S4-NeRF: Neural Radiation Fields Based on Structured State Space Sequence Model

Neural Radiation Field (NeRF) is a popular method that achieves a highly accurate representation of a 3D scene by optimizing a continuous volumetric scene function. NeRF's success is largely due to the use of volumetric rendering (VR) technology, but this is also its main shortcoming when it comes to generating viewing angle-dependent effects. It can have problems with blurring when rendering transparent objects or high-gloss surfaces. To solve this problem, the S4-NeRF model is proposed in this paper. The model not only breaks through the limitations of traditional volume rendering methods by introducing a structured state space sequence model to optimize the volume rendering process of rays, but also draws on the ideas of existing banked models to design a baked module based on the linear attention mechanism to more accurately capture the view-dependent effects in the scene. Experiments on Blender and Shiny Blender datasets show that the model significantly improves the visual quality of transparent objects and highlighted surfaces.