cua-som 0.1.4

Computer Vision and OCR library for detecting and analyzing UI elements

Som (Set-of-Mark) is a visual grounding component for the Computer-Use Agent (Cua) framework powering Cua, for detecting and analyzing UI elements in screenshots. Optimized for macOS Silicon with Metal Performance Shaders (MPS), it combines YOLO-based icon detection with EasyOCR text recognition to provide comprehensive UI element analysis.

Features

• Optimized for Apple Silicon with MPS acceleration
• Icon detection using YOLO with multi-scale processing
• Text recognition using EasyOCR (GPU-accelerated)
• Automatic hardware detection (MPS → CUDA → CPU)
• Smart detection parameters tuned for UI elements
• Detailed visualization with numbered annotations
• Performance benchmarking tools

System Requirements

• Recommended: macOS with Apple Silicon
  • Uses Metal Performance Shaders (MPS)
  • Multi-scale detection enabled
  • ~0.4s average detection time
• Supported: Any Python 3.11+ environment
  • Falls back to CPU if no GPU available
  • Single-scale detection on CPU
  • ~1.3s average detection time

Installation

# Using PDM (recommended)
pdm install

# Using pip
pip install -e .

Quick Start

from som import OmniParser
from PIL import Image

# Initialize parser
parser = OmniParser()

# Process an image
image = Image.open("screenshot.png")
result = parser.parse(
    image,
    box_threshold=0.3,    # Confidence threshold
    iou_threshold=0.1,    # Overlap threshold
    use_ocr=True         # Enable text detection
)

# Access results
for elem in result.elements:
    if elem.type == "icon":
        print(f"Icon: confidence={elem.confidence:.3f}, bbox={elem.bbox.coordinates}")
    else:  # text
        print(f"Text: '{elem.content}', confidence={elem.confidence:.3f}")

Docs

• Configuration

Development

Test Data

• Place test screenshots in examples/test_data/
• Not tracked in git to keep repository size manageable
• Default test image: test_screen.png (1920x1080)

Running Tests

# Run benchmark with no OCR
python examples/omniparser_examples.py examples/test_data/test_screen.png --runs 5 --ocr none

# Run benchmark with OCR
python examples/omniparser_examples.py examples/test_data/test_screen.png --runs 5 --ocr easyocr

License

MIT License - See LICENSE file for details.