What orbis-reframe knows about the cameras' recorded files, kept up to date with its development. On GitHub.
DJI 360 cameras file format — .osv
Desk research (2026-10-15); Avata 360 clips examined 2026-11-14 to 11-16.
The desk research comes from public specs and Gyroflow's telemetry-parser
(src/dji/, MIT OR Apache-2.0), which ships the protobuf schemas the
cameras write. The rules in README.md on what may be taken
from others' reverse engineering apply here in particular, because
telemetry-parser's DJI stitching code partly cites DJI Studio's internals.
Markers as in README.md: ✅ is checked on our Avata 360
clips (Modes and sample clips), all from one drone; nothing about the
Osmo 360 family has met a real file yet.
Covers the DJI Osmo 360 (2025), Osmo 360 II, Osmo 360 Air and the Avata 360
drone. The app opens the Avata 360's .OSV files, experimentally
(camera::DjiReader, orbiscore::DjiMeta): the facts marked ✅ below are read
by that code and checked on our clip; the Osmo 360 family isn't opened as
measured, only by the Avata 360's reading (logged).
1. Files and streams
📖 360 video is recorded as
.OSV; single-lens and flat modes as.MP4. Final Cut Pro doesn't open.OSVdirectly (DJI Studio exports MP4).📖 Osmo 360: two square ~1" sensors, f/1.9. 360 video up to 8K 50 (8K 30 for long takes), 6K, and 4K 100. 10-bit is available in every mode, with D-Log M as a log profile. ❓ Codec, presumably H.265; ❓ whether 8-bit stays the default.
✅ A take is three files with one name, in upper case: the
.OSV, a.LRFproxy and a.SRTsubtitle file. The camera names themDJI_<YYYYMMDDhhmmss>_<NNNN>_D_<…>(on our clip the last part is2026; ❓ whether that's always the year) inDCIM/DJI_001/.✅ The
.OSVand.LRFeach keep the full path they were recorded to, name included, in afsidbox inmoov/udta. A renamed file can be matched with the rest of its take by that name.✅ The container is MP4 (
isom; tagsencoder=DJI Avata360,comment=EIS:RS;FOV:Linear;). On the Avata 360 its streams are:# kind what 0, 1 video hvc1the two lenses, one per track: 1920², 3000² or 3840² by the mode (Modes and sample clips), HEVC Main 10, yuv420p10le2, 3 data djmd(handlerCAM meta)the metadata, one track per sensor (§2) 4, 5 data dbgi(CAM dbgi)~47 Mb/s each at 100 fps, ❓ debug data, not read; only in the 24, 30 and 100 fps modes -- at 59.94 fps the file has none and the streams below move up two 6 data camd(CAM camd)3 samples, ❓ not read 7 video MJPEG 960×540 a cover picture, flagged attached_pic: an already reframed flat view, not a fisheyeThere's no audio track.
✅ FFmpeg's mov demuxer (9.0) warns "stream 0, timescale not set" on every open of an
.OSV. It comes from the cover: any MP4 with anattached_picstream gets it (a test file made withffmpegalone does too), and the demuxer names the wrong stream. Harmless; the app keeps FFmpeg's warnings in its verbose log.✅ The lens tracks are tagged BT.709 (primaries, transfer and matrix), limited range, in Normal and D-Log M alike: the tags don't tell log from standard. The colour mode is in the metadata (§2).
📖 Which track holds the master lens (the one at yaw ≈ 0) differs by product, per telemetry-parser: track 1 on the Osmo 360, track 0 on the Osmo 360 II, Osmo 360 Air and Avata 360. Its source is DJI Studio's product table, so we only use what our own files show. ✅ On the Avata 360 clip the seam confirms it: the master lens is on track 0 (§3).
✅ The
.LRFproxy: one HEVC Main (8-bit) track, 1920×960 in every mode, at the clip's rate up to 30 fps (25 fps from the 100 fps mode, 29.97 from 59.94), with both lenses side by side (lens track 0 on the left), plus the same twodjmdtracks and acamd. It's 6 % of the.OSV's size and carries the same calibration, but only one frame'sdjmdsample in four (736 against 2944 on clip0008, stamps 40 ms apart), each with its frame's 40 attitude quaternions: 10 ms of every 40 ms of the attitude. The drone's part changes only on frames the proxy keeps too (every 80 or 160 ms on0008), so the GPS reads the same from either. Reading only thedjmdtrack of the.OSVtakes about as long (45 against 33 ms on0008: the video samples are skipped, not read), so the reader takes every frame's metadata from the.OSVand needs neither file beside it.✅ The
.SRThas one entry per frame up to 59.94 fps, and one every 20 ms (50 Hz, every second frame) at 100 fps: frame count, wall-clock time, ISO, shutter, f-number, EV, colour temperature,color_md(dlog_m, ordefaultfor Normal), metering mode, GPS latitude / longitude and relative / absolute altitude. It has no attitude.
