Motion data axes¶

Trinet cameras record motion data in the motion sensor's own frame, exactly as the sensor measures it. The camera does not rotate or re-map the axes. Because the sensor sits differently on each camera's board, the direction of x, y and z relative to the lens depends on the camera model. The calibration's camera–IMU rotation is what links the two frames, and this page lists it for every model.

Conventions¶


Frame	Right-handed
Accelerometer	m/s², gravity included. At rest, the axis pointing up reads about +9.81
Gyroscope	rad/s. Positive is counter-clockwise about the axis, seen from its tip (right-hand rule)
Magnetometer	µT, in its own frame (see below)
Same frame for	`.imu` files, the motion data inside the video file, the live USB stream, the SDK and the MCAP `/imu` topic

Camera frame, used in the table below: as seen from behind the camera, looking where the lens looks, x points right, y down and z forward out of the lens. It's the image frame: x along the image rows, y down the image. On stereo cameras it is the left eye (_L, cam0).

Axes per camera¶

Where each motion-sensor axis points, in camera terms:

Camera	Hardware version	IMU +x	IMU +y	IMU +z	Lens up, at rest
Trinet Mono (original)	V1, V2	up	left	backward	z ≈ −9.8
Trinet Mono	V3	left	up	forward	z ≈ +9.8
Trinet Mono	V4	up	right	forward	z ≈ +9.8
Trinet Stereo	V5	down	left	forward	z ≈ +9.8
Trinet Stereo GS	V6	right	down	forward	z ≈ +9.8

The same as rotation matrices, R_cam_imu, which maps a vector from the IMU frame into the camera frame (v_cam = R_cam_imu · v_imu):

Camera	`R_cam_imu` (rounded)
Mono V1, V2	`[[0,-1,0], [-1,0,0], [0,0,-1]]`
Mono V3	`[[-1,0,0], [0,-1,0], [0,0,1]]`
Mono V4	`[[0,1,0], [-1,0,0], [0,0,1]]`
Stereo V5	`[[0,-1,0], [1,0,0], [0,0,1]]`
Stereo GS V6	`[[1,0,0], [0,1,0], [0,0,1]]`

These are the rounded rotations of the reference calibrations. The true mounting is within a few degrees of them. For exact work, use the rotation from the calibration itself. Stereo and Stereo GS cameras carry their own per-unit calibration; see Get your camera's calibration. Not sure which hardware version you have? See Which camera do I have?

Quick check

Record a few seconds with the camera on a table, lens pointing at the ceiling, and read the accelerometer with the toolkit. z reads about +9.8 m/s² on V3–V6, and about −9.8 on V1/V2.

Rotate motion data into the camera frame¶

With the exact rotation from the camera's calibration:

import json
import numpy as np
from trinet_tools.reader import read_imu

calib = json.load(open("calibration.json"))
T = (calib.get("T_cam0_imu")                                 # stereo
     or calib.get("T_cam_imu")                               # mono, reference file
     or calib["extrinsics"]["T_cam_imu"])                    # mono, read from the camera
R_cam_imu = np.asarray(T)[:3, :3]

imu = read_imu("take0004.imu")
accel_cam = np.asarray(imu.accel) @ R_cam_imu.T              # m/s², camera frame
gyro_cam  = np.asarray(imu.gyro)  @ R_cam_imu.T              # rad/s, camera frame

The calibration also gives the lever arm — where the motion sensor sits relative to the camera — and the small camera–IMU time offset. See Use a calibration. VIO and SLAM tools such as OpenVINS, Kalibr, VINS-Fusion and Basalt take the IMU data in the IMU frame together with this transform, so you normally pass both unchanged (export).

Magnetometer¶

The magnetometer (Mono V3 and later, Stereo and Stereo GS) is a separate sensor, recorded in its own frame. That frame differs from the accelerometer/gyroscope frame, and its readings are reported with the opposite sign to the field direction. To bring it into the IMU frame, apply m_imu = R_imu_mag · m:

Camera	`R_imu_mag`	In words
Mono V4	`[[0,1,0], [-1,0,0], [0,0,-1]]`	x = +m_y, y = −m_x, z = −m_z
Stereo V5	`[[0,-1,0], [-1,0,0], [0,0,1]]`	x = −m_y, y = −m_x, z = +m_z
Stereo GS V6	`[[-1,0,0], [0,-1,0], [0,0,-1]]`	x = −m_x, y = −m_y, z = −m_z

These mappings were measured from recordings: the gyroscope's rotation was matched against the change in the magnetic field, over several units and recordings per model, and the sign was set by Earth's field direction. They are reliable for V4 and V5. Treat V6 as provisional: the fit there is weaker. Mono V3 is not yet measured. Mono V1/V2 have no live magnetometer, so the field reads zero.

The magnetometer is not calibrated for hard- and soft-iron effects, so readings include fixed offsets from the camera's own electronics. Calibrate them out, for example with a sphere or ellipsoid fit over a recording that turns the camera through many orientations, before you use the data for heading.