Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.32k
1.6k

Stereo + IMU Egocentric Capture

Technical Evaluation · Field Note #03

ORIGIN DATA LAB · Real-World Data Production for Physical AI

Physical AI Stereo IMU Evaluation

English Field Note · 한국어 기술노트 · GitHub evidence repository

Short answer. In a 120 s head-mounted stereo + IMU recording, all 3,600 / 3,600 stereo pairs shared the device frame tick, the device-recorded left/right exposure-centre offset had a median of 20.5 µs, and the IMU delivered ≈501.7 Hz with no gaps on the same clock as the video. Under auto-exposure, using the frame tick instead of the exposure centre biases frame time by about 5 ms.

Publication scope: technical evaluation, measurement figures and documentation only. The underlying raw stereo videos and IMU time series are not distributed here, and this is not a production calibration certificate or a robot-learning benchmark.

At a glance

Evaluated recording Stereo streams IMU Frame coverage
120 seconds 2 × 1600 × 1200 @ 30 fps 6-axis · ~501.7 Hz 3,600 frames per camera

A head-mounted stereo egocentric recording of an indoor two-handed workbench assembly task was evaluated for frame correspondence, timing, sensor continuity, image quality and calibration evidence. Measurements and limitations are presented separately to avoid conflating recorded evidence, estimates and unverified claims.

Evaluation summary

Metric / claim Observation Evidence status
Stereo frame tick correspondence 3,600 / 3,600 matched pairs Verified in evaluated record
Frame counter continuity No missing increments observed Verified in evaluated record
Visual motion alignment Peak at zero-frame lag; r = 0.997 Evaluated
Left/right exposure-center difference Median 20.5 µs; mean 21.1 ± 2.9 µs Device-reported; not independently verified
IMU sample rate Approximately 501.7 Hz Verified in evaluated record
IMU continuity No observed gaps, duplicates or saturation in tested range Verified in evaluated record
Camera–IMU timing offsets Left −1.1 ms; right −1.2 ms Estimate — consistent within a few ms only
Stereo calibration for evaluated capture mode Mode-matched Kalibr calibration: median angular epipolar error 0.41 px Verified (update, see Field Note #04)
Camera–IMU extrinsics Kalibr values exist; not verifiable from this recording Calibrated, not independently verified
Robot-learning performance Not tested Not assessed

Update, 23 October 2026. A Kalibr calibration produced in the same 1600 × 1200 mode was validated on this recording without adjustment (Field Note #04 · dataset card). The same target-free video–gyro method, applied to a recording with a Kalibr camera–IMU reference, missed the reference time shift by several milliseconds — so the −1.1 / −1.2 ms figures indicate alignment within a few milliseconds, not a measured offset.

Visual evidence gallery

01 · Stereo imagery and exposure timing

Stereo frame pair

Representative stereo image pair from the evaluated recording. Monitor screens are blurred for public release.

Stereo exposure timing

Timing analysis based on recorded camera metadata. Exposure-center differences are device-reported rather than externally certified.

02 · IMU signal quality and sampling

IMU raw signals

IMU interval and head rate

Inertial signals and sampling interval diagnostics for the evaluated segment.

03 · Visual–inertial temporal relationship

Visual rotation vs gyroscope

Camera IMU offset windows

Visual motion and gyroscope comparison. Camera–IMU offsets are diagnostic estimates, not certified hardware synchronization.

04 · Stereo rectification with the originally supplied (mode-mismatched) file

Rectification residual map

Rectification residual summary

These residuals come from a calibration file produced in a different (16:9) mode. A mode-matched calibration has since been validated at 0.41 px (Field Note #04).

05 · Frame-level image quality

Image quality per frame

Image-quality variation across the evaluated recording.

What is verified — and what is not

Supported by the evaluated record: video decoding, frame continuity, recorded stereo tick pairing, IMU continuity, diagnostic motion comparison and (since 23 Oct 2026) mode-matched stereo calibration fit.

Requires additional validation: independent optical stereo timing confirmation, metric depth accuracy, and independent camera–IMU extrinsic and time-offset verification. No downstream VLA/robot policy training performance is claimed.

See VALIDATION_STATUS.md for the full evidence matrix and TECHNICAL_REPORT.md for methods, interpretation and limitations.

Repository contents

Path Description
README.md Dataset card, measurement summary and evidence gallery
TECHNICAL_REPORT.md Technical evaluation methodology and findings
VALIDATION_STATUS.md Claim-by-claim validation boundaries
figures/*.webp Nine diagnostic figures from the published evaluation

Not included: source MP4 recordings, raw IMU logs, full per-frame sensor metadata, calibration parameters, training-ready annotations or dataset splits. The Dataset Viewer may not show sample rows because this is a documentation-only technical evidence release.

Cite

@misc{origindatalab2026fieldnote03,
  author       = {{Origin Data Lab}},
  title        = {Stereo + IMU Egocentric Capture: Measured Synchronization, Timing Model and Calibration Findings},
  howpublished = {Technical Field Note \#03},
  year         = {2026},
  url          = {https://origindatalab.io/technical-field-notes/stereo-imu-egocentric-sample.html}
}

Custom data production and evaluation pilots

Origin Data Lab supports scoped real-world data collection for robotics and Physical AI, including egocentric/stereo video, inertial metadata, field operations, quality review and engineering-oriented deliverables. Production specifications and acceptance criteria are agreed per project.

Request a technical evaluation or paid pilot · contact@origindatalab.io


Origin Data Lab · October 2026 · Text and figures: CC BY-NC 4.0 — attribution to Origin Data Lab.

Downloads last month
442