Load a video, step through it frame-accurately, and compute its median background image.
You have unsaved work on - from before the page reloaded.
Drag a video here, or
The maze is static and the mouse moves, so the per-pixel median across evenly-sampled frames gives a clean empty-platform image.
Hough circles on the background: one big circle for the platform, ~20 small ones on its rim for the holes, indexed in ring order by angle.
Drag a circle's edge to resize it, drag its middle to move it. Click a hole (no drag) to mark it as the target - or use the dropdown below if clicking precisely is awkward on a small screen.
Background subtraction against the median, gated to the confirmed platform circle and a plausible size range - largest matching blob wins each frame. A frame with no valid blob is marked failed, never guessed.
Scrub the frame viewer above - green crosshair = tracked, red X = failed frame.
Per-hole pixel-change signal, duration-gated. An entry is the signal plus the mouse's own tracked blob going invalid right after; a poke is the signal plus the mouse's speed dropping to at-or-below its own trial's median. Low-confidence candidates are shown, not hidden - disagreement gets flagged, not silently resolved.
Human review is a draft control: it works this session but does not yet survive a reload or a re-detect with new thresholds - full correction persistence is its own later piece, not built here on purpose. See KNOWN_LIMITATIONS.md.
| Hole | Frames | Type | Confidence | Avg speed | Human review (draft) |
|---|