Neural radiance fields as a regularization approach to inverse volume rendering

Inverse volume rendering poses a significant challenge in reconstructing density and color functions from real-world observations, particularly due to the non injectivity of the problem. This study addresses this issue by introducing a unique criterion to ensure a well-defined solution. We explore the connection between classical regularization, as well as the selection method, and modern neural network approaches. Our methodology integrates minimum support regularization with a novel technique to achieve sharper density estimates, complemented by a georeferencing pipeline that aligns reconstructed point clouds with planar cartographic data, specifically suited for urban sutanability.