pnerfp.md

Parallel Inversion of Neural Radiance Fields for Robust Pose Estimation

Yunzhi Lin .et al

Goal

• Parallelised optimisation for estimating 6-DoF poses with monocular RGB-only input.
• Approach:
  1. integrate a momentum-based camera extrinsic optimisation procedure into Instant NGP
  2. parallel Monte Carlo sampling to overcome local minima
  3. Robust pixel-based loss function
• Further explore the idea of pose estimation via NeRF inversion (iNeRF)

Related works

Generalisable 6-DoF Pose Estimation

Instant NGP

iNeRF

Methods

Momentum-based Camera Extrinsics Optimisation

Camera Pose Representation

Momentum-Based Optimisation

• $\mathbf{r}i = \mathbf{o} + t_i \mathbf{d}$ $$ \tau(\mathbf{d}) = \frac{1}{K} \sum^{K}{i=1} t_i \mathbf{d} \times \frac{ \partial \mathcal{L} }{ \partial \mathbf{r}_i } $$

Parallelised Monte Carlo Sampling

• A single camera pose is vulnerable to local minima during optimisation iterations

Pixel-based RGB Loss