DeepSharpen: An Unsupervised Sharp Feature Recovery Method Based on Quadric Projection
Jianhui Nie, Yaning Wang, Zhiwei Shen, Xiaohui Lu · IEEE Transactions on Instrumentation and Measurement · 2024
Sharp features are integral components of point cloud data, containing rich surface information that is often compromised due to ambiguous light reflection during optical measurements, such as structured light and RGB cameras. This article aims to address the critical issue of recovering distorted sharp features, which is essential for enhancing the accuracy of measurements from such devices. To achieve this, we propose DeepSharpen, an unsupervised method that first performs feature region point cloud segmentation. Next, we fit the segmented point cloud to quadratic surfaces and infer the intersection lines as the true positions of the features. Finally, we refine additional points within the feature regions through loss functions based on shape and adjacency consistency to accomplish feature sharpening. Our experimental results demonstrate that DeepSharpen not only accurately recovers the damaged sharp features but also exhibits strong generalization capability and noise robustness.