Skip to main content

Non-isometric 3D shape registration.

Jiang, T., 2020. Non-isometric 3D shape registration. Doctoral Thesis (Doctoral). Bournemouth University.

Full text available as:

[img]
Preview
PDF
JIANG, Tao_Ph.D._2019.pdf

54MB

Abstract

3D shape registration is an important task in computer graphics and computer vision. It has been widely used in the area of film industry, 3D animation, video games and AR/VR assets creation. Manually creating the 3D model of a character from scratch is tedious and time consuming, and it can only be completed by professional trained artists. With the development of 3D geometry acquisition technology, it becomes easier and cheaper to capture high-resolution and highly detailed 3D geometries. However, the scanned data are often incomplete or noisy and therefore cannot be employed directly. To deal with the above two problems, one typical and efficient solution is to deform an existing high-quality model (template) to fit the scanned data (target). Shape registration as an essential technique to do so has been arousing intensive attention. In last decades, various shape registration approaches have been proposed for accurate template fitting. However, there are still some remaining challenges. It is well known that the template can be largely different with the target in respect of size and pose. With the large (usually non-isometric) deformation between them, the shear distortion can easily occur, which may lead to poor results, such as degenerated triangles, fold-overs. Before deforming the template towards the target, reliable correspondences between them should be found first. Incorrect correspondences give the wrong deformation guidance, which can also easily produce fold-overs. As mentioned before, the target always comes with noise. This is the part we want to filter out and try not to fit the template on it. Hence, non-isometric shape registration robust to noise is highly desirable in the scene of geometry modelling from the scanned data. In this PhD research, we address existing challenges in shape registration, including how to prevent the deformation distortion, how to reduce the foldover occurrence and how to deal with the noise in the target. Novel methods including consistent as-similar as-possible surface deformation and robust Huber-L1 surface registration are proposed, which are validated through experimental comparison with state-of-the-arts. The deformation technique plays an important role in shape registration. In this research, a consistent as similar-as-possible (CASAP) surface deformation approach is proposed. Starting from investigating the continuous deformation energy, we analyse the existing term to make the discrete energy converge to the continuous one, whose property we called as energy consistency. Based on the deformation method, a novel CASAP non-isometric surface registration method is proposed. The proposed registration method well preserves the angles of triangles in the template surface so that least distortion is introduced during the surface deformation and thus reduce the risk of fold-over and self-intersection. To reduce the noise influence, a Huber-L1 based non-isometric surface registration is proposed, where a Huber-L1 regularized model constrained on the transformation variation and position difference. The proposed method is robust to noise and produces piecewise smooth results while still preserving fine details on the target. We evaluate and validate our methods through extensive experiments, whose results have demonstrated that the proposed methods in this thesis are more accurate and robust to noise in comparison of the state-of-the arts and enable us to produce high quality models with little efforts.

Item Type:Thesis (Doctoral)
Additional Information:If you feel that this work infringes your copyright please contact the BURO Manager.
Uncontrolled Keywords:shape registration
Group:Faculty of Media & Communication
ID Code:34203
Deposited By: Unnamed user with email symplectic@symplectic
Deposited On:24 Jun 2020 15:49
Last Modified:25 Jun 2020 10:08

Downloads

Downloads per month over past year

More statistics for this item...
Repository Staff Only -