| In the field of computer graphics and computer vision,3D reconstruction is one of the most important research directions.Taking RGB pictures as input,generating a 3D model of the target is challenging but beneficial.Compared with stereo 3D reconstruction,monocular RGB pictures are not only more convenient to obtain,but also more prominent when it comes to practice.Therefore,monocular 3D reconstruction is more popular by researchers currently.However,a single RGB image contains insufficient information,which is more difficult to infer the surface of an object.With the rapid development of deep learning technology and the emergence of large-scale 3D data sets,a series of algorithms based on generative adversarial networks have been proposed and have achieved good results.However,most deep learning models did not consider the impact of the lighting information about the scene on the rendering of the target object,and the neural network model is affected by the amount of network parameters,which can only process image data with lower resolution,relying on large scale 3D data sets for longer time training at the same time.Aiming at the above-mentioned problems,this thesis proposes an iterative 3D reconstruction algorithm for monocular RGB images under Lambert’s illumination model based on SFS,which can calculate the normal vector and pixel depth of the corresponding sampling point iteratively according to the pixel value of the image,and can generate a mesh model of the target object.It does not rely on any data set for training,with the ability to process monocular RGB images of larger resolution.What’s more,the method constructs the relationship between the skeleton and edge of the object and supports custom the vertices or meshes of the mesh model.Experiments show that the method proposed in this thesis effectively reconstructs the mesh of simple target objects,and achieves more accurate reconstruction results than other methods in quantitative analysis.The method which can be combined with neural networks will handle more complicated images even with texture in the future. |
