Three-dimensional geometric modeling and shape representation
Why do you want to talk about three-dimensional shape representation? There are two important directions in the field of multimedia information processing, which are also two important directions in the field of computer science and technology. One is computer vision and the other is computer graphics. Computer vision is to get the description of objects from images by recognizing or reconstructing them. On the contrary, graphics are descriptions of an object and a scene, and images and videos are generated by drawing and animation. As far as processing is concerned, they are inverse processes, and there is a close relationship between these two fields.
What we are considering now is computer three-dimensional modeling or model description, not only the content of computer vision or computer graphics. Here is a brief list, such as model-based three-dimensional object recognition and scene recognition, which are typical computer vision applications to consider. Besides, you can also use the model to do other things, such as drawing, including design, deformation, animation and so on based on this model. In addition to the two applications studied separately, how to combine them and seamlessly connect virtual and real things involves three-dimensional models.
Compared with before, the application field of the three-dimensional model we consider is much larger. To achieve this requirement, there are higher requirements for the expressive force of models and modeling.
How to find some newer and more effective expressions to meet these needs? There are four aspects here. Is it possible to establish a shape space now, and express the considered object completely in the shape space? The second point is that expression should be flexible according to the characteristics of modeling. The surface of an object is not all continuous, but there are also some discontinuities and irregularities. How to treat these characteristics flexibly? The third is that some descriptions need to be partial, while others need to be holistic. How to combine the two well? Fourthly, there are many editing tools in a large number of calculations, including deformation and animation, so we have higher requirements for the current shape expression.
A long-standing problem in computer vision is to identify existing models, but what we see in front of us is a part of three-dimensional data. How to identify it by matching this part of data with model data? This involves local matching. What description is needed to solve local matching? We should have a model, build a model base from a large number of data, extract many features from it, and organize these features. This involves three major issues. The first is how to construct the physical space. Without a good expression or arrangement of morphological features, it is difficult to do a good job of subsequent matching. The second is how to define similarity according to this feature. The third is how to find the best match in the shape space.
In order to solve these problems, we have done some work in recent one or two years, and recently proposed a generalized shape distribution description method, which has two main purposes. Can we find some key descriptions, find the relationship between them and combine the whole and the part in the process of three-dimensional shape representation? If we describe the relationship between them, we can increase the local structural information from the overall point of view. These two tasks are the main purpose of doing so.
We still have a lot of work to do on the problem of shape description. In addition to local and overall descriptions, we should try to combine them. In the next few years, more pattern recognition methods will be used to solve many problems in graphics or virtual reality and computer vision.