神经辐射场的研究现状与展望

Current Status and Prospects of Research on Neural Radiance Fields

基于2D图片的视点合成一直是计算机视觉和计算机图形学领域中的一个关键问题,旨在通过一组目标场景的2D图片合成新视角下的场景图片.神经辐射场作为一种新颖的隐式场景表达方式,因其出色的视觉效果备受研究者的关注.对神经辐射场的发展历程进行梳理,从理论基础、优化与扩展以及应用等方面介绍了相关研究.在神经辐射场的优化与扩展方面,通过优化网络结构、模型压缩等方式加速训练及渲染过程,还有一些工作致力于降低对输入图片的要求以及提高渲染质量;在应用方面,神经辐射场在人、物及场景的建模中展现出巨大的潜力,并有工作将其扩展到动态场景的表达.此外,神经辐射场与生成式模型结合,可以实现通过文本或者图像来引导三维模型生成的目标.最后,总结了现有研究工作的不足,并指出加速神经辐射场的训练与渲染,优化渲染结果,以及进一步扩展应用场景仍然是未来相关工作的研究方向.

Novel view synthesis based on 2D images has always been a key issue in the fields of computer vision and computer graphics,aiming to synthesize the images from a novel viewpoint utilizing a set of 2D images of the target scene.Neural radiance field,as a novel implicit scene representation,has attracted great attention from researchers due to its excellent visual effects.The development of neural radiance fields is traced,introducing relevant research from aspects of theoretical foundation,optimization and extension,and applications.In terms of optimization and extension,some work focuses on accelerating training and rendering processes through optimizing network structures,model compression,and reducing the requirements for input images as well as improving rendering quality.Neural radiance fields demonstrate great potential in people,objects,and scene modeling,and some work extends their application to dynamic scene representation.Additionally,by combining neural radiance fields with generative models,it is possible to guide the generation of 3D models through text or images.Finally,the shortcomings of existing research are summarized,pointing out that accelerating the training and rendering of neural radiance fields,optimizing rendering results,and further expanding application scenarios are still the research directions for future work in this field.