Signal Quality Scorer — SNR + consistency gate for training data #2

Open
opened 2026-04-26 21:43:01 -07:00 by pyr0ball · 0 comments
Owner

Deterministic scorer (MIT licensed) that evaluates each recording before it enters the training set.

SNR estimator: for landmark sequences, compute signal-to-noise ratio per landmark coordinate over the gesture window.

Consistency checker: inter-trial variance across the 10–20 repetitions for a gesture class. Flag high-variance recordings for re-record.

Artifact detector: blink contamination in face landmarks, sudden motion spikes, MediaPipe dropout frames (all-zero landmarks).

Confidence gate:

  • High confidence (SNR > threshold, variance low, no artifacts) → self-certify, add directly to training set
  • Low confidence → flag for re-record or manual review

Output: per-sample score dict {snr: float, consistent: bool, artifacts: list[str], certified: bool}.

Tests: synthetic clean landmark sequences score certified=True. Gaussian noise injection degrades SNR below threshold. Landmark dropout triggers artifact flag.

Deterministic scorer (MIT licensed) that evaluates each recording before it enters the training set. **SNR estimator:** for landmark sequences, compute signal-to-noise ratio per landmark coordinate over the gesture window. **Consistency checker:** inter-trial variance across the 10–20 repetitions for a gesture class. Flag high-variance recordings for re-record. **Artifact detector:** blink contamination in face landmarks, sudden motion spikes, MediaPipe dropout frames (all-zero landmarks). **Confidence gate:** - High confidence (SNR > threshold, variance low, no artifacts) → self-certify, add directly to training set - Low confidence → flag for re-record or manual review Output: per-sample score dict `{snr: float, consistent: bool, artifacts: list[str], certified: bool}`. Tests: synthetic clean landmark sequences score certified=True. Gaussian noise injection degrades SNR below threshold. Landmark dropout triggers artifact flag.
pyr0ball added this to the Phase A — Camera-Only milestone 2026-04-26 21:43:01 -07:00
Sign in to join this conversation.
No labels
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: Circuit-Forge/raven#2
No description provided.