跳转至

4. Capture workflow

Back to contents · 中文

Ten tasks, one rig, one calibration protocol. Top row: six laboratory procedures; bottom row: four household and industrial tasks.

4.1 A capture day

  1. Wear the rig and run the checks (1.4 and 2.5).
  2. Record calibration. Every donning starts with the 01–17 calibration sequence (chapter 5); one or two spare takes are a good idea. Calibration recordings must be recognisable by name, for example with the task name "calibration".
  3. Record the task list (chapter 3). Pick the task on the page, r to start, e to save; d discards a bad take.
  4. Spot-check. Replay a few recordings every day with visualize_session.py (4.4).
  5. Upload (optional, 3.6).
  6. Keep a log. The capture program writes recording_log_<date>.csv next to the data folder with each recording's start and end time, duration and frame counts. Also note by hand who made each recording and in which donning: processing groups recordings by exactly that, one calibration per group (7.3).

4.2 Naming

Each recording is a folder named session_<date>_<time>_<task>, for example session_20260901_154755_物料称量. It holds five files, described in chapter 6:

data/session_20260901_154755_物料称量/
├── session_20260901_154755_物料称量.mp4
├── session_20260901_154755_物料称量.video.npz
├── session_20260901_154755_物料称量.npz
├── session_20260901_154755_物料称量.tracker.npz
└── session_20260901_154755_物料称量.session.json

Recordings that deliberately capture interruptions and recovery get _恢复 ("recovery") after the task name, for example session_20260901_164121_过柱_恢复.

4.3 Recording rules

  • Stand still for 2–3 seconds before starting and hold 1–2 seconds at the end. The side checks and the projection both use the opening stance.
  • Do not touch trackers or straps while recording. If one gets knocked, note it in the log and recalibrate if needed.
  • Make sure the wrist trackers move. A recording without wrist tracker data cannot be reconstructed.
  • Keep recordings a manageable length. Laboratory tasks usually take 2–7 minutes; split longer ones.
  • Mind the top of the frame. The GoPro's HDMI output carries its on-screen status (battery, SD card and so on) along the top, and the release crops the top 110 pixels. Keep important content away from the very top.

4.4 Checking a recording

Start with session.json: video.queue_dropped_frames, read_failures, writer_failures and tracker.dropped_frames should all be 0, and overlap_s should be close to the recording length.

Then replay it:

python visualize_session.py data/session_YYYYMMDD_HHMMSS_<task>
# optional: --no-open --port 8000 --max-fps 30 --build-only

The replay page detects the glove type and puts video, the five trackers, both hand skeletons and touch on one timeline (the page is shown in 3.4). What to look for:

Check Normal
Trackers All five have coordinates and move, and turn together when the wearer turns
Hand skeletons Both move; the hand moving in the video moves on its own side
Touch (Wuji) Lights up during grasps; large areas that stay on or stay dark point to a glove or calibration problem
Duration and rates Match the recording