DS0705 - Fondements du numérique

Patch-aware Processing of Surfaces – PAPS

Patch Aware Processing of Surfaces

Using self-similarities for high resolution point sampled surface analysis and synthesis

Context

This project aims at taking advantage of recent technological evolutions. Indeed, professional surface acquisition become ever more precise and efficient but remain very expensive while cheap low-resolution scanners have appeared. We aim at devising method to enhance the resolution of low-resolution laser data but also to correctly process high resolution data.

Our work is based on the development of methods to describe locally a point-sampled surface. Several reasons make it a challenging task
- the type of data can vary a lot. For example some point sampled surfaces can entail areas that have lower dimensions at the processing scale. We have therefore been working on a descriptor that adapts to the local dimension. This work is based on the careful optimization of the position and orientation of deformation fields.
- it is also important to be able to recognize similarities when they are lying on surfaces with different curvatures. We have developed descriptors that are robust to a change in the curvature of the underlying surface.
- from a more theoretical point of view, we have analyzed the local differential properties of the shape in order to better understand the geometrical meaning of the frequency decomposition of closed paths defined on the surface.

Thanks to our defined descriptors we have been able to obtain promising results for surface super-resolution, denoising and resampling.

Our descriptor based super-resolution method is now free from any registration step which is usually costly and not very robust. It can work directly on a single scan which is a consequence of the self-similarity principle which we were able to emphasize.
Our denoising and resampling methods are very efficient to handle data with heterogeneous dimensions.

At this stage of the project, we have several future directions.
- the first is to finalize the publications of our first methods and applications.
- we aim at developing a probabilistic definition fo the self-similarity of signals be it for one-dimensionnal signals, images of shapes.
- finally we will pursue the application of our descriptors to various problems which require a better definition of comparison metrics between local descriptions.

Submitted publications:
*Sparse Geometric Representation Through Local Shape Probing ,
J. Digne, S. Valette et R. Chaine
*Super-resolution of Point Set Surfaces using Local Similarities , A. Hamdi-Chérif, Y. Béarzi, J. Digne, R. Chaine
*The Bilateral Filter for Point Clouds , J. Digne soumis à Image Processing OnLine (IPOL) 2016

The past decade has seen a radical evolution in 3D surface acquisition and processing which has been driven by two general directions: designing ever higher quality digital acquisition devices on one hand, and designing low cost acquisition devices on the other hand. This evolution is similar to the one witnessed for digital cameras: on the one side ever better reflex cameras and on the other side cheap cameras integrated into mobile phones or computers. This development calls for a variety of tools able to deal with this varying quality to generate the highest resolution possible out of low quality scans and to process high accuracy surfaces. This project proposes to tackle this problem by developing efficient and scalable methods taking advantage of an intrinsic property of surfaces : their natural self-similarity. Indeed, most surfaces, be it from a fine-art artefact or a mechanical object, are characterized by a strong self-similarity. This property stems from the natural structures of objects but also from the fabrication processes: regularity of the sculpting technique, or machine tool.

In the field of surface reconstruction from point clouds, existing approaches generally focus on reconstructing smooth closed surfaces. We have shown in recent works [Digne et al., 2011] that high precision data is most of the times irrevocably smoothed out by these approaches. When the size of the data increases, global methods fail to recover both the shape geometry and the local details. We propose to explore new local approaches, the only ones well suited for large and precise data. The key idea of the project is to avoid processing the point cloud as a global shape, as is usually done in the Geometry Processing literature. Instead, our proposed approach takes into account a smaller scale : the points and their neighborhoods (patches). Analyzing surfaces at such a local scale will permit to reveal surfaces self-similarity, which is the core of our project.

The first step towards a self-similarity aware processing pipeline consists in establishing ways to compare local variations of shapes. To do so we will first decompose the surface in local patches that will be described in an intrinsic way. Armed with these descriptions we will address problems such as high resolution scan registration, scan super-resolution, compression and segmentation of surfaces. A last application will be the design of an interactive tool for exploring surfaces described by point clouds.

Project coordination

Julie DIGNE (Laboratoire d'Informatique en Images et Systèmes d'Information)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

LIRIS - CNRS Laboratoire d'Informatique en Images et Systèmes d'Information

Help of the ANR 134,522 euros
Beginning and duration of the scientific project: September 2014 - 42 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter