Research on Lightweight NeRF SLAM Models through Knowledge Distillation

NeRF（neural radiance field） has great potential in high-quality 3D scene reconstruction. However, its high computational complexity, data requirements, and storage limitations pose numerous challenges in practical applications. To tackle this issue, an improved NeRF-based SLAM（simultaneous localization and mapping） system incorporating knowledge distillation is proposed to achieve rapid and efficient training. The experimental results show that the proposed system improves the point cloud accuracy by 18.21%, the point cloud completeness by 14.86%, and the completion rate by 14.09%in terms of reconstruction accuracy, compared with the original NeRF model. In terms of reconstruction efficiency, it reduces the total FLOPs（floating point operations） by 35.52%. While maintaining the reconstruction accuracy, it significantly reduces the training time and computational resource consumption. This research not only offers new insights for optimizing NeRF SLAM systems, but also opens new paths for the application of knowledge distillation to the 3D vision domain.