One per camera, passed to the follower. Each camera is sent as its own track and labelled separately in the dataset.

​
Fields

​
str
required
Which camera to open. The form depends on the operating system, see below.
​
str
What to call it: wrist, front, left. The dataset uses this name, as observation.images.<label>.
​
int
default:"640"
Capture width in pixels.
​
int
default:"480"
Capture height in pixels.
​
int
default:"30"
Capture rate. Frames are captured at this rate whatever the observation rate is, and the one nearest each observation is the one recorded.
​
str
default:"mjpeg"
Capture format. Linux only. On macOS the camera negotiates its own and this is ignored.
Keep pixel_format="mjpeg" on any host with more than one camera. Uncompressed video needs about 18 MB per second per camera, and two of those do not fit on one USB bus.

​
Finding a camera

The handle is whatever your platform’s capture backend understands.
That lists every camera and serial device with its by-path name, which is the handle to put in your config.
Use the by-path name, never /dev/video0 or /dev/ttyACM0. Those numbers change across reboots, so two identical cameras can swap places between runs.
Tired of long names? Register your own short ones once, then use those anywhere a handle is accepted.
Camera indexes and device names are not stable the way a by-path name is. On a machine with more than one camera, confirm which is which before a session rather than during one.

Follower

Where cameras are passed, and everything else in its constructor.

Observation

How a frame is paired with the joints from the same instant.