Hello all,
I'm a research student at Brunel university working in the field of online sketching. Dr Garner suggested that I post to this area I would like to tell you about my research and request contacts or collaboration with others that work in a similar area. To put it simply, a small group of us are trying to turn a 2D sketch into 3D geometry. We hope that this can help designers increase their productivity and save time. A detailed description of this can be found below.
I hope that you may find this interesting.
My research mainly consists of using an artificial intelligence technique, neural networks, for interpretation of on-line sketches. Simply, this is the inference of depth from a 2D sketch. To explain in more detail, neural networks are learning machines that can be trained to solve a particular problem based on other examples from the same domain. That is, examples that are of similar nature can be used to provide a solution for a problem that that has not been encountered before.
The aim of the research is intended for conceptual design. A designer would sketch their concept on paper and then pass this onto a CAD draftsman who would then interpret the design into a conceivable CAD model. This is a long-winded process where much of the information and intent of the designer is lost through errors related to perception. Such information such as the actual intent of the designer is very important in the feeling and form of the object under development.
Shapes of interest are those that belong to the category of freeform surfaces, primarily in the forms of patches. These can then be stitched together to form closed surfaces and can be manipulated to take the shape of small objects such as small consumer goods. In particular, patches that are constructed using planar boundary curves are under consideration here. Other forms of freeform surface will be looked at into the future. There has already been much research in reconstructing shapes that consist of planar faces such as cuboids, prisms and similar. One of the challenges that promotes this research is the application of neural network techniques for the reconstruction of freeform surfaces, as little progress has been made in this area so far.
The stages of development for this work involve developing a neural network solution that can be applied to reconstruct freeform surfaces from a freehand sketch. To be able to construct a freeform patch a suitable interface was developed that allowed a user to sketch on-line onto a graphics tablet and hence directly into software. This digital sketch although intended by the designer as a 3D shape was still currently in two-dimensional form. The aim of the neural network was to reconstruct its 3D representation from that single 2D input image.
It can be noted at this point from one 2D sketch based on a single viewpoint, that there can be an infinite number of 3D possibilities that can fit that shape. That statement is true, however there are only a finite number of possibilities that are valid and sensible to represent that surface. These are the ones that most likely represent the shape that the designer had intended and should therefore be reconstructed as so.
To reconstruct a freeform patch, the neural network has to be firstly trained with a series of similar shapes. This involves creating synthetic patches constructed from CAD curves, using mathematical splines. A large varied set of various curves are created in 3D where they are then projected back onto a 2D plane, simulating what the 2D sketch would be for that respective 3D shape. This set is then normalised and applied to the neural network using a supervised training algorithm.
After training, the network is then saved and then can be used for reconstruction of designer input sketches. An input sketch is applied and its reconstructed geometry displayed. This work is almost complete and is still under development to improve the reconstruction processes. Different methods of neural network training and training sets are being experimented with to find the most optimum training sets, training methods and reconstruction algorithms. Future developments will include the application of other information from the sketch such as highlights and shading on the patch. This will provide more information with regard to the topology and surface contours and would be able to improve its form.
Best regards,
Usman Khan.
PhD Student,
CAD Research Team,
Tower A, Room 303,
Brunel University,
Kingston Lane,
Uxbridge, UB8 3PH.
(018952) 66350.
|