RoadPainter introduces a pioneering model that, through semantic image synthesis, represents a substantial advancement in reducing data collection costs and annotation efforts, offering a transformative tool for researchers and practitioners in infrastructure management.
The
LIAISON project is proud to announce the release of a scientific paper introducing a
new generative model, called Road Painter, designed to generate synthetic road crack images from segmentation masks, offering a
significant advancement in reducing the need for extensive data collection and costly annotations for road crack detection.
As a matter of fact, computer-aided deep learning has greatly improved road crack segmentation, but supervised models struggle with the
scarcity of annotated images. Additionally, there is
limited focus on extracting pavement condition indices from the predicted masks.
Road Painter sets a new standard in image quality, outperforming existing generative models across multiple benchmarks. Through its architecture, which includes 134 million parameters optimized for efficiency, these optimizations were made possible through careful adjustments in the model's structure, the implementation of advanced noise schedules, and a comprehensive analysis of Classifier-Free Guidance.
In addition to producing superior images,
RoadPainter also enhances the performance of segmentation tasks across major architectures. Synthetic images are added to datasets of real images, enhancing their volume and the diversity of pavement cracks. After applying this data augmentation technique, it has been demonstrated that several semantic segmentation models improve their detection metrics across multiple open datasets.
Additionally, a
hybrid system has been developed that combines an object detection model with a segmentation model enhanced by RoadPainter. The results of this system enable the calculation of a refined pavement condition index, which is valuable for intelligent maintenance planning by road administrations.
RoadPainter represents a
significant leap forward in the field of semantic image synthesis, offering practical applications in infrastructure maintenance and beyond. With its advanced capabilities, this model is set to become an essential tool for researchers and practitioners alike.
This paper, authored by Saúl Cano-Ortiz, Eugenio Sainz-Ortiz, Lara Lloret Iglesias, Pablo Martínez Ruiz del Árbol and Daniel Castro-Fresno from University of Cantabria (Spain) and the Institute of Physics of Cantabria (IFCA-UC & IFCA-CSIC), has been produced within the
LIAISON Project. The LIAISON project develops and implements sustainable solutions throughout the life cycle of European roads and railways. It promotes circular infrastructure, energy efficiency and the adoption of renewable energy sources.
Read
the full article:
https://shorturl.at/Cd9n7
Know more about
LIAISON project:
Website:
www.liaison-transport.eu/
Vídeo presentation:
www.youtube.com/watch?v=gwecQdicXW0
Linkedin:
www.linkedin.com/company/liaison-eu-project