1.1 Photos
The app doesn't open DJI photos yet; what the Avata 360 writes (ten
November photos, exiftool):
- ✅ A 360 photo is a JPEG stitched in the camera, equirectangular,
with the panorama XMP tags (
ProjectionTypeequirectangular,UsePanoramaViewer): 15520×7760 or 7776×3888. One shot is a flat 4000×3000 JPEG instead, with no panorama tags; ❓ by its size the single-lens photo mode (Modes and sample clips). - ✅ With a 360 photo the camera can keep each lens's raw picture as its
own DNG,
…_UP_2026.DNGand…_DOWN_2026.DNGbeside the JPEG: 16-bit Bayer, 7680² per lens for the 15520 px photo, 3840² for the 7776 px one (the spec's 15360×7680 and 7680×3840 DNG are the pair side by side). In the 3840² pair the two lenses' colour filter patterns differ (GRBG up, GBRG down), so one sensor reads out mirrored against the other; the 7680² pair is RGGB in both. - ✅ The JPEG and the DNGs carry DJI's XMP block (
drone-dji): the drone's and the gimbal's attitude in degrees, absolute and relative altitude, speeds, sensor temperature, white balance,SensorFPS, the product nameAvata360. ❓ Whether a photo carries the lens calibration as the video'sdjmddoes: not looked for.
2. Metadata track
✅ The
djmdtracks carry one protobuf sample per video frame (8552 samples for 8552 frames). Only the first sample holdsclip_metaandstream_meta(~6.7 kB); the others onlyframe_meta(~1.4 kB).✅ The two tracks are one per sensor. Their samples are the same apart from the timestamps (16 µs apart), the sensor temperature and the subject-tracking data; both carry the calibration.
📖 The schema differs by product and is chosen by product name, not by the
proto_file_namethe clip declares (DJI reuses one schema under different declared names):product schema pano_dewarp_paramsatStreamMetafieldOsmo 360 dvtm_oq101.proto6 Osmo 360 II, Osmo 360 Air, Avata 360 dvtm_eagle4_wa530.proto5 Parsing with the wrong one silently drops the calibration (the wire types disagree).
✅ The Avata 360 clip declares
proto_file_namedvtm_AVATA360.proto(library 02.01.20, product 02.00.11) and decodes cleanly withdvtm_eagle4_wa530.proto.clip_meta_headernames the productDJI Avata360(no space) and the firmware.✅ Clip metadata on the Avata 360 clip:
imu_sampling_rate4000,sensor_fps100.019,sensor_res3840×3840,LRO_value2443,sensor_readout_time12312719 (❓ units; as ns that's 12.3 ms, against 9.4 ms forLRO_valuein ns × 3840 lines),eis_status1. In the 24-60 fps modesLRO_valueis 4207 (16.2 ms over 3840 lines) whatever the stream size, andsensor_resstays 3840×3840.✅ Stream metadata:
video_stream_meta1920×1920, 100 fps,bit_depth10;color_mode19, which the schema'sColorModeTypenamesCOLOR_MODE_DLOG_M(the.SRTagrees). Normal-colour clips write 0,COLOR_MODE_DEFAULT(the.SRT'sdefault). So the colour profile is in the file; nothing needs the.SRT.✅ Frame metadata, per frame: the camera's exposure (ISO, exposure time as a fraction, white balance, EV,
focal_length2.34 mm,f_number1.8 where the.SRTsays 1.9), the sensor temperature,sensor_shift_param(❓ EIS data); the fused IMU attitude (§4); and on the drone,drone_frame_metaat 10 Hz: attitude as Euler angles in 0.1° (NED), velocity, GPS, altitude, andtracking_infowhile the drone follows a subject.
2.1 Field map (values from the Avata 360 clip)
Field numbers are paths from ProductMeta, names are telemetry-parser's
(dvtm_eagle4_wa530.proto and dvtm_library.proto). Everything listed is
✅ present on our clip; fields the schema has and the clip doesn't write
are left out. Serial numbers and positions are not reproduced here.
clip_meta (1), first sample only
| field | name | Avata 360 value | notes |
|---|---|---|---|
| 1.1.1 | clip_meta_header.proto_file_name |
dvtm_AVATA360.proto |
not the schema to parse with (above) |
| 1.1.2 / 1.1.3 | library_proto_version / product_proto_version |
02.01.20 / 02.00.11 |
|
| 1.1.5 | product_sn |
