GoPro 360 (.360)

Facts about the files the GoPro MAX and MAX2 write, verified on the committed MAX2 fixture (test/fixtures/gopro/max2-eac.360, firmware H24.02, 8K mode, structure and telemetry only) and on full MAX2 recordings. The original MAX has no fixture: its facts come from the public references below and are provisional until a recording of it is measured. A fact verified on one body only says so. Public references used for orientation: GoPro's GPMF specification (github.com/gopro/gpmf-parser, Apache-2.0 or MIT) and the MAX geometry published with max2sphere (Apache-2.0).

Container

An ISO base media file (see FORMAT-iso-bmff.md) holding:

• two HEVC video tracks of the same size, each one half of an equi-angular cube map (below). MAX 5.6K: 4096 by 1344. MAX2 8K: 5888 by 1920, Main profile, level 6.0, 8-bit full-range (yuvj420p), 29.97 fps;
• stereo AAC audio, and a four-channel PCM ambisonic track (24-bit on the MAX2);
• a timecode track (tmcd);
• the telemetry track: sample format gpmd, handler "GoPro MET".

Tracks are told apart by handler and format, not position: the order differs by model and mode.

Picture: an equi-angular cube map stitched in camera

The camera stitches its two lenses itself; each video track holds three cube faces side by side.

• Face size is the track height (1344 on the MAX at 5.6K, 1920 on the MAX2 at 8K).

• The first and last face of each track are wider by a band of (width - 3 * height) / 2 pixels (32 on the MAX at 5.6K, 64 on the MAX2 at 8K): the camera's own seam between its lenses runs down the middle of those faces, and the band holds the same columns twice, once from each lens. On the MAX2 the seams sit at columns 992 and 4896 of both tracks (a sharp step in the picture), and the 64 columns before each seam repeat the 64 after it (correlation 0.998 at a shift of 64 on a textured frame). The middle face carries no seam, and the edges between neighbouring faces (columns 1984 and 3904 on the MAX2) are continuous.

• Faces are equi-angular: pixels are evenly spaced in angle across a face, not in tangent, so a face coordinate maps to a direction through atan rather than linearly.

• Which face is where. Measured on MAX2 frames by matching every face edge against every other (the pixel line just inside each edge; the true neighbours correlate at 0.93 to 0.999, and every pair matches both ways): the first track holds the left, front and right faces, the second the down, back and up faces, each stored with its columns as u and rows as v. In the camera's cube frame (front along +x, +z up, -y to the front's right), a face's direction for face coordinates u, v in [-1, 1] is centre + tan(u * pi/4) * across + tan(v * pi/4) * down:

  Track	Face	Centre	across (u)	down (v)	Seam band
  1	left	+y	+x	-z	yes
  1	front	+x	-y	-z	no
  1	right	-y	-x	-z	yes
  2	down	-z	-x	+y	yes
  2	back	-x	+z	+y	no
  2	up	+z	+x	+y	yes

  The second track's faces are therefore turned a quarter against the first's. The original MAX is stated by its public geometry to share this order; it is not yet measured on its frames.

• The band in a seamed face sits at its middle: the face's logical columns [F/2 - b/2, F/2 + b/2) (face size F, band b) are stored twice, first as seen by one lens and then, immediately after, as seen by the other, and the rest of the face follows; a seamed face is F + b columns wide in the track.

Telemetry (GPMF)

The gpmd track's samples are GPMF payloads: big-endian key-length-value records (four-character key, a one-byte type, a one-byte structure size, a big-endian two-byte repeat count, the data padded to four bytes), with DEVC containing STRM streams. Each stream carries sticky labels (STNM name, SIUN unit, ORIN axis order, SCAL divisor, STMP start time) and one data record, its last child. A payload covers about a second, and its samples are spaced evenly over the sample's duration on the track's media clock.

• DVNM, in the DEVC: the camera's name (MAX2 on the MAX2).
• GYRO (rad/s, 800 Hz on the MAX2), ACCL (m/s^2), MAGN: signed 16-bit triples over SCAL. ORIN names, per channel, the camera axis it measures, lower case meaning negated (ZXY on both bodies: Z, X, Y in that order). The reader applies it, then turns the result onto the orientation frame's axes (x, y, z to -x, -z, -y, measured on the MAX2, ADR 0023). Steered by the gyro, the picture shook least with each frame matched to the gyro 10 ms after it.
• SHUT: exposure time in seconds, 32-bit float, one per frame, no SCAL.
• CORI, IORI: signed 16-bit quaternions over SCAL 32767, one per frame. CORI is the camera's orientation since the recording began, IORI the turn the camera applied to the image relative to its body. On the MAX2 IORI is a small turn about a single axis (a few degrees). The stored order is w, x, y, z. Measured on the MAX2 (ADR 0021): CORI is the world's turn to the body, in a frame of x right, y down, z forward; the cube's frame is that frame turned by { w: 0.5, x: -0.5, y: 0.5, z: -0.5 }. CORI is the gyro sampled once a frame (ADR 0023): its turn rates match the gyro's to a scale of 1.000, about 5 ms ahead of it.
• GRAV: the gravity direction, a signed 16-bit unit vector over SCAL 32767, one per frame, in the same body frame as CORI (measured on the MAX2). Negated, it is up.
• Also present and not used: WBAL, WRGB, ISOE, UNIF, PRJT, MNOR, OCNF, WNDM, MWET, AALP, MSKP, LSKP, CSCT, CSCB, LRVO, LRVS, TMPC. Of the camera's attitude streams only GRAV is read for a rendition; CORI and IORI are decoded for the measurement harnesses.
• Personal data, which a fixture holds only as zeros (tools/checks/src/fixtures.ts): GPS5 / GPS9 (location), FACE (detected face boxes), and in the settings block of the moov/udtaGPMF box CASN (the camera's serial), CINF, MUID and CPID (camera and media IDs) and CDAT (the creation time). FIRM/FMWR (firmware) and MINF (the model) identify the model only and stay.

Metadata boxes

Besides the GPMF box, moov/udta holds the camera's own boxes: FIRM (firmware), LENS, SETT, HMMT, BCID, and three that identify the camera or the recording, CAME, MUID and GUMI, which a fixture zeroes too.

Files

A recording longer than about 4 GB is split into chapters named GS<chapter><number>.360: two digits of chapter, then the recording's four-digit number. .LRV files beside them (also named GL...) are low-resolution proxies and .THM files thumbnails.