Skip to content

milestone6_dataset_pipeline.md

Latest commit

 

History

History
84 lines (61 loc) · 2.6 KB

milestone6_dataset_pipeline.md

File metadata and controls

84 lines (61 loc) · 2.6 KB
Raw

Milestone 6: Dataset Pipeline and Offline Analysis Platform

Scope

Milestone 6 focuses on building the data loop, not improving detector logic:

  • Keep GunshotEventDetector unchanged
  • Build event-centric dataset capture pipeline
  • Export audio + features + metadata for offline analysis

Pipeline

AudioCapture -> STFTProcessor -> FeatureExtractor -> GunshotEventDetector -> dataset_recorder

Recorder stages:

  1. Capture each detector peak and attach its final accepted/rejected decision
  2. Wait for full 200 ms post context
  3. Pull 400 ms PCM window from AudioHistoryBuffer (200 ms pre + 200 ms post)
  4. Pull matching feature rows from FeatureHistoryBuffer
  5. Persist asynchronously into a private staging directory as:
    • audio.wav
    • features.csv
    • metadata.json
  6. Atomically publish the complete event directory

Recorder sessions use collision-resistant event IDs and never overwrite an existing event. Failed writes are removed from the staging area instead of leaving partial dataset entries.

New core components

AudioHistoryBuffer

  • 3-second rolling stereo PCM cache
  • Supports absolute-sample window extraction
  • Supports time-centered extraction (center ± window)

FeatureHistoryBuffer

  • 3-second rolling feature cache
  • Stores per-frame AudioFeatures + detector score/confidence
  • Supports time-range extraction for CSV export

Dataset layout

Recorder creates:

dataset/gunshot, dataset/footstep, dataset/reload, dataset/switch, dataset/ambient, dataset/unknown

Current auto-save target:

dataset/unknown/<event_id>/audio.wav|features.csv|metadata.json

Each physical detector candidate produces one clip. metadata.json records the final detector decision, recorder session, wall-clock timestamp, and human review status. Manifest export covers all six labels, not only unknown.

Recorder UI

  • Live statistics:
    • Total Events
    • Saved Events
    • Discarded Events
    • Disk Usage
    • Recording Time
  • Recent events (last 10)
  • Waveform viewer
  • Spectrogram viewer
  • Replay selected event
  • Manual ambient snapshot
  • Manifest export (dataset/manifest.csv)

Dataset Studio review flow

  • Unreviewed clips are shown by default
  • Keys 1 through 6 label/review the current selection
  • Auto-advance can select and optionally play the next clip
  • Review state is independent from the label, so a clip may remain unknown while still being marked as reviewed
  • Manifest export, soft delete, restore, and undo route through DatasetManager
  • Recorder and Dataset Studio both accept --dataset-root <path> so capture and review can share a non-default dataset location