edessa/new-multiview-recordings

New MultiView recordings

Nineteen sessions (7–100 min) recorded 2026-11-03 to 2026-11-08: up to six synchronized Basler daA1280-54uc colour cameras (1280×960, 24 fps, software-triggered) with simultaneous 128-channel EMG from an Intan RHX recording controller (2 kHz).

Session	Length	Frames / camera	Rounds with all cameras	Calibration recorded
subject02_session01	67.9 min	~97,730	97,720	16:08 (calib_20261103_1608)
subject03_session01	61.1 min	~87,925	87,919	17:43 (calib_20261103_1743)
subject05_session01	70.5 min	~101,510	101,485	19:06 (calib_20261103_1906)
subject06_session01	35.2 min	~50,709	50,703	2026-11-05 13:27 (calib_20261105_1327)
subject07_session01	65.4 min	~94,222	94,197	2026-11-05 14:54 (calib_20261105_1454)
subject07_session02	70.6 min	~101,651	101,639	2026-11-05 16:26 (calib_20261105_1626)
subject09_session01	12.8 min	~17,830–18,390	17,825	2026-11-06 11:19 (calib_20261106_1119)
subject09_session02	19.8 min	~28,533	28,532 (5 cameras)	2026-11-06 11:19 (calib_20261106_1119)
subject10_session01	60.6 min	~87,236	87,223	2026-11-06 13:14 (calib_20261106_1314)
subject11_session01	22.7 min	~32,592	32,590	2026-11-06 14:42 (calib_20261106_1442)
subject11_session02	22.2 min	~31,988	31,981	2026-11-06 15:09 (calib_20261106_1509)
subject12_session02	52.4 min	~75,429	75,407	2026-11-06 16:11 (calib_20261106_1611)
subject13_session01	7.4 min	~10,649	10,647	2026-11-06 17:24 (calib_20261106_1724)
subject13_session02	52.0 min	~74,824	74,809	2026-11-06 17:24 (calib_20261106_1724)
subject14_session01	23.5 min	~33,782 (cam2 33,616)	33,616	2026-11-07 14:19 (calib_20261107_1419)
subject14_session02	49.5 min	~71,211	71,193	2026-11-07 14:19 (calib_20261107_1419)
subject15_session01	67.3 min	~96,928	96,890	2026-11-07 16:17 (calib_20261107_1617)
subject16_session01	100.0 min	~143,997	143,927	2026-11-08 20:22 (calib_20261108_2022)
subject16_session03	33.3 min	~47,843	47,825	2026-11-08 20:22 (calib_20261108_2022)

Layout

<session>/
  session.json                 capture settings, per-camera stats, clock anchor, EMG summary
  triggers.csv                 every trigger: seq, host_trigger_ns
  synced_frames.csv            one row per trigger: n_cameras + each camera's image path
  cam<i>_<serial>_timestamps.csv   one row per frame: frame_idx, image, trigger_seq,
                               device timestamp, device timestamp mapped to host clock, ...
  frames/cam<i>_<serial>/*.tar the JPEG frames, 10,000 per tar (see below)
  calibration/                 the calibration used for this session
    intrinsics/camera_<i>.npz  K, dist (OpenCV 5-coefficient), image_size, serial
    extrinsics/camera_poses.json   R, t per camera: X_cam = R @ X_world + t, metres,
                               world = cam0 frame
    calibration_summary.json   quality report
  emg.h5                       EMG on the host clock, with per-frame index
  emg_frame_index.csv          EMG sample range for every camera trigger
  timing_v2/                   corrected timing -- use instead of the above (see below)
  emg_frames.bin, emg_blocks.csv   raw EMG blocks as streamed
  intan_rhd/                   Intan's own .rhd recording

Corrected timing: timing_v2/ (use this)

Every session also has a timing_v2/ folder with refitted clock timing. Use it in place of the original timing files. The original post-pass mapped the Intan sample counter and each camera's clock onto host time with one straight line per session. When the host clock changed rate mid-session, the line could not follow, and EMG-to-video alignment drifted:

Worst EMG-to-video error in the original files	Sessions
17.5 ms	subject16_session01
7.0 ms	subject10_session01
4.5 ms	subject07_session02
3.5 ms	subject13_session02
1.5–2 ms	subject07_session01, subject12_session02, subject16_session03
≤ 1 ms	all others

timing_v2/ refits both clocks with a correction re-estimated every 60 s. The EMG and camera corrections, fitted independently, agree to ~0.3 ms rms.

• emg_frame_index.csv replaces emg_frame_index.csv.
• emg_timestamps.h5 (host_ns, timestamps, one row per row of emg.h5) replaces emg.h5's host_ns and timestamps. The EMG samples are unchanged.
• cam<i>_<serial>_timestamps.csv replaces the camera timestamp CSVs; only dev_ts_host_ns differs.
• timing_v2.json holds the fits and how much each quantity moved.

Restoring the frames

Each tar holds paths relative to the session folder (cam0_22865794/000/frame_000000.jpg), so extracting in the session folder restores the layout that synced_frames.csv and the timestamp CSVs refer to:

cd subject03_session01
for t in frames/*/*.tar; do tar -xf "$t"; done

Frames are stored as <camera>/<NNN>/frame_<frame_idx>.jpg (1,000 per folder). The trigger each frame answers is in the trigger_seq column; frames with the same trigger_seq were exposed together (median cross-camera spread ~0.15 ms).

Cameras

Index	Serial
cam0	22865794
cam1	22865834
cam2	22873121
cam3	22873159
cam4	24702671
cam5	24702677

Images are upright (cameras flipped in hardware).

Camera indices follow the cameras present in each session. Always use the serial in the folder name (cam<i>_<serial>), not the index, when matching across sessions: subject09_session02 has no 24702677 (cam0–cam4 as in the table above).

Known issues

• subject02_session01: cam4 (24702671) was physically moved at ~11:03 into the session (trigger 15912). Its calibration is valid only before that; drop cam4 afterwards.
• subject05_session01: by the next calibration (20:56) cam5 (24702677) had shifted relative to the other cameras (~8 px); when this happened could not be determined from the video. cam3 also moved, but only after the session ended.
• subject13_session02: cam0 (22865794) was bumped at trigger 59749 (2489.8 s, ~92 px shift); disregard cam0 frames with trigger_seq >= 59749. Details in subject13_session02/NOTES.md and exclusions.json. The other cameras did not move.
• subject09_session01: cam5 (24702677) was on a faulty USB port: ~560 missed triggers and some late exposures (up to ~20 ms), and the camera dropped out near the end. Check cam5_24702677_timestamps.csv (dev_ts_host_ns vs host_trigger_ns) and drop late cam5 frames.
• subject14_session01: cam2 (22873121) skipped 164 triggers and hung at 1407 s (missing for the last ~7 s); ~35 cam2 frames started more than 5 ms late (up to 21 ms). Drop late cam2 frames by comparing dev_ts_host_ns with host_trigger_ns in cam2_22873121_timestamps.csv.
• subject09_session02: recorded with 5 cameras (24702677 unplugged).
• A handful of frames per camera were lost to USB hiccups (listed as missing in synced_frames.csv, n_cameras < 6).