[caglarpir]

Ricoh Theta X videos carry a CAMM track of type 5 samples: positions without timestamps. So they have no GPS clock, and their capture times come from the camera's clock. The RMKN box that the camera writes into udta is an EXIF block with a GPS fix: date, time and position. That fix is the first GPS sample of the video. This reads the fix and stamps the samples with its time, plus each sample's time in the video after the first sample. That gives these videos a GPS clock, which frame sampling, GPX alignment and uploads already use, as for every video with GPS timestamps.

Changes

• camm_parser: CAMMInfo.rmkn_gps_fix (time and position), read from the RMKN box with the existing ExifReadFromEXIF. None when there is no GPS time or position, when GPSStatus is V (no fix), or when the date is not a valid capture time.
• CAMMVideoExtractor: when a track has only type 5 samples, and the first sample is within 1 m of the fix (the EXIF block rounds the position to 0.01 arc seconds, about 0.3 m), stamps the samples as CAMMGPSPoint, built the way the description file decodes timestamped points, so a video reads the same from either.

Behaviour

• Positions don't change.
• Capture times move by however far the camera's clock is off.
• A GPX track is now placed by the GPS clock, so a GPX stamped by the camera's clock no longer lines up; mapillary_tools warns that the track covers only part of the video.
• A video without an RMKN GPS time (a camera without a fix when the recording starts writes none, and its track begins later), with GPSStatus V, whose first sample is not at the fix, or with type 6 samples reads as before.
• As before, no frames are sampled from the part of a video before its first GPS sample.

Verification

• On 8 Theta X videos, the position of the RMKN fix equals that of the first GPS sample (0.0 mm apart, same altitude), while the second sample is 0.3 to 1 m away. A video without a fix when recording starts (its track starts 63 s in) has no RMKN GPS time.
• main vs this branch on those videos (process, video_process by interval and by distance, with and without a GPX): positions identical; the first sample is stamped with the RMKN time; capture times about 23.6 s later (the camera's clock was behind); the video without an RMKN GPS time unchanged.
• Unit tests for the RMKN reading (fix, no fix, no GPS time, no position, not EXIF, no RMKN box) and for the extractor end to end (stamps type 5 samples from the first one, leaves them alone when the first sample is elsewhere or without RMKN, keeps type 6 timestamps).
• Full test suite on Python 3.9; ruff, usort, mypy on 3.9 and 3.13.

Limitations

• The GPS time is taken as UTC, as EXIF defines GPSTimeStamp.
• The video is assumed to be recorded in real time. A Theta time-lapse video, if the camera writes one with a CAMM track, would get a clock running at 1x, which overrides --video_duration_ratio. Nothing in the RMKN box marks a time-lapse.

Since the first version

Rebased onto 0.15.0 and rewritten: it reads the RMKN box with the existing EXIF reader instead of a hand-written TIFF parser, stamps the samples only when the first one is at the position of the fix, and skips a fix whose status is V.