VixitAi Labs ← SPARSE capability note
Live engine · runs on 2 CPU cores · no GPU

Explore the engine

Generate a labelled scene or feed a real neuromorphic recording, and the same pipeline that produces every number in our results log will run it — filtering, clustering and tracking in the event domain, never building a frame. You get accuracy against ground truth, the full latency distribution, and an offline viewer of the run.

Input

Pick a source. Synthetic scenes come with ground truth, so they score; a recording only scores if it carries labels.

The detection limit sits between magnitude 8 and 10 at 640×480. Try 6, 7.5, 9 for the canonical scene, or 11.5 to watch it fail honestly.

One run at a time, so the latency figures stay honest on a shared machine. Synthetic scenes are capped at 5 s and 640×480; recordings are capped at 6M events.

Run something and the numbers appear here.
Read the numbers as what they are. Synthetic scenes are generated by our own model, so they measure the pipeline against our assumptions, not against a sensor. Real recordings are the honest test, and on those the accuracy is weak — the capability note explains exactly where and why.