Skip to content

README.md

Latest commit

 

History

History
103 lines (72 loc) · 4 KB

README.md

File metadata and controls

103 lines (72 loc) · 4 KB
Raw

PaperCAD Edge: Intelligent Sketch-to-CAD Converter

Qualcomm Edge AI Hackathon Project

PaperCAD Edge is a powerful, local-first application that transforms both hand-drawn sketches and professional floor plans into clean, editable, and intelligent CAD models. By leveraging NPU acceleration, PaperCAD Edge delivers real-time performance, making it ideal for on-the-go design and digitization.

👥 Team Members

  • Nishanth Kotla - nk3968
  • Shreyas Bhaktharam
  • Rivujit Das
  • Anurup Naskar

🚀 Key Use Cases

  • Real-Time Conversion: Use your device's camera to capture a live image of a sketch and watch it instantly become a clean CAD model on your screen.
  • Hand-Drawn Sketch Digitization: Transform rough, back-of-the-napkin sketches into geometrically perfect and editable DXF files. Our "Geometric Intelligence" engine cleans up imperfections and understands your design intent.
  • Professional Floor Plan Processing: Upload existing high-quality floor plans or scanned drawings to quickly convert them into layered, parametric CAD models.

✨ Features

  • Multi-Input Support: Processes PNGs, JPEGs, and live camera feeds.
  • Geometric Intelligence: Goes beyond simple tracing to infer constraints like perpendicularity and parallelism, cleaning up rough drawings.
  • AI-Powered Line & Symbol Detection: Uses a pipeline of advanced models to accurately identify walls, doors, windows, and other symbols.
  • Text & Dimension Recognition: Automatically reads text labels from drawings using OCR (Optical Character Recognition), preparing the geometry for clean, scaled export.
  • 3D Generation: Instantly extrudes 2D floor plans into 3D wall prisms with proper thickness and height.
  • NPU Acceleration: Seamlessly utilizes the NPU for core AI tasks to ensure real-time performance and energy efficiency.

🧠 Our AI Pipeline

Our system uses a combination of specialized, lightweight models for maximum accuracy and speed.

  • Structure Detection: A lightweight CNN + Transformer architecture (like M-LSD) turns the image into clean line segments.
  • Symbol Detection: A quantized YOLOv8-Nano model identifies specific symbols like doors and windows.
  • Text Recognition: We use PaddleOCR for its high accuracy and efficiency in reading dimensions and labels.
  • Geometric Refinement: A Graph Neural Network (GNN) analyzes the relationships between detected lines to refine alignment and snap junctions, ensuring a geometrically perfect output.

🛠️ Installation

  1. Install Python 3.8 or higher.
  2. Install dependencies: 
    pip install -r requirements_local.txt

💻 Usage

Command Line

# Convert a hand-drawn sketch
python local_png_to_dxf.py my_sketch.png -o sketch.dxf

# Convert a professional floorplan with custom scale
python local_png_to_dxf.py professional_plan.png --pixels-per-meter 150 --wall-thickness 0.1 --wall-height 2.5

Python API

from local_png_to_dxf import process_png_to_dxf

# Convert PNG to DXF
output_path = process_png_to_dxf(
    "input_floorplan.png",
    output_path="output.dxf",
    show_preview=True
)

⚙️ Configuration

You can modify the Config class in local_png_to_dxf.py to adjust:

  • Scale: PIXELS_PER_METER
  • Wall Properties: WALL_THICK_M, WALL_HEIGHT_M
  • Processing: SNAP_TOL_DEG, MERGE_GAP_PX
  • NPU: USE_NPU, NPU_PROVIDER

📂 File Structure

├── local_png_to_dxf.py      # Main converter logic
├── app_local.py             # Gradio web interface for real-time demo
├── test_local_converter.py  # Test script
├── requirements_local.txt   # Dependencies
├── README.md                # This file
└── outputs/                 # Default output directory
    └── *.dxf                # Generated DXF files