Pocket 3 original media

Part of the Osmo Pocket 3 reference. Read the overview for firmware, survey scope and evidence levels.

Color, recording and preserved media

Pocket 3 color requests use 0x02/0x42: Normal 00, HLG 3C, D-Log M 3D. HLG and D-Log M selections were accepted and corroborated by separate image-effect status. Do not substitute Pocket 4 color values.

General → Video Compression offered Efficiency (HEVC) and Compatibility (H.264). H.264 appeared dimmed in the D-Log M sequence, became selectable in Normal, and its selection correlated with an accepted two-byte 0x02/0xAB request. The Low-Light General panel later showed HEVC. This is a partial availability comparison, not a complete color/codec matrix or proof of the live-preview codec.

Short takes were started and stopped in HLG, D-Log M, Normal/H.264, Med-Tele and square 3K/60. The known record request 0x02/0x02 (01 start, 00 stop) was accepted, and recording status followed. Mimo additionally sent a 62-byte glamour parameter SET after record start; those ancillary requests returned DF while recording continued. A rejected glamour write must not be mistaken for a rejected record command.

The following 20 Mimo downloads preserved from the phone over USB have independent File evidence. Capture settings identify the experiment; file metadata supplies the measured output. Each import was size-checked, hashed and parsed. Subsequently, 20 complete camera HTTP originals matched their phone imports by SHA-256, covering every row below, including the later Glamour-on take. The listed times are container durations unless a video-only duration is explicitly identified.

Experiment	Inspected downloaded output
Video HLG, 4K/25	HEVC, 3840×2160, 10-bit, 25 fps, 180 frames, 7.200000s; BT.2020/HLG tags
Video D-Log M, 4K/25	HEVC, 3840×2160, 10-bit, 25 fps, 181 frames, 7.240000s; BT.709 tags
Video Normal/H.264, 4K/25	H.264, 3840×2160, 8-bit, 25 fps, 184 frames, 7.360000s
Med-Tele, 2.7K/25	H.264, 2688×1512, 8-bit, 25 fps, 126 frames, 5.040000s
Square Video, 3K/60	HEVC, 3072×3072, 10-bit, 60000/1001 fps, 305 frames, 5.098667s
Square Video, 3K/60, Glamour master on and Smooth zero	HEVC Main 10, 3072×3072, 10-bit, 60000/1001 fps, 480 decoded frames; 8.008s video, 8.021333s container
Low-Light, 4K/30	HEVC, 3840×2160, 8-bit, 30000/1001 fps, 174 frames, 5.805800s
Full five-minute Raw+Video Timelapse, 4K/30, 2s	HEVC, 3840×2160, 8-bit, 30000/1001 fps, 151 frames, 5.038367s
Early-stopped Raw+Video Timelapse	HEVC, 3840×2160, 8-bit, 30000/1001 fps, 9 frames, 0.300300s
Custom Motionlapse	HEVC, 2688×1512, 8-bit, 25 fps, 14 frames, 0.560000s
4K/120 HLG Slow Motion	HEVC, 3840×2160, 10-bit, 30000/1001 playback fps, 476 frames, 15.882533s; BT.2020/HLG tags
4K/120 D-Log M Slow Motion	HEVC, 3840×2160, 10-bit, 30000/1001 playback fps, 476 frames, 15.882533s; BT.709 tags
1080P/240 Slow Motion	HEVC, 1920×1080, 10-bit, 30000/1001 playback fps, 1174 frames, 39.172467s
Auto Hyperlapse	HEVC, 3840×2160, 10-bit, 30000/1001 fps, 63 frames, 2.133333s
2X Hyperlapse	HEVC, 3840×2160, 10-bit, 30000/1001 fps, 198 frames, 6.613333s
Photo, 16:9	JPEG, 3840×2160
Photo, 1:1	JPEG, 3072×3072
Panorama, 180°	JPEG, 4096×1536
Panorama, 3×3 grid	JPEG, 4000×3840
Panorama, 3×3 grid with RAW selected	JPEG, 4000×3840; nine full-resolution DNG components subsequently preserved from the card

The square 3K/60 take is a useful dependency to investigate: the preceding camera color status was Normal and the last explicit compression SET was H.264, but the preserved output is HEVC 10-bit. An automatic format-dependent codec change is a candidate explanation, not a rule established by this sample. The later Glamour-on take also reported Normal before recording and has the same square HEVC 10-bit encoding. Its 480 video frames decode successfully, but Smooth was zero and the scene contained no face: this camera recording does not establish a nonzero Smooth effect. Processed local-editor exports are documented in the Mimo album and exports reference.

