Demo Pipeline

This page describes how to build a self-contained proof-of-concept demonstration using open datasets and the Inspect CLI. It is intended for use in technical evaluations and field demos before production data is available.

Open datasets

Verify any IFC download URL before use. The buildingSMART gallery is the authoritative source; third-party mirrors may serve modified files.

Pipeline steps

Step 1 — Generate design point cloud from IFC

Use IfcOpenShell to sample the IFC surface geometry into a reference point cloud (the “ground truth” design). Each IfcProduct element is triangulated and its vertices collected into a flat (N, 3) array representing the design surface.

Step 2 — Simulate a damaged scan

Use Open3D to introduce controlled deformations into a copy of the design cloud, producing a simulated “as-built” scan with known defects. A representative demo deforms a localised region by 15 mm to simulate a surface depression, then saves the result as a PLY file.

Step 3 — Compute deviation (M2)

Run the eds inspect scan CLI command, pointing it at the IFC design file and the simulated scan PLY. The CLI calls src/ifc.rs to load the design reference cloud, then src/deviation.rs to compute per-point nearest-neighbour deviation and emit a JSON report containing compliant_pct, max_deviation_mm, and mean_deviation_mm.

This step exercises src/ifc.rs and src/deviation.rs (M2).

Step 4 — Project 3D → 2D (trilink-core)

trilink-core::project_to_depth_map converts the scan point cloud into a depth map image for AI inference input. This is handled automatically by the CLI using the camera intrinsics in config.toml — no manual step is required.

This step exercises trilink-core::project_to_depth_map (foundation #31).

Step 5 — AI defect detection

A detection model runs over the depth map via the HTTP inference path (inference.mode = "http"). For demos, YOLOv8 can be used as the external inference server. The CLI sends the depth map image to the configured HTTP endpoint and receives bounding-box detections in return (M4).

Step 6 — Back-project 2D → 3D

Detected 2D bounding boxes are back-projected to world coordinates using trilink-core::unproject, then overlaid on the 3D model and included in the deviation report (M4).

Deviation engine in the demo

The deviation engine (M2) is the quantitative centrepiece of the demo. It answers the question “by how many millimetres does the as-built structure deviate from the IFC design?” — not just “is there an anomaly?”. Make sure Step 3 is demonstrated explicitly, as it differentiates this pipeline from a generic defect detector.

Tech stack summary