconsistent under perspective reprojection. PGCM, in this sense, measures whether the geometry itself survives viewpoint transfer.

3. Projection model, warping, and visibility

The consistency mechanism is built on a perspective camera model and inverse-depth parameterization. The 2019 paper motivates inverse-depth for three reasons: it emphasizes surfaces closer to the camera, background or non-surface pixels can be assigned zero corresponding to infinite depth, and resampling inverse-depth is simpler than resampling depth (Săftescu et al., 2019). The last point is directly relevant to PGCM because the consistency term depends on dense warping.

For target view $p$8 and nearby view $p$9, let $d^{*}$0 and $d^{*}$1 be CNN predictions, and define corrected inverse-depth images

$d^{*}$2

With target pixel coordinates $d^{*}$3, intrinsic matrix $d^{*}$4, and $d^{*}$5 transform $d^{*}$6, the paper defines

$d^{*}$7

$d^{*}$8

$d^{*}$9

$d$0

$d$1

The rationale is explicit: “Since we are warping inverse-depth images, where the pixel values depend on the viewpoint, we need to compute the absolute difference in the same camera frame” (Săftescu et al., 2019). PGCM therefore compares $d$2 and $d$3, not quantities that remain expressed in different viewpoint-dependent parameterizations.

Visibility is equally central. The 2019 method states that “views cannot be consistent in the presence of occlusions” and therefore computes occlusion masks from the reference high-quality reconstruction (Săftescu et al., 2019). The mask construction is mesh-based. Each mesh triangle is assigned an index by hashing its world-frame coordinates, triangle-index images are rendered, and for each pair of views the target-view triangle ID is reprojected into the nearby view. Instead of interpolating triangle IDs, the four nearest samples are returned separately; if the target ID matches at least one of those four samples, the pixel is considered unoccluded, otherwise occluded (Săftescu et al., 2019).

Two clarifications are methodologically important. First, the paper does not use a depth threshold, z-buffer difference threshold, or learned occlusion confidence; the occlusion test is discrete triangle identity matching (Săftescu et al., 2019). Second, the visibility set $d$4 has “no relation to the set of valid pixels ($d$5) from the previous losses,” because the consistency term is computed between predictions and can therefore supervise regions with no valid label (Săftescu et al., 2019).

Dense warping uses linear interpolation of inverse-depth under the “mild assumption that surfaces between pixels are planar,” which the paper describes as another advantage of the inverse-depth formulation (Săftescu et al., 2019). This gives PGCM a differentiable resampling interpretation even though the original paper presents it as a loss rather than a benchmark score.

4. Relationship to edge preservation and standard evaluation

The 2019 system does not rely on geometric consistency alone. It couples $d$6 to a data term, a gradient term, and regularization:

$d$7

with $d$8, $d$9, $\Delta$0, and $\Delta$1 (Săftescu et al., 2019).

The data loss uses the berHu norm, and the gradient loss preserves local structure and sharp transitions. The system also introduces edge-aware weighting: $\Delta$2

$\Delta$3

with $\Delta$4 and $\Delta$5 (Săftescu et al., 2019). Pixels near edges receive larger weights.

The relationship between these terms and PGCM is explicit but indirect. The paper treats edge preservation and geometric consistency as distinct mechanisms, yet states that excessive smoothing causes geometrical inconsistencies (Săftescu et al., 2019). Thus edge-aware weighting does not appear inside $\Delta$6, but it improves the corrected inverse-depth fields supplied to the consistency term. A plausible implication is that a PGCM computed on oversmoothed predictions will often increase because occluding boundaries and sharp geometric transitions shift differently across viewpoints.

The paper also reports standard inverse-depth evaluation measures that are not PGCM measures. These include thresholded accuracy

$\Delta$7

with $\Delta$8, as well as

$\Delta$9

(Săftescu et al., 2019). These are per-view prediction quality metrics rather than cross-view consistency metrics.

This distinction motivates a common correction in interpretation: PGCM is not equivalent to gross-error correction, iMAE, iRMSE, PSNR, or any other single-view fidelity statistic. In the original study, the model “reduces gross errors by 45.3\%–77.5\%, up to five times more than previous work,” and adding geometric consistency improves performance “at all error scales,” but those gains remain indirect evidence for PGCM rather than the metric itself (Săftescu et al., 2019).

5. Generalizations across later research

Later work preserved the central PGCM intuition—agreement under viewpoint change—but instantiated it differently depending on geometry model, overlap regime, and task.

In multi-view stereo, GC-MVSNet defines consistency by forward-backward reprojection agreement between a predicted reference-view depth map and multiple source-view ground-truth depth maps. Its explicit per-source quantities are the Pixel Displacement Error

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$0

and the Relative Depth Difference

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$1

followed by thresholded inconsistency masks and multi-view aggregation through a normalized inconsistency frequency (Vats et al., 2023). This formulation makes PGCM more explicitly cycle-consistent and multi-scale than the 2019 inverse-depth loss.

In 3D Gaussian inpainting, PAInpainter replaces direct reprojection residuals with a dual-feature verification mechanism. Neighbors are selected from a perspective graph built from LoFTR match confidence, anchor content is projected into adjacent views using depth and camera transforms, and candidate inpaintings are scored by

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$2

with acceptance threshold $$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$3 in iterative refinement (Cheng et al., 13 Oct 2025). This does not define a standalone PGCM by name, but it operationalizes local perspective-geometry consistency as graph-conditioned RGB-depth agreement.

