Distance and Edge Transform for Skeleton Extraction

Xiaojun Tang, Rui Zheng, Yinghao Wang; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2136-2141

The shape skeleton or medial axis, is a concise shape description, defined by the centers of the maximally inscribed circles. Skeletonization algorithms support many applications, including optical character recognition, object recognition, pose estimation, shape matching, biomedical image analysis, etc. Usually classical algorithms tend to produce redundant skeleton branches at edge noise regions and require a branch pruning post process. Recently many CNN based algorithms achieved significant performance improvements compared with classical algorithms. Most deep learning algorithms directly used the shape image as input data and it's complex for end to end learning algorithms to fit the transformation from shape to skeleton. In this work, we proposed to use Smooth Distance Estimation (SDE) and Edge transformation to preprocess the input shape. Combined with a modified U-Net model and multiple models ensemble, the proposed method achieved 0.8129 F1 score in the Pixel SkelNetOn validation set, 1.5752 symmetric chamfer distance in the Point SkelNetOn validation set and 6407.4 squared distance score in the Parametric SkelNetOn validation set.