Parallel Inversion of Neural Radiance Fields for Robust Pose Estimation

• DOI: 10.48550/arxiv.2210.10108
• Publication Date: 2022-01-01
• Authors: Yunzhi Lin; Thomas Müller; Jonathan Tremblay; Bowen Wen; Stephen Tyree; Alex Evans; Patricio A. Vela; Stan Birchfield
• Keywords: Robustness; Maxima and minima; Global illumination

We present a parallelized optimization method based on fast Neural Radiance Fields (NeRF) for estimating 6-DoF pose of camera with respect to an object or scene. Given single observed RGB image the target, we can predict translation and rotation by minimizing residual between pixels rendered from NeRF model in image. integrate momentum-based extrinsic procedure into Instant Graphics Primitives, recent exceptionally implementation. By introducing parallel Monte Carlo sampling estimation task, our overcomes local minima improves efficiency more extensive search space. also show importance adopting robust pixel-based loss function reduce error. Experiments demonstrate that achieve improved generalization robustness both synthetic real-world benchmarks.