Can the counter
keep up?
Test whether events survive the journey from physical occurrence to available frames to recorded annotation.
This independent instrument combines synthetic calibration, manual video-event review, and imported mask/identity/occlusion evaluation. It does not automatically segment subjects, run a video model, or establish production annotation accuracy.
EVENT
FRAME
MARKER
Where did the events go?
Three measurement layers
The simulator assumes perfectly accurate event definitions and a perfect reviewer. Markers can disappear because the synthetic visible window never contains a sampled frame or because the simulated interface rejects a second timestamp.
Compare two configurations
Save the current configuration as baseline A. Change inspection FPS or marker spacing to compare B against exactly the same known synthetic events.
Who did what, when?
Load authorized video, select the target, and review any repeatable action by its observable key point and reset. Count decisions stay separate from evidence.
Local observation
Frame-step buttons seek by 1/30 second as a navigation convenience; that is not a verified source-frame interval. Use the video's actual visible evidence. The M key records general marks only. In action-cycle mode use the evidence buttons; replaying the clip does not add events by itself.
What counts as one repetition?
Choose an example or define your own. The same rules work whether someone moves forward, backward, or stays in place: identify the action, its reset, and the exact new key point.
Pause at the source-video time and record what you observe. **Ready/reset** arms the next repetition. A key point counts only if both reset and key point were directly observed for this subject, part, and action definition.
Counts are decisions from manual observations, not verified ground truth. The export keeps each action definition separate; switching presets does not mix old and new events. Replay or a sustained pose never adds another count.
Does the target survive the obstruction?
Import per-frame model masks and compare them with independently reviewed subject outlines and visibility labels. Tracking IDs alone never establish that the same person or animal was recovered.
Mask shapes are displayed only when the imported file names the same local video and matches its dimensions and duration. The demo is synthetic and never overlays an unrelated recording.
To label a visible frame, first use Draw / correct polygon at that timestamp. Hidden/out-of-view labels do not invent an invisible mask. Labels are saved locally and reported separately from model predictions.
All metrics are conditioned on reviewed reference masks. IoU is approximate; identity-switch warnings are suspects requiring confirmation. Reappearance candidates are not automatically verified. No training or automatic segmentation happens here.
Accepted markers and rejected attempts
Manual marks are reviewer observations—not independent ground truth. Source video is never embedded in the export. Polygon anchors are manual, while any between-frame polygon is only a rigidly translated outline using provisional rectangle motion—not semantic segmentation or person reidentification.