ERC Junior grantee Sylvain Lefebvre presents ShapeForge, a project with style

Changed on 25/03/2020
Sylvain Lefebvre is one of four Inria candidates selected in the young researcher category for the 2012 ERC European call for projects. His project, named ShapeForge, will make it possible to produce, quite literally, everyday objects or pieces of furniture from given examples, taking inspiration from their style.

You are passionate about computer-generated images. What do you seek in these virtual worlds?

I've always been fascinated by the idea of actually being involved in producing the world around us and of building virtual worlds and exploring them interactively. I used to create video games as a hobby before getting inspired for my research work. My aim has always been to make it as easy as possible to design such worlds, while still producing richly detailed images. Until 2010, when I joined the Alice team in Nancy, I worked on improving the textures used in computer graphics—images that give the illusion of different materials such as wood or marble. More specifically, since my thesis at Inria Grenoble, during my postdoctoral work at Microsoft Research in the United States, and then as a member of the Reves team at Sophia-Antipolis, we have been developing methods of "example-based texture generation".

What does that mean?

Instead of providing tools to paint textures, these methods make use of algorithms capable of generating textures that resemble a given sample and reproducing it on a large scale. I initially worked on stochastic textures as these account for the majority of materials to be imitated. Since 2009, I have taken an interest in structured textures, such as building frontages with repetitive window layouts, doors with a specific design, ironwork, and so on. When these textures are produced by artists in graphics studios using retouched photographs, they require many hours of repetitive work to design entire towns, adapting the scale of the façades, door shapes, etc. Our algorithm is able to generate these various textures using a single example. It is already providing some good results.

How did you make the transition from texture to geometry?

The idea is quite simple: geometry is, by definition, structured, like the textures I work with. Just as we reproduce various structured textures using examples, I had the idea of reproducing the different geometries of objects, creating a bench in Louis XV style, for example, using a Louis XV chair. With this kind of example-based generation of objects, we can take inspiration from all kinds of everyday objects, such as pieces of furniture, crockery or even staircases and hand rails, and so on. We will, of course, be able to use these design methods in video games. However, my ERC project is more ambitious than that. Working on the basis that 3D printers, which are becoming more widespread, let just about anyone produce objects very simply, I want to design algorithmic methods to create new objects automatically while complying to a certain style. These objects won't be film props, but actual objects! We will be able to use them in our daily life.

Is this where the virtual world moves back into the real world?

Exactly. An increasing number of methods used to design and produce objects, until now reserved for professional designers, are being made available to the general public. This new bridge between the virtual and the real should have a considerable impact on our daily lives. My goal is to create a top-flight European team working in this field.

ShapeForge: a scientific and algorithmic challenge

The main difficulty with this project is combining approaches that are very different in nature: algorithms from computer graphics which are used to build forms and textures using examples are combined with digital optimisation methods which make sure that the real object complies with the function it is assigned. Thus, to produce a Louis XV bench, on the basis of a Louis XV chair, you need to not only capture the appearance of the example but also formalise the characteristics of a bench (the flat surface of the seat, its dimensions, and its height) as well as its mechanical properties to ensure that it is solid enough. You then need to find, from among all the shapes that can be produced from a single example, the one that best complies with the various criteria. This all has to be done rapidly enough for the user to be able to interact with the system. The transition from virtual to real is an amazing scientific and algorithmic challenge.!