20 characters | the aircraft's serial |
| 1.1.6 | product_firmware_version |
01.00.06.00 |
|
| 1.1.9 | clip_timestamp |
250472750 | device µs; equals the first frame's frame_timestamp |
| 1.1.10 | product_name |
DJI Avata360 |
picks the schema |
| 1.2.1 | clip_streams_meta.video_stream_num |
1 | although the file has two lens tracks |
| 1.3.1 | distortion_coefficients.coeffients |
four zeros | not the lens calibration (that's §3) |
| 1.4.1 | sensor_readout_time.readout_time |
12312719 | ❓ units (above) |
| 1.5, 1.6 | sensor_read_direction, digital_focal_length |
empty | |
| 1.7.1 | eis_status.status |
1 EIS_ROCK_STEADY |
|
| 1.8.1 | imu_sampling_rate |
4000 | Hz; matches the quaternions per frame (§4) |
| 1.9.1 | sensor_fps.sensor_frame_rate |
100.0188 | the container says exactly 100 |
| 1.10.1 | ITD_value |
−322 | ❓ |
| 1.11.1 | LRO_value |
2443 | ns per line, by the schema |
| 1.12.1 / 1.12.2 | sensor_res.sensor_width / height |
3840 / 3840 | = the calibration frame (§3) |
stream_meta (2), first sample only
| field | name | Avata 360 value | notes |
|---|---|---|---|
| 2.1.3 | stream_meta_header.stream_name |
video |
|
| 2.2.1.4 | camera_stream_meta.device_header.device_name |
DJI FCA188 |
the camera module |
| 2.2.1.5 / 2.2.1.6 | device_frequency / device_timestamp |
100.0 / 250472750 | |
| 2.2.2.1 / 2.2.2.2 | device_version.device_hw_version / device_sw_version |
AC Ver.03 / 10.01.08.95 |
the camera module's own firmware |
| 2.2.3.1 | device_sn.device_sn |
14 characters | the camera module's serial |
| 2.2.4.1 | color_mode.color_mode |
19 COLOR_MODE_DLOG_M |
the colour profile (above) |
| 2.2.5, 2.2.6 | sharpness, denoise |
empty | |
| 2.3.1 / 2.3.2 | video_stream_meta.resolution_width / height |
1920 / 1920 | per lens track |
| 2.3.3 | framerate |
100.0 | |
| 2.3.4 / 2.3.5 | is_bit_depth_valid / bit_depth |
1 / 10 | |
| 2.3.6 | bit_format |
4 BIT_FORMAT_TYPE_YUV420 |
|
| 2.3.8 | video_codec_type |
1 VIDEO_CODEC_TYPE_H265 |
|
| 2.4 | fov_type |
empty | |
| 2.5.1 … 2.5.24 | pano_dewarp_params.* |
24 DewarpParams |
only 2.5.3 / 2.5.4 filled (below, §3) |
DewarpParams, the filled pair: native_refine_far_slave (2.5.3)
and native_refine_far_master (2.5.4). Field numbers within the record.
| field | name | slave (2.5.3) | master (2.5.4) | notes |
|---|---|---|---|---|
| 1 / 2 | fx / fy |
1035.882 / 1036.047 | 1039.390 / 1039.506 | px in the 3840 frame |
| 3 / 4 | cx / cy |
1907.738 / 1924.519 | 1933.663 / 1912.346 | px in the 3840 frame |
| 5-8, 15 | k1 k2 k3 k4 k5 |
0.09104, −0.04372, 0.02720, −0.01079, 0.00132 | 0.07275, −0.02460, 0.01725, −0.00826, 0.00108 | θ-polynomial (§3); k6-k9 (16-19) not written |
| 9 | xi |
not written | not written | no such term in this model |
| 10 / 11 | width / height |
3840 / 3840 | 3840 / 3840 | floats |
| 12 / 13 / 14 | yaw / pitch / roll |
−179.117 / 90.090 / 0.254 | −0.073 / 90.084 / 0.861 | degrees, camera from rig (§3) |
| 20 | p |
0.00005, 0.00054 | 0.00022, −0.00184 | the tangential pair p1, p2: needed for the seam (§3) |
| 21 | q |
four zeros | four zeros | |
| 22 / 23 | occlusion_pt_x / _y |
zeros | zeros | |
| 24 | lens_model |
8 | 8 | a float; ❓ the code for this model |
| 25 | temperature |
23.0 | 23.0 | °C at calibration |
| 26 | cam_imu_extri_q (w x y z) |
0.0073, −0.0006, −0.0021, −1.0000 | 0.9999, −0.0075, 0.0024, 0.0111 | camera to IMU (§4) |
| 27 | tangent_coeff |
0, 0 | 0, 0 | not the tangential pair: that's field 20 |
| 28 | cam_extri_q |
not written | not written | the pose is only in 12-14 |
| 29 / 30 | cam_imu_cali_enable / temp_compen_enable |
1 / 1 | 1 / 1 | |
| 31 | temp_compen_k |
6.65e-5 | 6.65e-5 | ❓ by the name, focal change per °C: ~0.04 % at the clip's 29 °C |
| 32 / 33 | gimbal_yaw_h1 / _h2 |
48 / 49 | 48 / 49 | ❓ |
The native_refine pair (2.5.1 / 2.5.2) holds only temperature (23.0);
the other 20 records hold nothing but zeros.
