Matching Inspection Images to Reveal Infrastructure Change

Most infrastructure teams aren’t data-constrained. They’re engineering-time constrained.

Asset owners already have years of high-resolution inspection imagery. But the metrics engineers actually need—primary failure modes and rate of deterioration—still require significant manual effort to extract.

The hard part is not just detecting defects. Detection tells you what is visible in one image, but budgeting decisions depend on the delta: Did that crack grow 2% or 20% since the last inspection cycle? That makes this a spatiotemporal problem, not just a detection problem.

Our goal was to turn disconnected inspection photos into a structured, longitudinal asset record. To do that, we need to establish, with high confidence, that two pixels from two different images represent the exact same physical point on the asset.

Building a Correspondence Engine

At a high level, the engine uses transformer self-attention to extract spatial features from each image with broader structural context. It then uses cross-attention to compare each pair of images and propose dense matched points between them.

Instead of depending only on small local textures, this architecture uses wider spatial context to identify corresponding points. This allows it to match surfaces that may be smooth, repetitive, weathered, or partially degraded even when captured years apart, from different angles, distances, and lighting conditions.

Validation Before Acceptance

Because infrastructure imagery can create plausible but incorrect matches, we do not treat raw model output as truth. We treat it as a proposal. Before an image pair is accepted, its matches must pass through a calibrated validation funnel tuned against a Human-in-the-Loop ground truth dataset:

• Confidence filtering: Matches must meet thresholds tightly tuned against ground truth data.
• Statistical outlier rejection: Matches must agree with a consistent, image-wide geometry.
• Geometric coverage validation: Matches must be distributed across the image area, rather than clustered in one small region.

Why This Matters Operationally

The output is a validated match graph that transforms fragmented photos into a defensible longitudinal asset record. This changes the workflow:

• Engineers spend less time manually cross-referencing images.
• AI detections are connected across time instead of remaining isolated snapshots.
• Localized deterioration can be tracked more consistently across inspection cycles so engineers can better understand how it is progressing.