Road Crack

Joint Topology-Preserving and Feature-Refinement Network for Curvilinear Structure Segmentation (JTFN)

Research: We propose to tackle a new problem that we call curvilinear object segmentation, which unifies existing segmentation problems with curved lines and continuous edges, such as crack detection, road extraction, and blood vessel detection. We add extra supervision on the U-Net decoder to preserve topology, and add gated units to skip connections to enable feature refinement. Compared to other studies that handle each task separately (such as crack detection, road extraction, and blood vessel detection), our unified framework offers SoTA segmentation quality in every task.

Below illustrates an example of our solution. The input images are illustrated in the first row, and the ground truth in the second row. Our model, JTFN, can effectively capture fine-grained object topology with high precision close to the ground truth. Remarkably, our method occasionally surpasses human annotators. For instance, our model successfully detects the ring road in the aerial view, which was missed by human annotators. Or on the blood vessel, our model renders cleaner segmentation on blood vessel that are barely visible to human eyes. We also compare our method to SoTA methods in each task, and showed comparable or better performance. If you would be interested in these findings, I will be happy to share more offline or in follow-up discussions. Check our paper at ICCV21 and code

Application: The model was landed on Docomo AI platform to detect cracks on roads and bridges.