frame_meta (3), every sample
| field | name | Avata 360 value | notes |
|---|---|---|---|
| 3.1.1 | frame_meta_header.frame_seq_num |
0, 1, 2 … | not written on the first sample (0) |
| 3.1.2 | frame_timestamp |
250472750, +10000 per frame | device µs |
| 3.2.1 | camera_frame_meta.camera_dev_header |
as 2.2.1 | |
| 3.2.3.1 | iso.iso |
200.0 | |
| 3.2.4.1 | exposure_time.exposure_time |
[1, 3252] | seconds as a fraction; the .SRT rounds it to 1/3200 |
| 3.2.5.1 | digital_zoom_ratio |
1.0 | |
| 3.2.6.1 | white_balance_cct |
5429 | K |
| 3.2.8.1 | exposure_value |
0.15625 | EV |
| 3.2.9.1 | color_temp_atmosphere |
5415 | K |
| 3.2.10.1 | focal_length |
[2340, 1000] | 2.34 mm |
| 3.2.11.1 | absolute_altitude.absolute_altitude_m |
the take-off altitude | m |
| 3.2.12.1 | sensor_temperature |
29.0 / 28.0 | °C; differs between the two djmd tracks |
| 3.2.13.1 / 3.2.13.2 | sensor_active_size |
4596 / 4596 | µm |
| 3.2.14.1 | f_number |
[18, 10] | f/1.8; the .SRT says 1.9 |
| 3.2.16.1-4 | sensor_shift_param.sensor_shift_arr* |
four packed uint32 arrays |
❓ EIS data |
| 3.2.18.1 / .2 / .4 | liveview_hfov |
1.8151 / 1.8151 / 1.7778 | rad (104°) and 16:9: the live view, not the recording |
| 3.2.19 | not in the schema | 64 bytes, nested | ❓ |
| 3.3.2.1.1 | IMU_attitude_after_fusion.current_frame.timestamp |
the frame's stamp − 370…400 | device µs (§4) |
| 3.3.2.1.2 | current_frame.vsync |
10428 + frame | |
| 3.3.2.1.3 | current_frame.attitude (w x y z) |
39-41 quaternions | 4 kHz (§4) |
| 3.3.2.2 | IMU_attitude_after_fusion.prev_frame |
the previous frame's batch | first sample only |
| 3.3.3.1 | IMU_vsync_pos.vsync_pos |
3, 5 … | the IMU's downsampling phase, by the schema |
| 3.4.1.5 | drone_frame_meta.drone_dev_header.device_frequency |
10.0 | the drone part changes at 10 Hz |
| 3.4.2.1 / 3.4.2.2 | velocity.velocity_x / _y |
−0.6 / −1.8 | m/s, NED |
| 3.4.3.1-3 | attitude.pitch_decidegree / roll_ / yaw_ |
−5 / 19 / 1781 | 0.1°, NED |
| 3.4.4.1.2 / .3 | gps.gps_coordinates.latitude / longitude |
the position | degrees, doubles |
| 3.4.4.2 / 3.4.4.4 | gps_altitude_mm / gps_altitude_type |
the altitude / 1 GPS_FUSION_ALTITUDE |
|
| 3.4.5.1 / 3.4.5.2 | relative_altitude.relative_altitude_mm / is_relative_altitude_valid |
2200.0 / 1 | above take-off |
| 3.4.7.1 | rc_mode.rc_mode |
7 DJI_COMMAND_MODE_GPS_NORMAL |
|
| 3.4.8.1 / 3.4.8.2 | barr_dist.below_tof_state / below_dist_cm |
1 TOF_STATE_REFLECTED / 23 |
|
| 3.4.9.1 | fai_degree.fai_g_deg |
−503 | 0.1°, ❓ |
| 3.4.11.2 | gimbal_angle_deviation.y_deviation |
0.00035 | ❓ |
| 3.4.13 | tracking_info |
on some frames | the subject the drone follows |
| 3.4.14 | not in the schema | 58 bytes | ❓ |
3. Lens calibration
📖
PanoDewarpParamsholds 24 calibrations, a slave/master pair for each of 12 cases: native, native-refined, refined at infinity ("far"), refined at stated distances (far_07…far_16), lens guards, and the waterproof case above and under water.✅ On the Avata 360 clip only the
native_refine_farpair is filled. Thenative_refinepair holds just a temperature; the other 20 are empty. So there's no slot to choose on this body.📖 Each
DewarpParams:fx fy cx cy,width height,yaw pitch roll,k1…k9,xi, tangentialp/tangent_coeff,lens_model, an occlusion polygon (occlusion_pt_x/y: the body in the picture), temperature compensation,cam_extri_q(pose between the two camera modules) andcam_imu_extri_q(camera-to-IMU rotation).✅ What the Avata 360 clip fills in:
width/height3840 (the sensor, twice the 1920 stream: every pixel value halves),fx fy cx cy,k1…k5(non-zero),lens_model8,yaw pitch roll,cam_imu_extri_q, the tangential pairp(field 20, non-zero), temperature compensation (enabled,k6.65e-5) andgimbal_yaw_h1/h2(48 / 49, ❓).tangent_coeff(field 27),qand the occlusion polygon are zero; there's noxiand nocam_extri_q. Values in §2.1.📖 The lens model telemetry-parser reads is an equidistant θ-polynomial (Kannala-Brandt) with a tangential pair:
θ_d = θ·(1 + k1θ² + k2θ⁴ + k3θ⁶ + k4θ⁸ + k5θ¹⁰) x' = θ_d·X/ρ, y' = θ_d·Y/ρ xd = x' + 2p1·x'y' + p2(θ_d² + 2x'²), yd = y' + p1(θ_d² + 2y'²) + 2p2·x'y' u = fx·xd + cx, v = fy·yd + cy✅ With the Avata 360's values this model, in image coordinates with x right and y down, registers the two lenses at the seam (below), and the tangential pair is needed: without it the seam is off by ~1° (about 9 px at 1920).
