Voxel51/Egocentric_10K_subset

Dataset Card for Egocentric 10K (subset - Factory 51, first 51 videos)

[image]

This is a FiftyOne dataset with 416 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/Egocentric_10K_subset")

# Launch the App
session = fo.launch_app(dataset)

Here's a filled-out dataset card for your Factory 51 subset:

Dataset Details

Dataset Description

This is a curated subset of the Egocentric-10K dataset, focusing exclusively on Factory 51 with limited video sequences per worker. The subset contains egocentric video data captured from head-mounted cameras worn by factory workers during their daily tasks, providing first-person perspective footage of real manufacturing environments and hand-object interactions.

The subset includes the first 51 video clips (indices 0-50) from each worker in Factory 51, making it a more manageable dataset for research, development, and prototyping while maintaining the diversity of worker perspectives and temporal coverage.

• —Curated by: Build AI (original dataset)
• —Funded by: Build AI (original dataset)
• —Language(s) (NLP): N/A (video dataset, no speech/text)
• —License: Apache 2.0

Dataset Sources

• —Repository: https://huggingface.co/datasets/builddotai/Egocentric-10K (original dataset)

Uses

Direct Use

This dataset subset is suitable for:

• —Egocentric vision research: Developing and testing algorithms for first-person video understanding
• —Hand detection and tracking: Training models to detect and track hands in industrial environments
• —Action recognition: Recognizing manipulation actions and work activities in factory settings
• —Object interaction analysis: Understanding how workers interact with tools and materials
• —Temporal action segmentation: Segmenting continuous work activities into discrete actions
• —Prototyping and development: Testing computer vision pipelines on real-world industrial data with manageable dataset size
• —Educational purposes: Teaching egocentric vision concepts with authentic factory footage
• —Transfer learning: Pre-training or fine-tuning models for industrial or egocentric vision tasks

Out-of-Scope Use

This dataset should not be used for:

• —Worker surveillance or monitoring: The dataset is intended for research purposes, not for tracking individual worker productivity or behavior
• —Performance evaluation of individual workers: Videos should not be used to assess or compare worker performance
• —Biometric identification: The dataset should not be used to develop facial recognition or worker identification systems
• —Safety compliance enforcement: While useful for safety research, it should not be used punitively
• —Generalization to all factories: This is data from a single factory (Factory 51) and may not represent all manufacturing environments
• —Real-time production systems without validation: Models trained on this subset should be thoroughly validated before deployment

Dataset Structure

The dataset is organized as a FiftyOne video dataset with the following structure:

Fields

Each video sample contains:

• —filepath: Path to the MP4 video file
• —metadata: VideoMetadata object containing:
• —size_bytes: File size in bytes
• —mime_type: "video/mp4"
• —frame_width: 1920 pixels
• —frame_height: 1080 pixels
• —frame_rate: 30.0 fps
• —duration: Video duration in seconds
• —encoding_str: "h265" (H.265/HEVC codec)
• —worker_id: Unique identifier for the worker (e.g., "worker001", "worker002", etc.)
• —video_index: Sequential index of the video for that worker (0-50)
• —factory_id: "factory_051" (constant for this subset)

Statistics

• —Factory: 1 (Factory 51 only)
• —Workers: 8 workers (worker001 through worker008)
• —Videos per worker: Up to 51 (indices 0-51)
• —Total videos: 408 video clips
• —Resolution: 1080p (1920x1080)
• —Frame rate: 30 fps
• —Video codec: H.265/HEVC
• —Format: MP4
• —Field of view: 128° horizontal, 67° vertical
• —Camera type: Monocular head-mounted (Build AI Gen 1)
• —Audio: No

Dataset Creation

Curation Rationale

This subset was created to provide a more manageable version of the Egocentric-10K dataset for researchers and developers who:

• —Need a representative sample of factory egocentric video data
• —Have limited computational resources or storage capacity
• —Want to prototype and test algorithms before scaling to the full dataset
• —Require data from a single factory environment for controlled experiments
• —Need temporal coverage (51 sequential videos per worker) without the full dataset size

By limiting to Factory 51 and the first 51 videos per worker, this subset maintains:

• —Temporal diversity: Sequential videos capture different times and activities
• —Worker diversity: Multiple workers provide varied perspectives and work styles
• —Environmental consistency: Single factory reduces environmental variability
• —Manageable scale: Suitable for development and testing workflows

Source Data

Data Collection and Processing

Original Data Collection (by Build AI):

• —Videos captured using Build AI Gen 1 head-mounted cameras
• —Recorded in Factory 51 during normal work operations
• —Workers wore monocular cameras with 128° horizontal FOV
• —Captured at 1080p resolution, 30 fps
• —Encoded in H.265/HEVC for efficient storage
• —No audio recorded

Subset Curation Process:

1. 1.Downloaded Factory 51 data from Hugging Face: https://huggingface.co/datasets/builddotai/Egocentric-10K/tree/main/factory_051
2. 2.Extracted tar archives containing video and metadata pairs
3. 3.Filtered to retain only videos with video_index 0-50 (first 51 videos per worker)
4. 4.Deleted videos with video_index > 50
5. 5.Organized into FiftyOne dataset structure with metadata preservation

Recommendations

Users should:

• —Validate on diverse data: Test models on data from other factories, environments, and contexts before deployment
• —Consider ethical implications: Use data responsibly and avoid surveillance or punitive applications
• —Acknowledge limitations: Report the single-factory, limited-temporal nature of the subset in publications
• —Respect privacy: Implement additional privacy protections if sharing derived data or visualizations
• —Supplement with annotations: Consider adding task-specific annotations for supervised learning applications
• —Combine with other datasets: Use alongside other egocentric datasets (Ego4D, EPIC-KITCHENS, etc.) for robustness
• —Monitor for bias: Evaluate models for fairness across different worker characteristics and conditions

Citation

@dataset{buildaiegocentric10k2025,
  author = {Build AI},
  title = {Egocentric-10K},
  year = {2025},
  publisher = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/builddotai/Egocentric-10K}
}

APA:

Build AI. (2025). Egocentric-10K [Dataset]. Hugging Face Datasets. https://huggingface.co/datasets/builddotai/Egocentric-10K

More Information

For more information about the original Egocentric-10K dataset:

• —Dataset page: https://huggingface.co/datasets/builddotai/Egocentric-10K
• —Evaluation set: https://huggingface.co/datasets/builddotai/Egocentric-10K-Evaluation
• —Build AI: https://build.ai