origin-data-lab/egocentric-motion-blur-gyro-vs-sharpness

Egocentric Motion Blur — Gyro-Predicted Blur vs Measured Sharpness

Dataset Summary

Per-frame image-quality and motion measurements for 7,557 egocentric frames: 3,600 from a head-mounted wide-angle stereo rig with per-frame exposure times and gyro-predicted blur, and 3,957 from two body-worn phone recordings of cooking tasks with camera rotation rates. It accompanies Origin Data Lab Technical Field Note #06. All values were re-derived from the source recordings for this release (end-to-end re-run) and match the 120 values published in the first version of Field Note #06.

Intended Uses

• Studying how camera rotation and exposure time degrade egocentric frames.
• Building or testing IMU-based blur flags for capture QC.
• Reproducing the field note's numbers. Not intended as an image-quality benchmark for any specific model.

Data Structure

Splits: stereo_rig (3,600 rows), phone (3,957 rows; recording distinguishes the two phone sessions).

stereo_rig

Column	Description
recording	stereo_rig_left
frame	0-based frame index
time_s	Exposure-centre time from device timestamps, s
exposure_ms	Per-frame exposure time reported by the device, ms
sharpness_laplacian_var	Variance of the Laplacian, centre crop rows 300–900 × cols 400–1200
fast_corners	FAST corners (threshold 20), same crop
blur_px_pred	Gyro-predicted blur at the image centre during exposure, px
rotation_dps	Mean camera rotation rate during exposure, °/s
rel_sharpness	sharpness / median of frames with blur_px_pred < 0.35 within ±90 frames
rel_fast_corners	fast_corners normalised the same way

phone

Column	Description
recording	phone_knife_prep or phone_seasoning
frame	0-based frame index
time_s	Container PTS, s
sharpness_laplacian_var	Variance of the Laplacian, centre crop rows 270–810 × cols 480–1440
fast_corners	FAST corners (threshold 20), same crop
rotation_dps	Camera rotation rate, gyro averaged ±16.7 ms around the frame after +188/+200 ms video–IMU alignment, °/s; empty if fewer than 4 gyro samples
rel_sharpness	sharpness / median of frames with rotation_dps < 5 within ±90 frames; empty if fewer than 6 reference frames
rel_fast_corners	fast_corners normalised the same way
non_task	1 if the frame lies in a human-annotated non-task keystep (Field Note #05)

Collection Process

Recordings were captured by Origin Data Lab: a head-mounted stereo + IMU rig (left camera used) and a body-worn Android phone. Values were computed by reproduce_fn06.py from the source video, frame timestamps and IMU files.

Quality Control

• Full pipeline re-run from source recordings; 120 published values matched exactly.
• Public CSVs reproduce all 273 aggregate values in results_fn06.json (script summarize_from_csv.py).
• No duplicate frames; times strictly increasing; exposure 8.075–12.0 ms; no negative values.
• 113 phone frames have empty rel_sharpness (no low-motion reference within ±3 s).

Visual results

Predicted motion blur vs relative sharpness

Camera rotation rate vs relative sharpness

Phone recording: motion and sharpness timeline

Example: sharper and motion-blurred frames

Key Results

• Gyro → image map fit: R² = 0.974
• Spearman ρ (predicted blur vs relative sharpness): -0.62
• Gyro-only flag for frames that lost >40% sharpness: AUC 0.92 (rig), 0.88 / 0.93 (phones)
• Frames above 20 °/s: 6.8% (rig), 13.2% / 21.6% (phones)

Limitations

Camera-motion blur only; Laplacian variance and FAST corners on H.264 video are proxies; three recordings with different devices, tasks and operators; rolling shutter and translation not modelled; phone exposure not recorded.

Licensing

Code: MIT License (LICENSE), Origin Data Lab’s own analysis code.

Data (per-frame statistics CSV) and documentation: CC BY 4.0 is prepared (LICENSE-DATA-CC-BY-4.0.md) but CC BY 4.0 permits commercial reuse.

Personal and Sensitive Information

The files contain only numeric per-frame measurements and a non-task flag. No images, video, audio, faces, locations or personal identifiers. Source videos and manufacturer calibration files are not distributed.

Citation

Origin Data Lab (2026). How Head Rotation Blurs Egocentric Video. Technical Field Note #06. https://origindatalab.io/technical-field-notes/motion-blur-egocentric-video.html

Public research resources

• GitHub reproducibility repository
• Hugging Face dataset
• Kaggle dataset
• Executed Kaggle notebook