pis field 20 inp1, p2order; swapping or negating the pair breaks the seam. There's noxi: the model has no such term, so the lens isn't an MEI lens withxi = 0. The app's stitch implements MEI only (X5 §5.3), so this needs a second model.📖
yaw pitch rollandcam_extri_qdescribe the same pose (the quaternion is the triple composedRy(yaw)·Rz(roll)·Rx(pitch)), camera from rig, in a frame whose up isz(pitch ≈ 90° is a level lens).✅ On the Avata 360 clip: slave
yaw −179.12 pitch 90.09 roll 0.25, master−0.07 / 90.08 / 0.86, so the two optical axes are 179.04° apart. Read as camera-from-rigRy(yaw)·Rz(roll)·Rx(pitch − 90), the file's angles register the seam; an exact back-to-back pair leaves a 0.94° tilt, and dropping the roll a 0.6° twist besides. Other rotation orders, a negated roll, the transposed rotation and mirrored image axes all match worse.✅ The master lens is on track 0, the slave on track 1, as telemetry-parser says for the Avata 360. Measured in a seam band 88-92° off-axis over 16 frames, with the whole model (tangential pair and the file's angles): the two lenses correlate at 0.62 with no shift, and what's left once around the seam is 0.21° radial and 0.08° tangential. With the slave on track 0 the same model correlates at 0.02. Without the tangential pair neither order registers (0.15-0.17, ~1° off), which is why the order looked undecided before that pair was read. The image circle's centre isn't a usable check: the lens housing's glare bounds the picture, and on track 0 the circle runs off the frame's sides.
✅ Track 0 is flagged as the default video track; track 1 isn't.
✅ The 1920 stream is the whole 3840 calibration frame at half size: taking it as a central crop instead (3776 px, scaled) breaks the seam (sectors that match fall from 96 % to under a third).
✅ Every mode carries the same 3840² calibration, and every stream size is the whole frame scaled: 3840 as it is, 1920 at half, and 3000 at 0.78125, not a whole factor. The 3000 px frame's image circle sits where the 3840 px one does, and the seam measures the same on all three sizes (
stitch_source_smoke --measure-seam, 1-6 s of six November clips: seam offset 0.03-0.10° before the per-clip fit, 0.02-0.04° after, its lens-1 corrections up to 0.16°). The reader lists 3000 as a known scale for the Avata 360 and logs any other size that isn't a whole factor, in case it is a crop.✅ Left after the calibration: a constant 0.2-0.3° radial offset that changes from frame to frame with the scene. It's parallax between the two lenses' centres, which sit a few centimetres apart along the axis.
📖 No lens translation is stated: the stitching distance is baked into the "far" ladder of focal lengths instead.
Ours to measure (see
README.md): which of the 24 slots to stitch with by default and per accessory, the blend and seam widths, and how far off-axis each lens's picture is usable. telemetry-parser's values for these come from DJI Studio and aren't used.✅ How far off its axis each picture reaches (2026-11-15, the averaged lens pictures of 16 frames, by the file's model): ~101° to the frame's nearest edge (the circle runs off track 0's sides), the housing's glare beyond. The stitch takes 200° as the usable field of view, as the X5's bare lens.
✅ A part of the drone reaches into the top of each picture (track 0's top left, track 1's top), from ~91° off the axis outwards: inside the seam band, over the ~100° of the seam behind the drone, in both lenses. There the two lenses share only a ~6° wide clean strip round the seam circle (measured with
seam_regress: the lenses agree at NCC 0.96-1.00 along the rest of the seam, 0.7 or less there). The lenses' parallax is large for near things: the pilot ~2 m behind the drone is ~1° apart across the seam at a seam at infinity, far scenes 0.1-0.2° (DJI_…_0008, last frame); the seam flow and the depth-aware seam both join it. Left in, the part spoils the stitch there: a cross-fade over the whole band blends its edge in as a thin dark line along the seam, and the new stitcher's seam colour blend measures it as if both lenses saw the sky (its veiling light and gain reach their bounds there; a darker, bluish block in the sky). The stitch leaves it out by a mask per lens (§3.2).
3.1 The mount (ours)
- ✅ The app's stitch places the two lenses by the file's angles (above:
camera from rig
Ry(yaw)·Rz(roll)·Rx(pitch − 90), our lens frame the image's turned half a turn about the axis) and turns the rig into the sphere by a mount per product (camera::DjiReader's table): on the Avata 360 the master's axis to the sphere's up and the drone's nose to its forward, which is towards the bottom edge (image y down) of the master lens's picture: yaw 90°, pitch −90°, roll 90° in the stitch'soutYaw/outPitch/outRollterms. - ✅ Measured on our clip (2026-11-15) from the sun: in 23 frames over the
85 s, the sun's direction in our stitched sphere against its direction in
the drone's body (its azimuth and elevation from the GPS position and the
UTC time,
creation_time, through the drone's own attitude, 3.4.3, NED Z-Y-X). The best rotation between the two puts the nose at −90.4° in the sphere before the mount's quarter turn (heading alone: −89.9°, 7° rms); the frames the sun sits behind a tree or a roof are the outliers. Frame 0 then looks where the cover picture looks. - ⚠️ The same fit tilts the camera ~4° against the body (its forward 4.3° below the body's), within what the sun's flare and the attitude's timing allow; not applied. The stabilization step, with a turning clip, measures it properly.
