Description

Current NRSfM methods are limited in not providing a fine and dense reconstruction with images recorded from a single monocular camera. In this work, we take advantage of the latest state-of-the-art research in NeRF and non-rigid body priors to improve the results of NR-SfM works in long sequences with large deformations. NeRF-based NRSfM is a recent idea that has not been explored in-depth so far, which we decide to explore due to its scene representation potential. We propose to separate deformation into coarse and fine deformations, which are more adequate to represent articulated objects. Previous methods only consider a linear blend skinning model or an as-rigid-as-possible assumption, therefore being very limited in the deformations they can capture. Introducing fine elastic deformation refinements over the coarse estimation allows the network to correctly refine the coarse shape, modeling finer details of the geometry. We obtain better qualitative and quantitative results by following this approach compared to previous work. Our work also shows better view synthesis in rendered images from the learned neural radiance field.