Research And Implementation Of Differentiable Physical Rendering Method Based On Edge Sampling

Three-dimensional graphics rendering has always been a key research content in the field of computer graphics,and there has been a lot of mature research work on real,accurate and fast graphics rendering.The deep neural network technology that has emerged in recent years has also achieved remarkable success in the field of computer vision.The differentiable rendering module can perform forward rendering in the deep neural network,and can calculate the derivative of rendering parameters,which allows the neural network to learn parameter information in the 3D scene.In this paper,the following work has been done to solve the problems of slow forward rendering,slow convergence speed of reverse rendering,and the inability to achieve a proper balance between the realism and speed of the rendering effect: First of all,this article adds support for the physically-based lighting model based on the microfacet theory in the differentiable renderer.The physics-based lighting model is more realistic than the experience-based lighting model,and is faster than the measurement-based method.It has achieved a proper balance between rendering speed and picture quality,and also allows the neural network to learn the roughness of the rendered object;Secondly,this article adds support for ambient light occlusion,and the ambient light term can also be derived from this.Ambient light occlusion can better express the light occlusion relationship between the parts of the object,making the picture more realistic like global illumination;finally,this paper also designs and implements the importance edge sampling method optimized according to the light direction and the normal direction,This method speeds up the convergence speed of reverse rendering and improves the stability.Based on the above work,this paper implements a simple differentiable physical renderer Kay based on the edge sampling method.It supports loading of 3D models composed of triangular fragments,setting of camera and object parameters,forward graphics drawing,and back propagation calculation.It can calculate the derivative of the input rendering parameters in the network by embedding it as a rewritten pytorch custom function.In addition,this article conducted experiments on classic 3D models in the field of graphics to verify the capabilities of the renderer.The algorithm model in this paper can perform effective differential rendering for the geometric figure plane composed of triangular fragments.It cannot effectively achieve complex lighting effects such as volume rendering and subsurface scattering simulation,and it cannot effectively achieve discontinuous textures and procedurally generated material parameters.