3.2 The body in the pictures (ours)
- ✅ Where each lens's picture shows the drone (2026-11-15), as the clear
radius off the lens's axis every 2° of image angle (counted from the
image's +x towards +y, y down;
camera::DjiReader's table, the stitch'sorbiscore::LensMask). Measured on our six clips, 12 frames each: each lens's rim unwrapped by image angle and off-axis angle (80-103°) through the file's calibration; per clip each pixel's spread over its frames; the body is where the median of that over the clips stays below 31 levels (the scene above ~32 -- the drone moves, the body in its picture doesn't), traced from the rim inwards. Stretches narrower than 8° dropped (scene streaks at the rim), gaps narrower than 8° filled, then drawn half a degree in and to the nearer of the next 2° either side. - ✅ The master's (track 0) runs from image angle 220° over the top (270°) to 22°, in to 92.5°; the slave's from 180° to 356°, in to 91.2°. Behind the drone the lenses then share only ~4° round the seam circle: the master sees the scene up to 2.6° past it, the slave 1.2-2°.
- ✅ The stitch takes the masked stretch from the other lens (its feather
goes over to that lens over the last 2° before the mask, or half the
shared strip where that's narrower) and measures nothing there -- the
seam's statistics, and with them the colour blend, veiling light and
lens colour; the depth costs; the seam alignment; the flow's strips show
the other lens's picture there. Measured: the thin line at
DJI_…_000816.7 s (colour blend off, and the fixed seam) and the sky block atDJI_…_000516.6 s andDJI_…_000143 s are gone; the veiling light measured in those stretches falls from 0.37 / 0.22 to 0.12 / 0.08 (linear). Cameras without a mask stitch as before, to the bit. - ❓ The file's
occlusion_pt_x/ypolygon is zero on the Avata 360; where a product fills it, it may say the same (§5).
4. IMU
- 📖 On the Osmo 360: raw gyro (rad/s) and accelerometer (g), stated at 1000 Hz, plus DJI's own fused attitude quaternion per reading, packed per video frame.
- ✅ The Avata 360 clip has only the fused attitude: no raw gyro or
accelerometer (
IMU_raw_attitudeis absent). So for the app it's an orientation track, like the Antigravity A1's, not gyro samples. - ✅ Per frame,
IMU_attitude_after_fusion.current_frameholds a timestamp, a vsync count and 39-41 quaternions (w x y z): 4 kHz, asimu_sampling_ratesays. Its timestamp sits 370-400 µs before the frame'sframe_timestamp; frame timestamps step 10000 µs (9983 at the least). In the slower modes the batch grows with the frame interval and starts earlier: 66-67 quaternions and ~800 µs before the frame at 59.94 fps (stamps 16700 µs apart), 133-134 and ~800 µs at 29.97 (33367 µs). By the schema the stamp is the batch's first sample's; it's auint32(it wraps after 71.6 minutes of the drone's uptime), the frame's auint64. Field 4 (offset, first sensor row to first sample) isn't written. - ✅ The batches run on without a seam: against the running sample count
the stamps lie on one straight line of exactly 250 µs per sample, each
within ±125 µs of it (rms 71 µs: a 250 µs grid), on all six clips; the
step from one batch's last sample to the next one's first is the same as
within a batch. No jumps: the largest step between samples on five clips
is 280°/s, a real turn; on
0001one step of 2718°/s in its last 0.1 s, where the recording ends. - ✅ The quaternion turns IMU axes into the world, NED (world down is
+z, north +x): world down in the IMU's axes is −x while the drone is
level, and a hand-eye fit of its turns against
drone_frame_meta's attitude (10 Hz, lag 0) puts the IMU's x up, y right and z forward on the drone, to ~1.2°. - ✅ Against the picture: the rotation over each frame interval of four
unstabilized 90° views (optical flow, the rotation of the tracked
features), fitted against the attitude's over the same interval, on all
six clips. In the sphere's axes (+X right, +Y up, −Z forward) the IMU is
y,x,-z, gain 0.98-1.03. The fit leaves 0.004-0.06° per frame, against 0.08-0.48° of motion. ❓ It also puts a turn of +1.6° about the forward axis on top (+1.3 to +2.0° on five clips, under 0.4° about the other two; 2.6° on0008, whose fit near the ground is the weakest), which neithercam_imu_extri_q(either way round) nor the lens pose predicts (the closest gives −0.35°). Six November clips in the other modes (same drone) give the same map and sign: +0.6 to +3.9°, median +2.3°. The level picture doesn't bear it out: Horizon lock with the plain map leaves the horizon of0005(1.5 km up) at −0.7° and the verticals of0001(street level) at +0.1°; with the 1.6° turn, at +0.6° and +1.4°. So the reader applies the plain map; where the turn comes from is open (§5). - ✅ Timing: the picture's rotation lines up with the attitude 4.6-5.8 ms
after
frame_timestamp(six clips), about half the readout below (4.7 ms): so a frame's stamp is where its readout starts, and the reader takes each frame at its middle row, half the readout after its stamp. The longer readout of the 24-60 fps modes moves it as it should: 7.4-9.7 ms (median 7.8) on five clips at 29.97 and 59.94 fps, against half of 16.2 ms; 4.4 ms on a sixth at 100 fps. - ✅ The fused heading holds: against the drone's own compass heading it
stays within 1-6° over each clip (30 s to 3:40). On
0010the two part by 56° in a hard turn (5.4-5.7 s, roll to 42°) and stay apart; the picture follows the fused attitude through it (0.01-0.02° per frame left), so it's the drone's heading that went wrong there. - ✅ The
.LRFproxy carries the same batches for one frame in four only (§1): the attitude comes from the.OSV. - ✅ Each lens's
cam_imu_extri_qon the Avata 360: about a half turn aboutzfor the slave (w 0.007, z −1.000), about identity for the master (w 1.000, z 0.011, a 1.56° turn in all). - ❓ What frame
