CVSat-NeRF: Neural Radiance Fields Supervised by Cost-Volume Information for Few-View Satellite Imagery

Abstract

Digital Surface Models (DSMs) derived from satellite imagery are traditionally generated using Multi-View Stereo (MVS)
pipelines, which rely on patch-matching techniques such as Semi-Global Matching (SGM). While effective in few-view
settings, these methods exhibit known limitations, for instance when rendering homogeneous surfaces or building borders.
Recently, deep learning-based volume rendering techniques, particularly Neural Radiance Fields (NeRF), have emerged as
a promising alternative for their performance and their compact, flexible continuous representation of 3D scenes. Although
primarily developed for multi-view scene reconstruction, their ray-casting mechanism makes them appealing for sparse-
view settings. SparseSat-NeRF (SpS-NeRF) is among the few frameworks that adapt few-view NeRF methodologies to
remote sensing. It leverages SGM-derived depth priors to define a depth supervision loss. In the process, patch similarity
is interpreted as a proxy for the confidence placed in the prior and used to regulate the attachment to the supervision
and to guide ray sampling. In this work, we introduce CVSat-NeRF to refine the exploitation of such priors. First, we
replace patch similarity with confidence derived from ambiguity, a more informative proxy that aggregates cost-volume
information more effectively. A histogram rescaling step is also proposed to tune the strength of the supervision weighting.
Second, we introduce a sampling strategy that directly derives sampling distributions from SGM cost curves. This approach
naturally adapts to the scale of the scene with minimal parameterization, while making better use of the SGM cost volume
information. Experiments on a two-view scene demonstrate improvements over SpS-NeRF, achieving a better balance
between attachment to the depth prior and detachment from its flaws.