Camera-based attention tracking is currently verified only on desktop Chrome.

FocusFlow · Stage 1 diagnostics

Local webcam gaze spike. No video recording, uploads, or saved calibration. Gaze estimates visual engagement, not mental attention.

History

Camera: idle · Permission: not requested · Model: not loaded

Browser tab visibility / window focus: unknown

“Visible” and “focused” describe the browser APIs, not your face or gaze. Looking away does not blur the window. Gaze classification appears separately below.

Use good lighting, remove glare, and keep posture steady. Calibration covers the viewport. Recalibrate after resizing, moving the camera, or changing posture.

Live readout

Focus level (high/medium earn credit)
unknown
Gaze classification (not browser focus)
unknown
Classification reason
Camera inactive; no inference
Face detected (MediaPipe)
no current inference
Last completed inference (monotonic ms)
unavailable
Raw regression (x, y), unclamped
unavailable
Smoothed median (x, y)
unavailable
Smoothing samples
0 / 5
Active bounds (both axes, inclusive)
-0.080 to 1.080
Base margin / retention hysteresis
0.08 / 0.05
Previous class used for bounds
unknown
Quality gate reasons
Missing or unusable eye landmarks / pose
Minimum eye aperture (gate <0.08)
unavailable
Feature RMS / spread rejection thresholds
>3 / >1.6
Live features: Lx, Ly, Rx, Ry, yaw, pitch
unavailable
Tracking quality
unavailable
Inference FPS (5s)
0.0
Median latency
unavailable ms
Median interval
unavailable ms
Median / max completion gap
unavailable / unavailable ms
Median / max decoded-frame interval
unavailable / unavailable ms
Median newly decoded frames between inferences
unavailable
Median scheduled-loop interval
unavailable ms
Maximum gap (5s)
0.0 ms
Gaps >500 ms (5s)
0
Calibration
not calibrated
Valid samples
unavailable
Held-out median error
unavailable
Feature distance / stability
unavailable / unavailable

Error is Euclidean distance in normalized viewport coordinates. Validation holds out each entire target group; nine targets provide limited evidence. Poor calibration disables definite gaze classification. FPS counts completed real inference calls, including calls with no detected face. The rolling display uses a full five-second denominator.

Test center → left/right/top/bottom → beyond each edge, holding each for two seconds. Compare live features, raw and median coordinates. If raw moves but gaze is unknown, inspect quality reasons. If raw stays centered, inspect fitted target medians and training versus held-out error. Exit from on-screen requires a median beyond −0.13 or 1.13; re-entry uses −0.08 to 1.08. High and medium are both focused; medium includes screen edges. Large or unreliable gaze changes can still trigger quality gates; unknown is uncertainty, not focused.

Background feasibility experiment

Start the camera, choose a scenario, then start a run. Spend at least 30 seconds in the target condition before returning to stop. Export includes inference timestamps, latency, visibility/focus events, and interruptions. It excludes facial features and calibration.

No experiment recording

Document PiP: unsupported. This probe does not guarantee background reliability.

  1. A: keep this tab visible for at least 30 seconds.
  2. B: switch to another Chrome tab for at least 30 seconds, then return and stop.
  3. C: switch to another desktop app for at least 30 seconds, then return and stop.
  4. D: open Document PiP first, return focus here to start the D run, then switch tabs/apps with PiP visible for at least 30 seconds. Return and stop. Record unsupported if unavailable.
Measured completed runs — no results until recorded
ScenarioDurationFPSMedian intervalMax gapGaps >500msMedian latency
Engineering thresholds and limitations

Both eyes required; aperture/width 0.08–0.65. Head yaw limit 1 radian, pitch 0.9. Calibration: 800 ms settling, 24 samples per target, 16 minimum after a 3.5-MAD filter (scale floor 0.015). Ridge λ=0.1, unpenalized intercept, training-only standardization. Poor: median held-out error >0.20 or any target median >0.35.

Unavailable: unusable landmarks. Low: uncalibrated/poor fit, standardized feature RMS >3, five-prediction spread >1.6, or aperture <0.08. High requires error ≤0.12, feature RMS ≤1.5, spread ≤0.08, aperture ≥0.15; other valid tracking is medium. Low/unavailable gives unknown immediately. Stale inference over 500 ms clears classification when execution resumes. Screen margin 8%; on-screen retention adds 5% hysteresis. Quality-distance scale floors: iris X 0.025, iris Y 0.020 eye-width units per eye; yaw 0.10 rad, pitch 0.08 rad, centered on calibration means. Regression training scales remain unchanged. High uses the calibrated inner region [0.1, 0.9] on each axis, retained to [0.05, 0.95] with high quality. Other accepted on-screen estimates are medium; both earn credit. Smoothing requires five valid predictions. These are engineering assumptions, not validated confidence scores.

Synchronous CPU VIDEO inference is capped at 15 calls/second, using unique decoded frames at 640×480 ideal camera size. A worker needs latency evidence and frame-transfer testing; this spike uses a measurable main-thread pipeline. Suspension can prevent UI updates until resume. Document PiP leaves inference in the source page, so results apply only to this arrangement.