Menu
- 1Upload
- 2Verify metadata
- 3Annotate
- 4Export
- 5Next video
FREE TOOL / DATA WORKBENCH
Turn video into temporal data.
Convert manipulation video into human-verified temporal data.Source: Human annotated
Frame index: Derived from timestamp × FPS
01 / VIDEO INPUT
Choose a manipulation video.
The selected file stays in this browser session. It is not automatically added to the nAIve physics commercial dataset.
02 / AUTO CAPTURE
Source file
03 / BEHAVIOR TRACK
Timeline
No gaps or overlaps detected.
04 / HUMAN VERIFICATION
Annotation controls
Experimental interaction labels are being evaluated and are not part of the published v0.2 ontology.
- Start
- Not marked
- End
- Not marked
- Duration
- —
06 / SESSION RECORDS
Annotation records
Choose a video, mark a start and end point, then create the first human-verified behavior segment.
RGB ANALYSIS / v0.1
Wrinkle / crease response inspection
Deterministic RGB image-structure analysis. Derived from RGB · Experimental.INTERNAL / EXPERIMENTAL WORKBENCH
Auto analysis
Independent browser-side layers. Failure in one module does not block temporal annotation.Load an episode to begin.
INTERPRETABLE TEMPORAL PROPOSALS
Temporal review
Automation proposes intervals. Only an explicit human acceptance adds a verified annotation.Run Auto Analysis, seed the garment and iron where needed, then generate temporal candidates. Nothing is generated or accepted automatically.
SEMANTIC + TRACKING SEED
Source-pixel inspection
Capture a frame, choose an interaction, then click. MediaPipe loads only when analysis or garment segmentation is requested.
Mask: not seeded. Hands: 0 detections. Pose-support samples: 0. Tool: 0 tracked positions. Model proposals and tracked outputs remain reviewable estimates.
CONSOLIDATED REVIEW
Review required
No automatic review items yet. Run analysis to generate interpretable candidates.
EXPERIMENTAL OUTPUTS
Export only what exists
Canonical temporal CSV remains nine columns and excludes unaccepted candidates.MediaPipe Tasks Vision 1.0.1; Hand Landmarker, Pose Landmarker, and Interactive Segmenter models load remotely from Google-hosted assets. Model URLs are recorded in provenance: hand, pose, segmenter.
Thermal analysis is not available from RGB video. It requires thermal / IR sensor data.