Augmenting with NeRFs: Fast Relocalization on Densified Datasets

We reinterpret NeRFs as a resource for extreme data augmentation to improve the field of camera relocalization. Our approach enables us to automatically render a massive, densified dataset of novel views, even if given only sparse ground-truth viewpoints. Compared to other realtime relocalization methods, training a lightweight off-the-shelf vision backbone as a pose regressor on our expanded datasets significantly improves accuracy, uniquely enables relocalization of spatially-novel views, and performs well on portable-scale hardware.