Detection demo — guide

Ask in your own words or tap a sample question. The assistant runs the geocanvas-detect CLI on the server over the high-res ACT Urban imagery and draws the result as a new layer. A run takes ~20 s; the spinner shows progress and the controls lock until it settles.

Options

• Model — each model carries its own class table and box style (oriented or axis-aligned); switching models switches both.
• Confidence — minimum score kept. Higher = fewer, more certain detections.
• NMS IoU — how aggressively overlapping duplicates are merged. Lower merges more; raise it to keep tightly-packed objects (e.g. parked cars) separate.
• Levels — z21 alone is fastest; adding z20 also scans a coarser zoom to catch larger objects, at roughly double the time.

Layers

Every run adds a layer on the left. Use 👁 to hide it, Edit to rename, recolour, or change its opacity, and ✕ to delete it. Runs stack, so you can compare models side by side.

Tips

• Zoom in first — smaller areas run faster (the request is clamped to a block at the view centre).
• Live detection needs you to be signed in; otherwise a bundled precomputed sample is shown and labelled as such.
• The server runs at most two detections at once.