一种基于MRF的单幅图像数据的三维重构方法研究

综合分析了常见的基于图像的三维重构方法的优缺点,提出了一种基于单张图像,采用马尔科夫随机场(MRF)推断3D位置和方向的3D重构方法。该算法首先将图片分割成多个小的区域(超像素块),并假定空间场景由许多很小的平面组成,超像素块与平面相互对应,对图像中每个超像素块求取出一组特征向量(纹理、颜色等),使用MRF模型化平面参数之间、超像素特征向量与平面参数之间的关系,采用监督学习的方式求取相关参数,求解MRF模型,并根据平面参数进行场景重建。这种算法不需对场景结构做明确的假定,因此较之以前的方法可以获得更多3D结构细节信息。用该方法对200张图片样本进行验算,样本中有60%实现了较为准确的3D重构。

3 D reconstruction method based on single image data by MRF

The mainstream methods of image-based 3D reconstruction are overviewed. The advantages and disadvantages of these methods are analyzed. A new 3D-reconstruction method is put forward; it is based upon a single image of an unstructured environment. For each small homogeneous patch in the image, a Markov Random Field (MRF) is used to infer a set of ‘parameters’ that contain both the 3D location and 3D orientation of the patch. The space is assumed being made up of many tiny planes, and each plane corresponds with a super-pixel. For each super-pixel in the image, a series of characteristic vectors such as color and texture are calculated. The relations between different planar parameters, as well as between super-pixel vectors and planar parameters are probed by using the MRF model. The MRF is trained via supervised learning. The scene is reconstructed based on the planar parameters. This algorithm does not require explicit assumptions about the structure of the scene, and compared with previous methods it enable us to gain more detailed information about the 3D structure. Using this approach, we have conducted checking computations for 200 pictures, among which 60% have been correctly 3D reconstructed.