Audio-stream inspection also distinguishes menu visibility from saved output:

Preserved takes	Audio streams
Video HLG and D-Log M	AAC, 48 kHz, mono
Normal/H.264 Video, Med-Tele, both square Video takes, Low-Light, Auto and 2X Hyperlapse	AAC, 48 kHz, stereo
All three Slow Motion takes, both Raw+Video Timelapse takes, Custom Motionlapse	No audio stream

The mono/stereo results reflect the selected settings in those takes, not a color-mode channel restriction. Audio rows and meters remained visible in some modes whose MP4 outputs contain no audio stream. The card copy supplies separate AAC-LC, 48 kHz stereo ADTS files for all three inspected Slow Motion takes:

Slow Motion take	Separate AAC duration
1080P/240	4.906667s
4K/120 D-Log M	3.989333s
4K/120 HLG	3.989333s

All three audio files fully decode. Their durations are consistent with real-time capture alongside the longer slow-motion playback; sample-accurate audio/video alignment remains unverified. These AAC companions are separate from the external-microphone WAV backup feature, which remains untested.

The D-Log M recording-mode identity comes from the accepted color setting and camera state; BT.709 tags alone do not identify D-Log M or prove a Rec.709 recording. The Slow Motion files also show why the shooting-rate setting must be kept separate from container playback rate. Metadata alone does not calibrate the transfer curve or verify audio/video synchronization. Full hashes establish unchanged downloads for the 20 matched phone files; the later card copy also matches all 25 previously preserved camera HTTP files. File integrity does not establish every mode’s output behavior. Recording settings and the live-preview signal are separate; see live view and HTTP media.

Camera HTTP originals and companions

The initial independently verified camera-download set contains 24 complete files: the 20 matched originals, the square Photo DNG and three LRF previews. Each saved transfer has contiguous HTTP 206 ranges, a consistent total length and ETag, matching local length, and a recomputed SHA-256 matching its transfer manifest. Incomplete transfers are excluded. This upgrades the earlier partial Low-Light comparison, whose passive phone capture had missing body ranges; it does not make those earlier captures complete.

The three inspected LRFs contain H.264 8-bit video at 30000/1001 fps: the Low-Light preview is 1280×720, while both square 3K/60 takes’ previews are 720×720. Preview dimensions, rate and codec therefore must not stand in for the paired camera original’s properties.

On this firmware, requests using the camera’s exact slash-separated file path worked, while percent-encoding the path separators returned 404. The camera served byte ranges. Preserve separators when encoding individual path segments; verify response ranges and total length before treating a download as complete.

USB card copy and source sets

Selecting Transfer File/OTG Connection on the camera mounted its card over USB. The complete copy contains 361 files totaling 5,073,372,646 bytes; all source/copy SHA-256 comparisons matched, with zero copy-validation errors. The card was then ejected. This inventory includes older takes and auxiliary card files, including MISC contents. It is not 361 new survey recordings. All 25 previously verified camera HTTP files, totaling 927,102,530 bytes, independently match files in this card copy; the 20 matched phone imports are a subset. These overlapping collections must not be added together.

The preserved RAW inventory is:

Capture	DNG count	Full CFA dimensions	Total DNG file bytes
Square Photo	1	3072×3072	19,314,420
RAW-selected 3×3 Panorama	9	3072×3072	173,940,344
Short Raw+Video Timelapse	9	3840×2160	155,263,488
Full five-minute Raw+Video Timelapse	151	3840×2160	2,605,540,864

All 170 DNGs contain five TIFF directories (IFDs) and one full-resolution RGGB 2×2 CFA array, stored as one uncompressed 16-bit sample per pixel. The square arrays each occupy 18,874,368 bytes and the landscape arrays 16,588,800 bytes; the complete files also contain previews and metadata. Standard referenced data ranges and image bounds pass validation, and all 170 RAW arrays have distinct hashes. Sample statistics on ten representative files confirm nonconstant image data. This does not establish effective sensor precision, demosaiced image quality or the meaning of opaque MakerNote fields.

The observed card layout separates rendered outputs from sequence components:

Observed location	Contents in the inspected takes
DCIM/DJI_001	MP4/JPEG outputs, the same-stem square Photo DNG, and same-stem Slow Motion AAC sidecars
DCIM/PANORAMA/001_<take>	PANO_<index>.JPG for the JPEG takes; PANO_<index>.DNG for the RAW-selected take
DCIM/TIMELAPSE/001_<take>	TIMELAPSE_<index>.DNG for both Raw+Video takes

Here <take> and <index> replace the observed take numbers and four-digit source indices. Directory suffixes, capture times and controlled recordings associate these sets with their outputs. They do not establish a universal naming algorithm or pixel-by-pixel source-to-video alignment. All 13 source JPEGs from the two JPEG panorama takes decode cleanly. A later Mac Wi-Fi pass retrieved two known nested source paths using /v2?storage=0&path=<camera-relative-path>, retaining the directory separators:

Known component	Complete bytes	Contiguous validated HTTP 206 ranges
Short Timelapse, TIMELAPSE_0001.DNG	17,250,304	17
RAW Panorama, PANO_0001.DNG	19,333,016	19

Both complete downloaded SHA-256 values match their preserved SD files and transfer manifests. This verifies retrieval of these exact known nested paths; automatic source discovery, a universal naming algorithm and every source layout remain unverified.