Analysis of Iterative Closest Point Algorithms

Iterative closest point (ICP) is a well-known algorithm for shape alignment. ICP is widely used in computer vision to integrate data from different viewpoints and sensors. The goal of this project is to provide an extensive review of existing algorithms and evaluate their performance.

We analyzed different ICP methods and compared between three correspondence methods (nearest neighbors with and without colors and normal shooting), three ICP objective functions (point-to-point, point-to-plane and symmetric), different types of point sampling and using both linearized and full variants. Our analysis was backed up by two datasets based on GREYC dataset and TUM RGB-D dataset.

Brief summary:

• Symmetric ICP is superior and converges faster than others. Point-to-plane ICP is only marginally worse than symmetric ICP.
• Normal Shooting and Nearest Neighbors with color information outperform default Nearest Neighbors correspondence method.
• Adding color gives a huge boost to all methods: weak point-to-point with color outperforms regular point-to-plane and symmetric ICP. Convergence is slower, but can converge on all meshes.
• Normal sampling allows to use 5% of points without significant loss of accuracy.