Hugging Face's logo

Datasets:
fpvlabs
/
stereo-550

Tasks:
Robotics
Video Classification
Image-to-Text
Modalities:
Video
Languages:
English
Size:
1K<n<10K
ArXiv:
Tags:
egocentric
first-person
stereo
stereo-calibration
vision-language-action
vla
License:
Dataset card Files
xet
 Community

Request controlled access to Stereo-550

By agreeing you accept to share your contact information (email and username) with the repository authors.

Log in or Sign Up to review the conditions and access this dataset content.

Stereo-550

Paper · Code · Build it yourself · 3D viewer · Blog

Collected with FPV Labs Open-Source Stereo Hardware

Dataset overview

A first-person calibrated stereo RGB video dataset capturing everyday human manipulation across objects, materials, tools, and multi-step activities. Every session is recorded as a synchronized left/right camera pair with per-session stereo calibration, giving the visual geometry of hands, object interaction, state change, and task progression, signals directly relevant to embodied perception, video-language learning, and human-to-robot representation research.

The dataset's strongest differentiator is dense, free-form action captioning at scale. Every recording is segmented into second-scale spans, each carrying a descriptive natural-language caption (e.g. "place small metal cup on kitchen counter", "peel garlic clove", "cut sewing thread with scissors"). Sessions were selected for hand visibility, so the corpus is consistently rich in hand-object interaction, combining repeated coverage of foundational skills with a long tail of rare task expressions.

  • ≈550 hours of synchronized stereo recording per camera (≈1,100 stereo camera-hours)
  • 1,462 stereo sessions (left + right = 2,924 video files), each with per-session stereo calibration
  • 209,315 labeled action segments covering ~100% of the recorded timeline
  • 460 action verbs and 32,630 object phrases, 57,104 distinct verb-object combinations
  • Top 20 expressions = only 1.5% of all instances, a genuine long tail for open-world learning
  • A median of 94 labeled segments per session; 95.8% of sessions show 10+ distinct task expressions
  • Per-session 6-axis IMU synchronized to video in 1,271 of 1,462 sessions (86.9%)

At a glance

Dimension Dataset evidence User value
Video scale ≈550 h / camera (≈1,100 stereo cam-h) Large calibrated stereo RGB corpus
Capture geometry Synchronized pair + per-session calibration Metric binocular depth cues, not just RGB
Delivery format MP4 (H.264, 1280×720, 30 fps) Straightforward ingestion into video pipelines
Label density 209,315 segments; median 94/session Dense temporal supervision, not clip-level tags
Action vocabulary 460 observed verbs Strong coverage of manipulation primitives
Object vocabulary 32,630 object phrases Wide object, material and tool interaction coverage
Effective breadth 57,104 verb-object combinations Compositional, resistant to rare-label inflation
Temporal structure 192,509 ordered task transitions Rich sequence structure for world models

Activity domain coverage

High-level activity families derived from the caption vocabulary (keyword-based, approximate). "Labeled hours" is the recorded time spent in each family; "primary in" counts the sessions where that family accounts for the most labeled time. Because captions are free-form, most sessions touch several families.

Activity domain Labeled hours Primary activity in
Cooking and food preparation 187 h 664 sessions
Dishwashing and kitchen cleanup 90 h 258 sessions
Textile and craft (sewing, tailoring, flowers) 54 h 214 sessions
Laundry and clothing care 45 h 145 sessions
Organizing and storage 39 h 87 sessions
Cleaning and housekeeping 29 h 87 sessions
Generic manipulation and transitions 106 h 7 sessions

The dataset is kitchen-centric: cooking and dishwashing together account for the largest share of labeled time and sessions, with meaningful secondary coverage of textile and craft work, laundry and clothing care, and general home cleaning and organizing.

Session length

session_length

Sessions range from short focused tasks to hour-plus continuous activity (median 14.9 min, longest 211 min), giving both clean short episodes and long-horizon sequences from the same capture pipeline.

Dataset structure & modalities

Each session directory (<user_id>/<session_id>/) is self-contained: left video, right video, action labels, calibration.json, and (where available) a synchronized IMU stream.

<user_id>/
└── <session_id>/                       # e.g. stera006_8zlo11pxcsqd
    ├── <session_id>_left.mp4           # left camera, 1280×720 @ 30 fps, H.264
    ├── <session_id>_right.mp4          # right camera, 1280×720 @ 30 fps, H.264
    ├── <session_id>_action_labels.json # per-segment action captions
    ├── calibration.json                # per-session stereo camera calibration
    └── imu/
        └── <session_id>/
            └── imu_synced.csv          # 6-axis IMU, synced to video (~180 Hz)

Modalities:

  • Stereo RGB video: synchronized left/right pair, 1280×720 @ 30 fps, H.264 (yuv420p).
  • Stereo calibration (JSON): per session; pinhole + radtan intrinsics/distortion and stereo extrinsics (baseline ≈42 mm).
  • Action labels (JSON): video_id + ordered segments, each with start_time, end_time, and a free-form text caption. Uniform schema across all sessions (no extra keys).
  • IMU (CSV): 6-axis inertial data (ax/ay/az, gx/gy/gz) synced to video at ~180 Hz; present in 1,271 of 1,462 sessions.

Stereo camera calibration

Each session ships a per-session calibration.json (pinhole camera model with radial-tangential distortion), enabling metric, geometry-aware use of the stereo pair:

  • Per-camera intrinsics: left/right camera_matrix and fx, fy, cx, cy, with 5 distortion coefficients (k1, k2, p1, p2, k3).
  • Stereo extrinsics: 3×3 rotation, translation, and a baseline of ≈42 mm between the two cameras.

What stands out

  • Calibrated stereo, not mono: synchronized left/right pair with per-session calibration, metric binocular geometry.
  • Foundation-scale volume: ≈550 labeled hours per camera across 1,462 stereo sessions.
  • Dense supervision: 209,315 action segments covering ~100% of the timeline; median 94/session.
  • Exceptional breadth: 460 verbs, 32,630 object phrases, 57,104 verb-object combinations.
  • Genuine long tail: top-20 expressions = only 1.5% of instances.
  • Hand-verified and multimodal: every session hand-visibility screened; 6-axis IMU synced to video; sessions up to 211 min.
  • Ready-to-ingest: standard MP4 + JSON + CSV, one directory per session.

Access & Licensing

Stereo-550 is released by FPV Labs for research use under the FPV Labs dataset license. Please read the complete license before use.

Contact

FPV Labs, contact@fpvlabs.ai
Downloads last month
299
Total file size:
1.86 TB

Paper for fpvlabs/stereo-550