PaRK-Detect

Patch-wise road keypoint detection from satellite imagery using deep learning.

PaRK-Detect is a deep learning-based pipeline for extracting road network maps from high-resolution satellite imagery using patch-wise keypoint detection and linking. It is built upon the work from PaRK-Detect: Towards Efficient Multi-Task Satellite Imagery Road Extraction via Patch-Wise Keypoints Detection (BMVC 2022).

What is PaRK-Detect?

PaRK-Detect formulates road extraction as a keypoint detection and linking problem instead of traditional pixel-wise segmentation. This formulation reduces redundancy, improves connectivity, and is highly efficient for large-scale road graph construction.

It performs:

• Keypoint classification: Classifies each 16×16 patch in a 1024×1024 image as containing a road keypoint or not.
• Keypoint localization: Determines the exact position of a keypoint within the patch.
• Keypoint linking: Predicts links between neighboring keypoints to construct a road network graph.

How It Works

1. Input Preprocessing:

   • Satellite masks are preprocessed into scribble labels via skeletonization and selective simplification.
   • 1024×1024 binary road masks are divided into 64×64 grid patches of 16×16 each.

2. Keypoint Extraction:

   • In each patch:
     • A keypoint is classified based on the patch's structure (intersection, endpoint, or line).
     • The keypoint's coordinates are computed from local road geometry.

3. Link Prediction:

   • Each keypoint is evaluated for connectivity to its 8 neighbors (N, NE, E, SE, S, SW, W, NW).
   • Valid links are stored in an anchor_link matrix.

4. Model Training:

   • The model is trained on 1024×1024 image-label pairs.
   • Loss functions optimize classification, regression (keypoint coordinates), and link prediction.

5. Postprocessing:

   • Outputs are reassembled into road network graphs.
   • Optionally converted to shapefiles for GIS applications.

Inputs

• Satellite road masks (1024×1024 .jpg)
• Naming Convention xxx_sat.jpg, xxx_mask.png, xxx_mask.mat
• Each image is preprocessed into:
  • A scribble mask (./scribble/)
  • Keypoint presence (if_key_points)
  • Keypoint coordinates (all_key_points_position)
  • Link predictions (anchor_link)

Installation

git clone https://github.com/lokrim/PaRK-Detect.git cd PaRK-Detect pip install -r requirements.txt

Citation

This repository is based on the official implementation of: PaRK-Detect: Towards Efficient Multi-Task Satellite Imagery Road Extraction via Patch-Wise Keypoints Detection Shenwei Xie (BUPT PRIS), BMVC 2022