Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Correcting Faulty Road Maps by Image Inpainting

Soojung Hong, Kwanghee Choi

TL;DR

The paper addresses the challenge of correcting faulty road map predictions by reframing automatic mending as an image inpainting problem. It introduces Globally and Locally Consistent Road map Completion (GLCRC), a data-driven architecture with a generator employing deep dilated convolutions and a multi-scale context discriminator, optimized with perceptual and RaLSGAN losses to produce geometrically coherent, sharp road maps. Evaluations on the Massachusetts Road Geometry dataset across straight, curvy, T-junction, and intersection geometries show that GLCRC+L outperforms the Vanilla GLCIC baseline in correctness, completeness, and quality, while maintaining efficient inference. The method decouples from upstream road extraction models, enabling robust, scalable production deployment for road map maintenance with minimal hand-tuning.

Abstract

As maintaining road networks is labor-intensive, many automatic road extraction approaches have been introduced to solve this real-world problem, fueled by the abundance of large-scale high-resolution satellite imagery and advances in computer vision. However, their performance is limited for fully automating the road map extraction in real-world services. Hence, many services employ the two-step human-in-the-loop system to post-process the extracted road maps: error localization and automatic mending for faulty road maps. Our paper exclusively focuses on the latter step, introducing a novel image inpainting approach for fixing road maps with complex road geometries without custom-made heuristics, yielding a method that is readily applicable to any road geometry extraction model. We demonstrate the effectiveness of our method on various real-world road geometries, such as straight and curvy roads, T-junctions, and intersections.

Correcting Faulty Road Maps by Image Inpainting

TL;DR

The paper addresses the challenge of correcting faulty road map predictions by reframing automatic mending as an image inpainting problem. It introduces Globally and Locally Consistent Road map Completion (GLCRC), a data-driven architecture with a generator employing deep dilated convolutions and a multi-scale context discriminator, optimized with perceptual and RaLSGAN losses to produce geometrically coherent, sharp road maps. Evaluations on the Massachusetts Road Geometry dataset across straight, curvy, T-junction, and intersection geometries show that GLCRC+L outperforms the Vanilla GLCIC baseline in correctness, completeness, and quality, while maintaining efficient inference. The method decouples from upstream road extraction models, enabling robust, scalable production deployment for road map maintenance with minimal hand-tuning.

Abstract

As maintaining road networks is labor-intensive, many automatic road extraction approaches have been introduced to solve this real-world problem, fueled by the abundance of large-scale high-resolution satellite imagery and advances in computer vision. However, their performance is limited for fully automating the road map extraction in real-world services. Hence, many services employ the two-step human-in-the-loop system to post-process the extracted road maps: error localization and automatic mending for faulty road maps. Our paper exclusively focuses on the latter step, introducing a novel image inpainting approach for fixing road maps with complex road geometries without custom-made heuristics, yielding a method that is readily applicable to any road geometry extraction model. We demonstrate the effectiveness of our method on various real-world road geometries, such as straight and curvy roads, T-junctions, and intersections.
Paper Structure (14 sections, 4 figures, 2 tables)

This paper contains 14 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Automatically Correcting Faulty Road Map Predictions
  3. Challenges of inpainting road maps
  4. Model Architecture
  5. Generator with Deep Dilated Convolutions
  6. Multi-scale Context Discriminator
  7. Loss function
  8. Avoiding Blurriness via Perceptual Loss
  9. Sharper Maps via Relativistic Least Square GAN (RaLSGAN) Loss
  10. Experiments
  11. Dataset
  12. Training details
  13. Experimental Results
  14. Conclusion

Figures (4)

  • Figure 1: GT, Input, and Output denotes the ground truth, the input image that our model receives, and its road inpainting result. Our method successfully reconstructs the real-world road types including T-junctions and intersections by capturing the geometric information of the surroundings.
  • Figure 2: Road geometry inpainting using RePaint. RePaint fails to yield geometrically sensible road maps and produces random road maps with different random seed.
  • Figure 3: Decreased road map blurriness by utilizing the perceptual loss: (a) GLCIC with MSE loss, and (b) GLCIC with perceptual loss. We mask the road maps (rectangles in the first row) so that the generator can reconstruct the roads within the mask (second row).
  • Figure 4: Road inpainting results of Vanilla GLCIC (Baseline), GLCRC (Enhanced Architecture), and GLCRC+L (Ours, GLCRC with perceptual loss and RaLSGAN loss) on various road types, (a) Intersection, and (b, c) Curvy road.