Visual localization using implicit representations and particle filtering-based pose refinement

This work proposes HAL-NeRF v2, a localization pipeline that couples direct pose regression with Monte Carlo–based refinement on neural scene representations. Building upon the original HAL-NeRF framework, the proposed system leverages Gaussian Splatting for fast, high-detail synthetic view generation during pose regressor training, and NeRFs for efficient pose refinement. The refinement stage is redesigned with Cauchy loss, systematic resampling, and maximum-likelihood estimation,instead of the HAL-NeRF steps, to improve convergence stability in scenes with transient elements and/or incomplete mapping. In addition, HAL-NeRF v2 evaluates multiple rendered views in parallel during refinement, allowing the particle filter to rapidly disambiguate candidate viewpoints. The experimental results in the Cambridge Landmarks dataset demonstrate an accuracy equivalent to that of HAL-NeRF v1 (0.09 m / 0.61° median translational/rotational error) while achieving a 25× reduction in run time.