Creating a dataset of 3D deformed shapes for evaluating 3D model registration techniques

Caleb Stride


Supervised by Paul L Rosin; Moderated by Yukun Lai

3D scanned data has become increasingly popular with the advent of low-cost devices such as the Kinect. There are many applications of 3D data, such as gaming, cinematography special effects, 3D printing, etc.

Due to unavoidable occlusion, multiple scans of an object are usually required to reconstruct the full shape. Many surface registration techniques have been developed to align multiple scans. However, so far, techniques for evaluating the registration techniques are still very limited.

The project aims to build a dataset of 3D deformed shapes (e.g. faces, hands, human bodies) with known correspondences. This benchmark dataset will be highly valuable for researchers to evaluate the effectiveness of algorithms that align surfaces based on geometry information.

The student will develop algorithms during the project to identify correspondences using the texture information, in particular markers placed on the surfaces. The student will also evaluate different registration techniques.

Initial Plan (04/02/2019) [Zip Archive]

Final Report (10/05/2019) [Zip Archive]

Publication Form