---
license: mit
task_categories:
  - video-classification
tags:
  - manual-labor
  - human-work
  - video-dataset
  - action-classification
  - physical-work
size_categories:
  - 1K<n<10K
language:
  - en
---

# Open Cortex FX v3

A curated dataset of videos depicting human manual labor and physical work, organized by task categories.

## Dataset Description

Standout's Cortex FX v3 is a video dataset focusing on human manual labor and physical work activities. Each video has been carefully annotated to identify work-related content and categorized into specific labor types.

### Dataset Statistics

- **Total Videos**: 708 videos showing human manual labor
- **Categories**: 28 distinct labor categories
- **Total Frames**: 204,329 frames across all videos
- **Average Duration**: 10.5 seconds (median: 6.4s)
- **Average Frame Count**: 289 frames per video (median: 170)
- **Frame Rate**: Most common FPS is 30.0 (388 videos), range: 15.0-30.0 FPS
- **Average Aesthetic Score**: 5.89 (out of 10)
- **Average Motion Score**: 7.64 (out of 10)
- **Format**: MP4 video files organized by category
- **Structure**: Videos are split into multiple zip files for easy distribution

#### Top Categories by Video Count:

- **Cooking**: 363 videos (51.3%)
- **Repair**: 100 videos (14.1%)
- **Automotive**: 58 videos (8.2%)
- **Gardening**: 28 videos (4.0%)
- **Construction**: 22 videos (3.1%)
- **Serving**: 19 videos (2.7%)
- **Assembly**: 18 videos (2.5%)
- **Crafting**: 17 videos (2.4%)
- **General Labor**: 14 videos (2.0%)
- **Sewing**: 12 videos (1.7%)

## Dataset Visualization

### Category Distribution

The dataset covers a diverse range of manual labor activities, with cooking being the most represented category:

![Category Distribution](category_distribution.png)

### Video Quality Metrics

All videos in the dataset have been evaluated for quality using three key metrics:

![Quality Metrics](quality_metrics.png)

- **Aesthetic Score**: Measures visual quality and composition (0-10 scale)
- **Motion Score**: Quantifies the amount and quality of motion in the video (0-10 scale)
- **Temporal Consistency**: Evaluates frame-to-frame coherence and stability (0-1 scale)

### Video Duration Distribution

The dataset contains videos of varying lengths, with most videos being short clips optimized for action recognition:

![Duration Distribution](duration_distribution.png)

### Frame Count and Frame Rate Distribution

The dataset includes videos with diverse frame counts and frame rates, providing rich temporal information for action recognition tasks:

![Frame and FPS Distribution](frame_fps_distribution.png)

- **Frame Count**: Videos range from short clips (~100 frames) to longer sequences (1000+ frames), with an average of 289 frames per video
- **Frame Rate**: Most videos are captured at 30 FPS (standard video rate), with some at 25 FPS and 23.976 FPS (cinematic rates)
- **Temporal Coverage**: The dataset provides over 200,000 total frames, offering substantial data for temporal modeling and action understanding

## Dataset Structure

The dataset is organized as follows:

```
final_0.zip
final_1.zip
final_2.zip
...
```

Each zip file contains:

- **Category folders**: Videos organized by labor type (e.g., `Cooking/`, `Repair/`, `Construction/`)
- **Metadata CSV**: Mapping file with video information

### Category Folders

Videos are organized into category folders based on the type of manual labor depicted:

- **Construction** - Building, construction work, using construction tools
- **Cooking** - Preparing food, cooking, chopping, mixing ingredients
- **Repair** - Fixing, repairing, maintenance work
- **Cleaning** - Cleaning, mopping, scrubbing, organizing spaces
- **Assembly** - Assembling products, putting things together
- **Gardening** - Gardening, landscaping, outdoor manual work
- **Painting** - Painting surfaces, applying paint or finishes
- **Sewing** - Sewing, textile work, fabric manipulation
- **Woodworking** - Working with wood, carpentry, wood crafting
- **Metalworking** - Working with metal, welding, metal fabrication
- **Moving** - Lifting, carrying, moving heavy objects
- **Serving** - Serving food or drinks, manual service work
- **Organizing** - Arranging, organizing physical items
- **Crafting** - General crafting, artisan work, making things by hand
- **Electrical** - Electrical work, wiring, electrical installation
- **Plumbing** - Plumbing work, pipe installation, water systems
- **Automotive** - Automotive repair, working on vehicles
- **Farming** - Farming, agriculture, working with crops/animals
- **Welding** - Welding, joining metals with heat
- **General Labor** - Other manual labor that doesn't fit above categories

## Metadata Format

Each zip file includes a `metadata.csv` file with the following columns:

| Column     | Description                               |
| ---------- | ----------------------------------------- |
| `name`     | Video filename (e.g., `aBc123XyZ789.mp4`) |
| `category` | Labor category                            |

## Usage

**Note:** The dataset viewer may not be available due to the zip file structure. However, the dataset can be loaded programmatically using the code examples below.

### Downloading the Dataset

The dataset is available on Hugging Face. You can download it using:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Standout/open-cortex-fx-v3",
    repo_type="dataset",
    local_dir="./open-cortex-fx-v3"
)
```

### Extracting and Using the Data

1. **Extract zip files**: Unzip the downloaded files to access the video content
2. **Read metadata**: Load `metadata.csv` from each zip to get video information
3. **Access by category**: Navigate to category folders to find specific types of labor videos

### Example: Loading Metadata

```python
import zipfile
import pandas as pd
import io

# Extract and read metadata from a zip file
with zipfile.ZipFile('final_0.zip', 'r') as z:
    with z.open('metadata.csv') as f:
        metadata = pd.read_csv(io.BytesIO(f.read()))

# Filter by category
cooking_videos = metadata[metadata['category'] == 'Cooking']
print(f"Found {len(cooking_videos)} cooking videos")

# Access video names
for video_name in cooking_videos['name']:
    print(f"Video: {video_name}")
```

### Example: Accessing Videos

```python
import zipfile

# Extract a specific video
with zipfile.ZipFile('final_0.zip', 'r') as z:
    # List all videos in Cooking category
    cooking_videos = [name for name in z.namelist() if name.startswith('Cooking/')]

    # Extract a video
    z.extract('Cooking/example_video.mp4', './output/')
```

## Dataset Characteristics

- **Focus**: Human manual labor and physical work activities
- **Quality**: Videos are filtered to show clear human action performing manual tasks
- **Diversity**: Multiple categories covering various types of physical work
- **Organization**: Structured by labor type for easy access and filtering

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{open_cortex_fx_v3,
  title={Open Cortex FX v3: A Classified Dataset of Human Manual Labor},
  author={Standout},
  year={2024},
  url={https://huggingface.co/datasets/Standout/open-cortex-fx-v3}
}
```

## License

Apache 2.0

## Contact

For questions or issues, please contact standout@standout.work.