---

Abstract

Monitoring of large structures for deformations can be performed by aligning the 3D point clouds of the structure scanned at different times. Existing methods based on the Iterative Closest Point algorithm either ignore features due to high outlier ratios or require measurements to adjust the estimations. In this paper, we propose to use a deep registration framework, not only for learning discriminative features but also their associated confidences. The confidences are directly informed by the registration in an end-to-end manner, resulting in state-of-the-art performance without any extraneous measurements. Our work is the first to use learned features and deep registration models for deformation monitoring; therefore, we plan to share our data, code, and the models to benefit further research.

---

Results

---

Paper