egocentric-stereo-rgbd / docs /DATA_CARD.md
tim404x's picture
Swap Stereo #01/#03 identities and synchronize metadata
370d3bf verified
|
Raw
History Blame Contribute Delete
6.76 kB

Hub Stereo RGB-D Sample Delivery · Data Card

Hub Data. July 2026. Format contract: README.md (same folder).

1. What this is

An egocentric (head-mounted) stereo RGB-D corpus of a human performing everyday manipulation and locomotion tasks in indoor and outdoor service environments, with dense machine-generated metric depth, per-frame IMU, 6-DoF visual-inertial pose, and stereo calibration. This is a human demonstration corpus; there are no robot actions, joint states, or teleoperation labels. It is intended for representation, depth, and world-model pretraining, for stereo-depth and egocentric-perception evaluation, and as egocentric human-manipulation demonstration data.

  • 9 clips, 29,758 frames, ~12 minutes, 1920×1200 per eye, mixed ~30 fps and 60 fps capture.

2. Per-clip contents

# Clip Frames Duration What's demonstrated
01 Restaurant · Chop Vegetables 1,800 60 s Chopping vegetables: fast periodic knife/tool motion in the near field.
02 Restaurant · Arrange Table 4,800 163 s Setting a table: plates, cutlery, repeated reach-place cycles; longest clip.
03 Restaurant · Arrange Glasses 1,920 66 s Arranging drinking glasses on a counter: close-range bimanual manipulation of transparent objects (a classically hard stereo case).
04 Restaurant · Wash Dishes 1,680 57 s Washing dishes at a sink: running water, wet specular surfaces, frequent very-close-range activity (highest near-field saturation).
05 Meeting Room · Arrange Chairs 3,627 60 s Moving and arranging chairs: room-scale motion, large low-texture walls. 60 fps.
06 Office · Photocopy 4,499 75 s Operating a photocopier: device interaction, uniform surfaces. 60 fps.
07 Open Space · Rack the Ball 3,967 66 s Throwing and racking balls: the extreme-motion case (fast small objects; see §5). 60 fps.
08 Outdoor · Sweeping 3,000 101 s Sweeping an outdoor seating area: natural light, long-handle tool use, ground-plane geometry.
09 Pantry · Wash Tea Set 4,465 74 s Washing a tea set: small-object manipulation at a sink. 60 fps; best temporal stability in the set.

Total: 29,758 frames across the 9 clips.

3. Collection hardware

  • Head-mounted global-shutter stereo camera (StereoLabs ZED X Mini): 1920×1200 per eye, 49.88 mm baseline, factory-rectified (D=0, R=I). Four clips captured at 60 fps, the rest at ~29-30 fps.
  • Built-in IMU, delivered frame-locked to the recording (so ~30-60 Hz here, i.e. at the video rate, not the sensor's raw high rate); on-device 6-DoF visual-inertial poses.
  • Delivered per clip as a Foxglove Data.mcap (stereo/mono RGB, IMU, VIO pose, camera calibration, static transforms), the side-by-side stereo MP4, a mono MP4, a colorized Hub metric-depth preview, and a self-contained LeRobot v3.0 dataset (LeRobot_v3.0/).

4. Depth provenance

  1. Frames are extracted from the rectified stereo pair; frame counts are asserted at every stage.
  2. Dense per-pixel depth is computed from the stereo pair using Hub's stereo depth model (July 2026 build), then temporally smoothed for frame-to-frame stability. Depth is metric via the calibrated conversion Z = f · B / disparity (f = 754.949 px rectified, B = 49.88 mm).
  3. Metric depth is delivered natively in each clip's LeRobot_v3.0/ dataset (mm on lerobot ≥ 0.6.0). The colorized Depth.mp4 is a preview only.
  4. The rectified calibration used for the depth conversion is cross-checked against the as-recorded sensor calibration; the two differ by at most 0.21 px in focal length and 0.0033 mm in baseline, confirming the rectification (full chain in each clip's LeRobot_v3.0/meta/calibration.json).

5. Accuracy and limitations (honest)

  • Coverage. Depth is dense: every clip retains ≥ 99.996% valid pixels, and unresolved (code-0) pixels are a negligible fraction. Density is not the same as reliability: the near-field and transparent/specular regions below are dense but low-confidence.
  • Metric basis. Depth is geometrically metric: it derives from the validated stereo calibration and the disparity solve. In a separate controlled bench (a fiducial target of known geometry), the same depth model measured ~0.93% RMS error. That figure is a bench reference, not an on-clip measurement: the delivered clips have no fiducial ground truth, so on-clip metric accuracy is not independently validated. Treat depth as metrically-scaled and geometrically consistent, and validate against your own reference if you need a certified number.
  • Far-range precision. This is a short 49.88 mm baseline, so metric precision degrades with range, roughly quadratically (about one disparity-pixel ≈ 8% depth error near 3 m, and coarser toward the 10 m clip). Near and table-scale depth is the strong regime; treat room-scale and outdoor mid and far depth as scale-correct but low-precision.
  • Near field (< 0.25 m). On close-range clips (02, 03, 04), roughly 10-16% of pixels fall inside the ~0.25 m near band, dominated by self-occlusion, the wearer's own body, and shrinking left/right stereo overlap rather than raw resolvability. They are reported at their computed values, not masked; treat them as low-confidence.
  • Invalid pixels. Unresolved pixels (disparity < 1 px; rare) are encoded as code 0, which decodes to the 0.01 m (10 mm) floor rather than NaN. Drop depth ≤ 0.01 m before training or metric use.
  • Extreme motion (clip 07). On the fastest-motion clip (thrown balls), brief whole-scene depth re-grades can appear on a small fraction of frames (~3%). Disclosed in the clip card.
  • Temporal variation (clip 03, Arrange Glasses). This clip has greater capture-motion/depth temporal variation than clip 01 (Chop Vegetables). Sequential review found no alternating-frame, freeze, blank, or decode defect. The media is retained unchanged for temporal-variation evaluation.
  • Edge softening. Temporal smoothing trades a small amount of edge sharpness at fast motion boundaries for frame-to-frame stability.
  • Transparent and specular surfaces (glassware, wet surfaces) produce locally unreliable depth, as they do for any stereo method.

6. Confidentiality

Shipped artifacts carry no capture-partner identity, operator IDs, or environment IDs. The recordings retain the camera unit's serial number and the native capture timestamps as ordinary sensor provenance (as Aria-class and DROID-class releases do); these identify the physical camera and the recording, not any partner or subject.

7. Terms

This sample is provided for evaluation. Contact Hub Data for data-use terms.