COLMAP-Based 3D Consistency Metrics

• COLMAP-based metrics are 3D-consistency evaluation measures that leverage classical geometric verification from SfM and MVS to assess coherent scene reconstruction.
• They integrate signals such as registration rate, dense photometric–geometric consistency (GPC), and coverage-weighted metrics (W-GPC) to identify failures and ensure geometric reliability.
• Empirical comparisons show these metrics correlate strongly with human judgments and outperform learned-backbone methods under noisy or pathological input conditions.

COLMAP-based metrics are a suite of 3D-consistency evaluation metrics that leverage classical geometric verification signals from the COLMAP structure-from-motion (SfM) and multi-view stereo (MVS) pipeline. These metrics address a key failure mode in multiview 3D evaluation: many learned-backbone or reference-free consistency metrics are insensitive to inconsistencies such as cross-scene mixtures, repeated/corrupted frames, or “hallucinated” geometry, yielding high scores even for inputs incompatible with a single static 3D scene. COLMAP-based metrics instead explicitly use matching robustness, registration, dense reconstruction quality, coverage, and outright reconstruction failure as interpretable consistency signals. They were introduced and analyzed in "Can These Views Be One Scene? Evaluating Multiview 3D Consistency when 3D Foundation Models Hallucinate" (Paul et al., 18 May 2026), where they are shown to correlate strongly with human assessments and to outperform learned baselines, especially under pathological or noisy inputs.

1. Sparse Feature Matches and Registration Success

COLMAP’s feature detector and exhaustive pairwise SIFT matching, followed by RANSAC-based geometric verification, allow direct assessment of global 3D consistency across a set of views. The central signal is the registration rate, defined as

$\mathrm{registration\_rate} = \frac{|R|}{|A|}$

where $A$ is the set of all attempted images and $R \subseteq A$ is the set successfully registered by SfM. A high registration rate indicates that most images could be placed into a globally consistent camera geometry, whereas a low rate reflects failure of geometric consistency—typically due to conflicting or outlier images, severe noise, or non-rigid content.

Optionally, match ratio per view pair—$$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$, where $M^{\mathrm{inlier}}_{ij}$ is the number of RANSAC-filtered feature correspondences—can be aggregated for a nuanced measure of pairwise geometric support.

A low registration rate (e.g., below 30%) reliably signals that the input set does not support a rigid multiview interpretation, and is thus flagged as 3D-inconsistent.

2. Dense Photometric–Geometric Consistency (GPC)

COLMAP’s MVS module outputs, for each registered view $v$, two distinct depth estimates: the photometric map $D_p^{(v)}$ and the geometric-consistency map $D_g^{(v)}$. The per-pixel quality for valid pixels $u \in \Omega_v$ (where both depths are finite and $D_g^{(v)}(u) > 10^{-5}$) is given by: $A$0 with $A$1.

For each view:

• $A$2
• $A$3
• $A$4

Scene-level GPC is the mean over all surviving dense outputs: $A$5

High GPC scores indicate agreement between independent MVS depth cues and dense multiview support, while low GPC signals inconsistent stereo evidence or geometric hallucination.

3. Integrated Consistency Mass (ICM)

ICM measures the total “verified” 3D mass in pixel–view space, with two denominator options:

• Survivor ICM: Normalized by area of successfully densified (D) images.
• All-attempted ICM ($A$6): Normalized by area of all attempted images (A), incorporating penalization for failed views.

$A$7

$A$8

$A$9 is especially failure-aware: if a large fraction of images are unregistered or undensified, the numerator collapses toward zero while the denominator remains constant, driving the metric to zero for inconsistent scenes.

4. Angular Coverage and Coverage-weighted GPC (W-GPC)

Geometric registration alone does not guarantee that the recovered cameras span the scene broadly. Angular coverage is measured by the largest circular gap between the azimuth angles $R \subseteq A$0 of the registered camera centers $R \subseteq A$1, relative to their median $R \subseteq A$2, in the horizontal plane. Specifically,

$R \subseteq A$3

where $R \subseteq A$4 are consecutive azimuthal gaps among sorted $R \subseteq A$5.

Coverage is used to weight GPC: $R \subseteq A$6

W-GPC downweights reconstructions that register dense geometry only over a limited viewpoint sector, ensuring penalization of partial or degenerate scene support.

5. Explicit Use of Reconstruction Failure

Unlike learned-backbone metrics (e.g., MEt3R, VGGT, DUSt3R), COLMAP’s pipeline fails explicitly on outlier, noisy, or inconsistent inputs, yielding no registration or no dense depth. These failures are incorporated both into $R \subseteq A$7 and $R \subseteq A$8. For images that fail SfM or MVS, their contribution to the ICM or GPC numerators is zero, but they remain in the denominator of $R \subseteq A$9. Thus, a high fraction of failure events reliably flags the whole input set as 3D-inconsistent. This offers superior robustness to pathological cases where learned methods may hallucinate plausible geometry.

6. Empirical Comparison and Application Guidelines

Scene-level rankings with COLMAP-based metrics, especially W-GPC, align with human judgments of consistency substantially better than leading neural metrics. For example, on DL3DV and MipNeRF360 datasets, Spearman $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$0 between W-GPC and human rankings reaches $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$1 (cf. MEt3R: $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$2 to $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$3; PRISM-MMD: $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$4, $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$5). On controlled benchmarks (SysCON3D), W-GPC achieves perfect separation and severity ordering ($$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$6), whereas learned metrics often rank synthetic noise or cross-scene mixtures as highly consistent.

Recommended usage:

• Preprocess with COLMAP to obtain features, SfM registrations, MVS depths.
• Compute and report: $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$7, primary ($$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$8), and diagnostics (e.g., $$\mathrm{match\_ratio}_{ij} = M^{\mathrm{inlier}}_{ij}/N^{\mathrm{raw}}_{ij}$$9).
• Low $M^{\mathrm{inlier}}_{ij}$0 or low registration rate denotes strong evidence against single-scene consistency.
• If COLMAP fails entirely, interpret consistency as zero; for potentially clean inputs, fallback to neural metrics remains possible.

By leveraging explicit geometric verification, dense stereo agreement, and failure detection, COLMAP-based metrics deliver interpretable, robust signals as to whether a set of views supports a coherent, static 3D scene, particularly outperforming neural consistency metrics under challenging or contaminated inputs (Paul et al., 18 May 2026).