For low-overlap surround-view driving, SurroundNEXO shifts the problem from dense cross-view correspondence to ego-centric geometric comparability. It introduces Ego-Ray Positional Encoding, in which image patches are mapped to ego-frame viewing directions,

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$4

and evaluates cross-view depth consistency by calibrated reprojection only inside valid overlap (Yuan et al., 15 Jun 2026). This suggests that in weak-overlap regimes a PGCM should not rely on dense correspondence density alone.

Video-oriented variants make the same notion temporal. SGC measures 3D spatial geometric consistency by separating static from dynamic regions, estimating local camera poses for depth-clustered static subregions, and quantifying divergence through local and global rotational and translational variances plus warped depth consistency,

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$5

(Dou et al., 19 Mar 2026). PDI-Bench instead defines an object-centric Perspective Distortion Index,

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$6

with default weights $$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$7, thereby turning perspective scale law, 3D trajectory smoothness, and rigidity preservation into an explicit video PGCM (Wu et al., 14 May 2026).

Ground-truth-free multiview verification introduces a different emphasis: failure-aware consistency. The COLMAP-based family in “Can These Views Be One Scene?” defines per-pixel bounded quality

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$8

and scene-level scores such as geometric–photometric consistency and coverage-weighted GPC (Paul et al., 18 May 2026). Here the crucial principle is that inability to verify geometry should lower the score rather than vanish as missing data.

Satellite multi-view evaluation under RPC geometry extends PGCM beyond pinhole cameras. The proposed protocol defines RPC projected 3D feature consistency through

$$L_{gc} = \sum_{n \in N}\sum_{p \in U_n} \left|d_{n,t} - \hat d_{n,t}\right|.$$9

and a geometry-constrained dense matching proxy using EPE and PCK@10 inside an epipolar band of $t$0 pixels around the curved, height-dependent RPC correspondence manifold (Luo et al., 16 Jun 2026). Its central finding is that semantic agreement and geometric localization decouple: high similarity at the correct 3D point does not guarantee sharp, usable matchability.

Adjacent work broadens the conceptual perimeter of PGCM rather than its exact formula. A diffusion-model study introduces a vanishing-point-based perspective loss over edge profiles, effectively a single-image perspective-structure consistency measure (Upadhyay et al., 2023). MMPerspective, by contrast, operationalizes perspective consistency in multimodal reasoning as correctness preserved across perspective-preserving transformations such as cropping, masking, flipping, and rotation (Tang et al., 26 May 2025). These are not cross-view reprojection metrics, but they indicate that PGCM-style reasoning can be instantiated as either geometric verification or invariance under geometry-preserving perturbation.

6. Limitations, misconceptions, and methodological boundaries

A first misconception is historical. In the 2019 mesh-correction work, the consistency term is clearly a training loss, not a standalone evaluation metric (Săftescu et al., 2019). Treating the paper as having introduced a named benchmark metric would be inaccurate. A PGCM emerges by reinterpretation: the same masked reprojection discrepancy can be normalized and reported post hoc.

A second misconception is that any geometric-looking scalar is automatically a perspective-geometry metric. Standard per-view measures such as iMAE, iRMSE, thresholded accuracy, PSNR, SSIM, LPIPS, or FID do not, by themselves, quantify cross-view perspective agreement (Săftescu et al., 2019, Cheng et al., 13 Oct 2025). They may correlate with consistency, but they do not define it.

A third misconception concerns visibility. The 2019 formulation does not use thresholded depth comparison or learned occlusion confidence; it uses triangle-identity agreement from a reference mesh (Săftescu et al., 2019). Other settings likewise require task-specific validity logic: overlap-only evaluation in low-overlap surround-view depth (Yuan et al., 15 Jun 2026), static-region masking in generated video (Dou et al., 19 Mar 2026), or geometry-constrained search manifolds in RPC satellite imagery (Luo et al., 16 Jun 2026). A plausible implication is that no PGCM is meaningful without a clearly specified valid-support set.

A fourth boundary concerns localizability versus agreement. The RPC study shows that very high projected 3D feature consistency can coexist with poor localization under matching inference (Luo et al., 16 Jun 2026). Thus a descriptor can look geometrically consistent at the true 3D point while still failing to produce a sharp, unique response over the physically valid search manifold. PGCMs based only on similarity at correspondences may therefore overstate practical usefulness.

A fifth boundary concerns hallucinated geometry. The multiview robustness study demonstrates that learned reconstruction backbones can hallucinate dense geometry and cross-view support for unrelated scenes, repeated images, and random noise, while classical failure-aware verification can correlate far better with human judgments (Paul et al., 18 May 2026). This suggests that a PGCM built entirely on learned residuals may inherit the failure modes of its backbone unless it explicitly accounts for registration failure, missing dense support, or narrow viewpoint coverage.

Finally, perspective consistency is not identical across domains. Inverse-depth mesh correction, MVS forward-backward reprojection, perspective-graph inpainting, ego-frame driving, static-background video audits, panorama stitching, and RPC satellite matching all preserve the same intuition—consistency under geometry-aware transfer—but use different camera models, visibility assumptions, and aggregation rules (Săftescu et al., 2019, Vats et al., 2023, Zhu et al., 12 Mar 2026, Luo et al., 16 Jun 2026). The most defensible general statement is therefore narrow: PGCM denotes a family of metrics derived from the question of whether geometry predicted or inferred in one view remains compatible with geometry in other views once the relevant camera model, valid-support set, and aggregation rule are made explicit.