Neural Triangular Mesh Compression Based Efficient Neural Radiance Fields

Topological polygon-based approaches have facilitated the rendering of neural radiance fields through traditional polygon rasterization pipelines. However, challenges persist in neural rendering, such as extended training durations and substantial storage demands. We introduce a multi-resolution dense voxel-based representation aimed at accelerating the model’s training process and a neural triangular mesh to compress the model in this paper. The dense voxels directly capture 3D geometry and learn each vertex’s feature, and other points’ features in the scene are obtained by trilinear interpolation, which reduces the size of the MLP and makes faster convergence. Only the feature vectors of the points that make up the surface of the scene need to be stored, thus greatly reducing the storage space required for the model. Experimental results on multiple public datasets demonstrate that our method substantially enhances training efficiency without compromising rendering quality. The model’s training time is reduced to 1/6 of the original and only 30% of the storage space is required. Visualized experimental results further confirm our proposed method’s high-quality novel view synthesis capabilities.