cam_imu_extri_qis written in. Both lenses' records name the same IMU (the twodjmdtracks carry the same attitude, §1), so the turn between the two quaternions should be the turn between the two lenses thatyaw pitch rollgive, whichever way round the quaternions are read. It isn't: the quaternions put the lenses 179.56° apart, a half turn about theirz; the lens poses, which the seam confirms (§3), 179.04° apart, a half turn about the image'sy. A turn's angle doesn't depend on the axes it's written in, so the 0.5° can't be a different axis convention; the axes (zagainsty) say the quaternions' camera frame isn't the image frame ofyaw pitch rolleither. Neither reading of the master's quaternion fits the rotation fit above (the closest leaves 0.35° about the forward axis against 1.6°, and adds about a degree about the other two). The app doesn't use it. - ✅ The Avata 360 flies its camera with the lens axes vertical: on our
clip track 0 (the master lens) sees the sky and track 1 (the slave) the
ground while the drone is near level (−0.5° pitch, 1.9° roll by
drone_frame_meta). - 📖 Timing isn't exact from the stamps: the Osmo 360 II delivers a reading every ~995 µs, packets carry a fixed-offset stamp while the reading count drifts, and the IMU clock wanders against the frame clock (5 ms over 5 minutes on an Osmo 360). telemetry-parser fits a straight line of anchor against reading index over neighbouring packets.
- 📖 Raw axes:
zyxrelative to the fused attitude in a 360 recording (Gyroflow's orientation notation, X5 §6.3). - 📖 Rolling shutter: from
LRO_value(ns per sensor line) × sensor height, 18.3 ms on the Osmo 360; the genericsensor_readout_time(24.0 ms) is not what DJI's 360 stabilization uses. ✅ On the Avata 360, 2443 ns × 3840 lines = 9.4 ms (sensor_readout_timesays 12.3 ms): the lag above fits half of the 9.4 ms; in the 24-60 fps modes 4207 ns × 3840 = 16.2 ms, and the lag fits half of that too. ❓ The direction of the readout isn't checked: on0010's fastest turn (250°/s) the seam shows no doubling with the per-row correction or without it.
Modes and sample clips
All the modes the Avata 360 offers 📖 (DJI's spec page, and its firmware notes for what came later, §6), with the clips we have and whether that is enough. Enough means the file is examined here and what the app reads per mode (lens size, calibration, attitude, readout) is checked on it; one clip means examined, but a second take would check what one clip can't (a moving camera for stabilization, the other colour); wanted means no clip yet.
All our files come from one Avata 360, a contributor's, each take with its
.LRF and (but for one) its .SRT; every video is 10-bit HEVC. Short
names in this document (0008, 0010) are the October clips; the
November ones are written 1115_NNNN. The spec gives the rates as
60 / 30 / 24, which the files write as 59.94 / 29.97 / 23.976 (only 100 is
exact); the names below are the spec's, the sizes per lens track.
| Mode | Spec | Files | Our clips | Status |
|---|---|---|---|---|
| 360 Video 8K (3840² per lens) | 60 / 50 / 48 / 30 / 25 / 24 | DJI_…_D_….OSV + .LRF + .SRT |
at 60: Normal 1115_0008; D-Log M 1115_0009, 0010, 0017-0020, near the ground and in flight. At 30 and 24: one Normal clip each (1115_0007, 0006); fw 01.00.08 |
enough at 60; one clip at 30 / 24; 50 / 48 / 25 wanted (do they write exact rates, and the same readout?) |
| 360 Video 6K (3000² per lens) | 60 / 50 / 48 / 30 / 25 / 24 | as above | at 60: Normal 1115_0004, D-Log M 1115_0005. At 30 and 24: one Normal clip each (1115_0003, 0002); fw 01.00.08 |
enough at 60 (the 0.78125 scale, §3); one clip at 30 / 24; 50 / 48 / 25 wanted |
| 360 Video 4K (1920² per lens) | 100 (added by firmware 01.00.03, 2026-06-23; not on the spec page) | as above | D-Log M: six October flights, 30 s to 3:40, fw 01.00.06 / .08; Normal: 1115_0001, 10 s |
enough |
| Colour: Normal, D-Log M | color_mode 0 / 19 (§2) |
both, at 100 and at 60 in 8K and 6K | enough; D-Log M at 30 / 24 not seen (a check, low) | |
| QuickShots, FocusTrack (360 mode) | 📖 by remote controller | ❓ presumably a normal .OSV |
FocusTrack: the October subject-following clips (tracking_info, §2.1) |
FocusTrack enough; a QuickShot wanted (low: its file, any marker in djmd) |
| Single-lens Video 4K | 3840×2160 at 60…24; 3840×2880 (4:3) at 60…24, only with the goggles and a motion / FPV controller | ❓ .MP4 (§1) |
none | wanted: a flat video, but does it carry djmd, which FOV (Natural Wide, Standard (Dewarp), Wide) and stabilization (RockSteady, HorizonBalancing) it was taken with |
| Single-lens Video 2.7K | 2688×1512 and 2688×2016 (4:3) at 120 / 100 / 60…24 | as above | none | wanted (low once one single-lens clip is in) |
| 360 Photo 120 MP | JPEG 15520×7760, DNG | DJI_…_D_….JPG (stitched, §1.1) + _UP / _DOWN .DNG at 7680² |
seven, each with its DNG pair, at 12-122 m above take-off | enough to start photo support |
| 360 Photo 30 MP | JPEG 7776×3888, DNG | as above, DNG at 3840² | two: one with its DNG pair, one JPEG only | enough |
| Single-lens Photo | (added by firmware 01.00.03, 2026-06-23; no size stated) | .JPG, flat |
❓ 1115_0011, 4000×3000, 2.3 mm (28 mm equivalent), no panorama tags: by its size a single-lens photo |
one; a single-lens DNG wanted (low) |
| Internal storage (42 GB) or microSD | ❓ the same files | ❓ which ours came from | wanted, low: whether the internal storage writes the same names and fsid path |
The spec page lists only Single Shot among the photo modes. Our
photos carry XPKeywords single or smart (smart on six of ten,
the 4000×3000 one among them; either kind comes with or without DNGs); ❓ what smart stands for. Each photo's
SensorFPS is one of the video rates (23.97, 47.95, 59.94, 25), as if the
sensor ran in the video mode last set.
