| # Time semantics |
|
|
| The release preserves timestamps and offers optional record-time matching. It does not certify exposure synchronization or force all streams onto one frame grid. |
|
|
| ## Three distinct time quantities |
|
|
| 1. Header time is the original timestamp written by the acquisition software. It may use a device clock or a host clock, and meanings can vary across batches. |
| 2. Bag record time is the recorder's timestamp. A common per-episode origin is subtracted to obtain record_time_ns for all streams. |
| 3. Video PTS controls playback within a single MP4. It starts at that stream's first valid image, is quantized to microseconds, and is mapped back to source time by the frame table. |
|
|
| For each valid frame, the current encoder satisfies: |
|
|
| ```text |
| video_pts * 1000 |
| = record_time_ns - stream_start_record_time_ns |
| + video_pts_adjustment_ns |
| ``` |
|
|
| The formula assumes the stored time base of 1/1,000,000 seconds. Original nanoseconds are unchanged. If equal timestamps or quantization would produce non-increasing PTS, the encoder increases playback PTS by one microsecond and records the difference. The complete exported set has a maximum absolute adjustment of 500 ns. This number measures encoding time quantization, not capture-time accuracy. |
|
|
| ## Optional matching |
|
|
| `nearest_by_record_time` finds the nearest valid auxiliary video frame using bag record time, with a configurable maximum difference (20 ms by default). It returns source/video indices and a signed record-time difference, or null for a missing camera, outside coverage or an excessive difference. |
|
|
| It does not establish a device-clock model, jointly pair left/right exposure groups, interpolate motion, or enforce causality. A selected auxiliary frame can occur after the main-frame record time. A future causal application should use a past-only policy. Missing images must not be represented by frame index zero. |
|
|
| Record-time differences and header-minus-record-time offsets cannot be relabelled as exposure-time errors. No fixed latency corrections were fitted during this export. All metadata states capture_time_validated=false. Users needing physical synchronization should establish timestamp semantics and independently validate a clock model using suitable evidence. |
|
|
| ## Playback and gaps |
|
|
| Starting several independent MP4 players together does not synchronize them: each file starts at its own first image. The frame table preserves each stream's offset from the common origin. Per-stream image counts can differ; in particular, left/right frame indices are not guaranteed to form exposure pairs. Source timestamps remain the basis for gap analysis. Any future composite preview is a derived visualization and should provide its selection rule separately. |
|
|
