Image modeling system based on photometric stereo algorithm design and implementation
|Course||Applied Computer Technology|
|Keywords||Image-Based Modeling Multi-view Stereo Photometric Stereo Orientation Consistency Virtual Reality|
Among the methods of 3D modeling, Photometric Stereo is one of the typical Image Based Modeling (IBM) methods. It is featured by its accurate result, low cost, simple modeling routine and high efficiency when compared with 3D-Scanning, geometry modeling and other Image Based Modeling techniques. There are prospective applications in Virtual Reality and Industry Manufacturing.On the other side, Photometric Stereo has many shortcomings. Its accuracy can not reach the expected value in many cases. And there are many restrictions on the appearance of object being modeled. Once not observed, the outcome will deviate a lot from the idea value. Thirdly, some of the algorithms are solved simply by a purely mathematic way, having a lot potential for improvement.This paper mainly discusses the principle and algorithm design of Photometric Stereo. Firstly, a practical solution aiming at more accurate results is proposed, to apply the high-light model in Photometric Stereo. Then, facing up to the poor results caused by the high-light and shadow, which are often inevitable, we put forward an algorithm to eliminate their side effect base on existing 4-source algorithms, In this algorithm, multi-photographs are used as a whole to validate whether all these photographs are under Lambertian shading model. This algorithm is easy to implement, and turn out to have strong robustness.On the basis of research on the algorithms using shading models, we further focus onto the Photometric Stereo algorithms using reference objects. These algorithms use nearest neighbor search algorithms (NNS) to search observation vectors on reference object when modeling an object, based on the orientation consistency principle. And their robustness and accuracy benefit from this solution. We propose a pyramid-based algorithm to accelerate the speed of searching, by exploiting the neighbor information of both the object being referenced and the object to be modeled. In addition, our pyramid algorithm reuses the NNS algorithms, to guarantee that the poorest performance is not worse than NNS.As Photometric Stereo shows more varieties than other modeling techniques, different algorithms of Photometric Stereo may require different kinds of input images, then may have different parameters, and eventually lead to different processing sequences and results. But there are rarely systems now accommodating these complex processes. Based on the algorithms we designed, and the possible requirement of different users, such as experts and ordinary users, we build a powerful, user-friendly and extensible Image-based Modeling system mainly based on Photometric Stereo, and the system has been proved to be of good performance by our testing on real models.