Stereo Egocentric Video + 501.7 Hz IMU: Real-World Sensor Evaluation

Stereo Egocentric Video + 501.7 Hz IMU: What We Validated—and What We Did Not

An engineering field note on real-world first-person manipulation data for Physical AI and robotics.

Capturing two camera streams and high-rate inertial measurements does not automatically produce a calibrated robotics dataset. Frame correspondence, timestamp integrity, inertial continuity, temporal alignment, and stereo geometry each require separate evidence.

At Origin Data Lab, we evaluated a 120-second head-mounted stereo recording of an indoor, two-handed case-assembly task involving a power drill and hand tools. This post summarizes the engineering observations and their limitations.

1. Real-world capture configuration

• Stereo: 1600 × 1200 per eye, 30 fps, 3,600 left–right frame pairs
• IMU: six-axis accelerometer and gyroscope, approximately 501.7 Hz
• Scene: first-person, indoor workbench assembly
• Timing: device frame timestamps and per-frame exposure metadata

Figure 1. Left/right stereo frame pair from the evaluated recording. The monitor was blurred for the public-release image.

2. Video quality and frame correspondence

Both 120-second streams decoded as 3,600 frames with no decoding errors in this evaluation. The left and right device frame ticks matched across all 3,600 pairs. We also evaluated image quality indicators across the recording rather than relying on a single attractive example frame.

Figure 2. Per-frame image quality measurements for the evaluation segment.

3. Stereo exposure timing

The two streams share device frame ticks, while exposure start and duration are recorded independently. The device-reported median difference between left and right exposure centers was 20.5 microseconds. This is a device-recorded measurement, not an independently certified optical synchronization accuracy.

Figure 3. Stereo exposure timing assessment from recorded device metadata.

4. IMU continuity and motion signals

The 120-second segment contained 60,242 paired accelerometer/gyroscope samples at an effective rate of 501.71 Hz. The observed sample intervals ranged from approximately 1.87 to 2.11 ms, with no missing or duplicate samples identified by the checks applied to this recording.

Figure 4. Measured accelerometer and gyroscope signals.

Figure 5. IMU sampling intervals and observed head-motion distribution.

These are results for this recording, not a blanket guarantee for every device or production session.

5. Camera–IMU alignment: an estimate, not certification

We compared visual rotation estimates with the gyroscope signal and estimated camera–IMU temporal offsets of approximately −1.1 ms (left) and −1.2 ms (right) under the report's sign convention. The fit was encouraging, but method resolution, camera modeling, and independent timing verification remain limitations.

Figure 6. Visual motion estimate compared with gyroscope measurements.

Figure 7. Windowed temporal-offset estimates; values are estimates rather than hardware-level timing certification.

6. Stereo calibration: a measurable limitation

The manufacturer-supplied calibration parameters were associated with a different sensor mode from this recording. When applied to the evaluated 1600 × 1200 mode, the independent validation subset showed a median vertical rectification residual of 3.409 pixels. That is evidence of a mode mismatch—not a production-grade calibrated stereo result.

Figure 8. Spatial distribution of vertical rectification residuals under the supplied parameters.

Figure 9. Residual summary illustrating why mode-specific calibration is still required.

For metric depth, 3D hand trajectories, or precision visual-inertial work, calibration must be performed and independently validated for the actual recording mode.

7. Why this matters for Physical AI

A useful engineering dataset should distinguish between verified integrity, estimated alignment, and calibration work that remains open. The 120-second sample supports investigation of stereo video, sensor timing, IMU continuity, and processing pipelines. It does not by itself establish metric-depth accuracy, certified camera–IMU synchronization, or broad generalization across tasks and environments.

Explore the public technical release

• Technical Field Note #03: https://origindatalab.io/technical-field-notes/stereo-imu-egocentric-sample.html
• GitHub documentation: https://github.com/origin-data-lab/stereo-imu-egocentric-evaluation
• Hugging Face evaluation materials: https://huggingface.co/datasets/origindatalab/stereo-imu-egocentric-evaluation

The public materials are technical documentation and supporting figures; the original video and raw IMU data are not included in the public release.

Origin Data Lab produces custom real-world data for robotics, embodied AI, and Physical AI projects, with field capture, metadata, QA/QC, and engineering-oriented delivery.

Discuss a custom collection or paid evaluation pilot: https://origindatalab.io/contact.html