forked from longtao.wu/droboticflow
1.6 KiB
1.6 KiB
Continuous annotation boundaries
The Fanyu Capture workbench uses hardware-frame-v1 for newly confirmed slices.
An ordered boundary grid maps one integer hardware timestamp to both camera frame
indices and the CFR left-video media time. Interior edges must advance both camera
indices. The common-range endpoints are included; an interval is half-open.
- Slices partition the common range. Descriptions partition each slice.
- New descriptions begin at the selected unannotated gap's start. The operator locates and explicitly confirms its end before saving the description.
- Adjacent descriptions share their boundary. Editing it updates both sides.
- Empty regions are not automatically assigned meaning. Explicit invalid-region marking is available; final submission rejects uncovered regions and overlaps.
- Paused media seek notifications do not overwrite the logical playhead.
- New materialized JSON preserves hardware boundary IDs. Packaging must not recalculate them from clip-relative milliseconds. Acceptance uses the exported media boundary grid for the same preview mapping.
- Legacy off-grid data is not silently rewritten. The annotator must explicitly confirm a reset/re-slice; its existing warning states that annotations are cleared. Already accepted legacy packages retain their previous export contract.
Checks: annotation_boundary_grid_test.go, human_slice_materialization_test.go,
annotation_draft_test.go, annotation_submission_conflict_test.go, frontend
unified-annotation-boundaries.test.mjs and annotation-playhead-draft.test.mjs,
and backend test_delivery_pose_contracts.py.