The file names no mode beyond its size and rate (no capture-mode name in the metadata); the app names the mode by the lens size, with DJI's names (3840² 8K, 3000² 6K, 1920² 4K).
Wanted beyond the table: clips from a second Avata 360 (for the IMU's turn, §4, and the camera's tilt, §3.1), and anything from the Osmo 360 family.
5. Open questions
- The same lens reading and mount on a second Avata 360 (every mode of our drone stitches with it, §3); the camera's tilt against the body (§3.1).
- Whether the IMU sits turned against the sphere about the forward
axis: the rotation fit says 1.6°, the level picture under a degree
(§4); a second Avata 360 would help. What frame
cam_imu_extri_qis written in: it doesn't agree with the lens poses (§4), so it can't give the lens-to-body turn until that's known. The other modes of the same drone write the same quaternions, so only a second unit can show whether it follows the picture. - Everything above on the Osmo 360 family: tracks, sizes, codec, the schema per product, the calibration slots and the raw IMU.
- D-Log M's transfer curve.
- Which calibration slot stitches best, per accessory, where a body fills more than one (measured on our footage).
- Whether another product or firmware fills
occlusion_pt_x/y, and in what terms: it would give the body's outline (§3.2) from the file.
6. Sources
- Our Avata 360 clips (Modes and sample clips): the streams by
ffprobe; thedjmdsamples decoded with telemetry-parser's.protofiles; the seam measured by sampling both lenses into a band around the seam and matching them, 16 frames, for each lens order, pose reading and tangential reading; the mount from the sun's position (NOAA's solar position formulas) against the stitched frames; the attitude's map and lag by fitting each frame interval's rotation of four unstabilized views against it. The photos' tags byexiftool. - Gyroflow telemetry-parser,
src/dji/(mod.rs,rig.rs,timing.rs,dvtm_oq101.proto,dvtm_eagle4_wa530.proto,dvtm_library.proto): https://github.com/AdrianEddy/telemetry-parser/tree/master/src/dji - AvataCraft-desktop, a fork of this project with an Avata 360 reader,
whose notes on the same clip agree with what's marked ✅ above, except
that they put the slave lens on track 0, leave out the tangential pair,
and give the
.SRTone entry per frame and ~16 quaternions per frame: https://github.com/pcwizardsinc/AvataCraft-desktop - DJI, Avata 360 specs (the 360 and single-lens video modes, photo sizes, colour modes, H.265 up to 180 Mb/s): https://www.dji.com/avata-360/specs
- The Avata 360's firmware notes as reported (01.00.03, 2026-06-23: 4K 100 in 360 mode, single-lens photos; 01.00.06 and 01.00.08 add no recording modes): https://www.drohnen.de/86591/dji-avata-360-firmware-v01-00-0300-4k-100fps-single-lens-fotos-update/, https://shop.coptrz.com/blogs/news/dji-avata-360-v01-00-0600-september-2026-firmware, https://dronexl.co/2026/10/25/dji-avata-360-firmware-v01000800-waypoints-rc-pro-2/
- Heliguy, features by control method and mode (QuickShots, FocusTrack): https://www.heliguy.com/blogs/knowledge-base/how-do-the-features-of-dji-avata-360-differ-across-various-control-methods-and-shooting-modes/
- Phototrend, Osmo 360 dual 1-inch sensors: https://phototrend.fr/en/2025/07/dji-osmo-360-dual-1-inch-sensors/
- Conns Cameras, Osmo 360 specs: https://connscameras.ie/cameras-video/dji-osmo-360-8k-revolutionary-360-camera-with-standard-combo/6937224110502/67191/
- Apple Community,
.OSVfiles in Final Cut Pro: https://discussions.apple.com/thread/256108341