AI & MLPythonJavascript

Testing Offline Depth Estimation in a Photo Walkthrough

A collection of room photographs can support a useful walkthrough without reconstructing a complete 3D scene. The house-photo project starts with navigation between viewpoints, then explores what additional vision processing might contribute.

The retained validation record includes an actual offline run of Depth Anything V2 Small. That gives this experiment a concrete model-inference result while keeping the viewer's basic operation separate from AI.

Start with navigation

The viewer organizes photographs through explicit viewpoints and connections. A person can move through a collection using those relationships without requiring a model to recover every camera position.

The processing tools also examine image metadata, duplicates, and candidate overlap. Those conventional steps help establish what the collection contains before attempting a more ambitious reconstruction.

Add depth as a separate experiment

The September 2026 record documents downloading and checking a Depth Anything V2 Small checkpoint, then running it with offline mode enabled on a sample entry image.

The output includes a grayscale depth map and provenance. It gives a visual estimate of scene depth; the recorded check does not establish metric reconstruction accuracy or the quality of a browser parallax effect.

That boundary matters. An attractive depth map is a useful intermediate artifact, but it is not the same as a correctly scaled model of a house.

Validation exposed practical problems

The retained notes describe an incomplete checkpoint transfer caught by integrity checking and a runtime configuration serialization problem corrected before the successful run.

Those are valuable parts of an offline-model workflow. A cached file needs to be the right complete checkpoint, and the recorded configuration needs to explain what actually ran.

The project also contains an optional semantic-feature path, but it was not exercised with a semantic checkpoint in that validation. Camera recovery, splat reconstruction, and generated scene completion remain separate ideas rather than demonstrated results.

Keep private imagery out of the demonstration

Later viewer work includes a real-house collection. Its navigation and file checks should not be confused with the earlier sample-image inference result.

For a public post, I would show only a permitted sample image, its derived map, and a synthetic navigation example. Addresses, interior photos, floor plans, and location metadata should stay private.

The useful lesson is incremental: build a useful viewer, retain source provenance, then evaluate one model capability at a time instead of labeling the whole workflow “AI reconstruction.”

Background