VRS-NeRF: Visual Relocalization with Sparse Neural Radiance Field

The implementation of VRS-NeRF: Visual Relocalization with Sparse Neural Radiance Field which provides a a new baseline of applying NeRFs to visual relocalization task. We use zip-nerf as implicit learning map (ILM) and 3D reconstruction as explicit geometric map (EGM) for efficient localization. In the localization process, we use SFD2 as local features, IMP as matcher, and online rendering of patches for matching. Results on 7Scenes and CambridgeLandmarks are promising and much better than previous LENS and NeRF-loc. However, the pose accuracy on Aachen dataset is not satisfying because of the poor quality of rendered images.

Results

Rendered video

Groundtruth and rendered images

Matching between query and rendered images

7Scenes dataset (patch size=15, median position (cm), rotation (deg) errors and average percentage of poses within error of 5cm, 5deg)

	chess	fire	heads	office	pumpkin	kitchen	stairs	Average (%)
LENS	3, 1.3	10, 3.7	7, 5.8	7, 1.9	8, 2.2	9, 2.2	14, 3.6	-
NeRF-loc	2, 1.1	2, 1.1	1, 1.9	2, 1.1	3, 1.3	3, 1.5	3, 1.3	89.5
ACE	2, 1.1	2, 1.8	2, 1.1	3, 1.4	3, 1.3	3, 1.3	3, 1.2	97.1
SP+SG	0, 0.1	1, 0.2	0, 0.2	1, 0.2	1, 0.1	0, 0.1	2, 0.6	95.7
SFD2+IMP	0, 0.1	1, 0.2	0, 0.2	1, 0.2	1, 0.2	0, 0	2, 0.5	95.7
VRS-NeRF	0, 0.1	1, 0.2	0, 0.2	1, 0.2	1, 0.2	0, 0.1	3, 0.8	93.1

CambridgeLandmarks (patch size=15, median position (cm), rotation (deg) errors and average percentage of poses within error of 25cm, 2deg)

	Kings College	Great Court	Old Hospital	Shop Facade	St Mary Church	Average (%)
LENS	33, 0.5	-	44, 0.9	27, 1.6	53, 1.6	-
NeRF-loc	7, 0.2	25, 0.1	18, 0.4	11, 0.2	4, 0.2	-
ACE	18, 0.4	42, 0.2	31, 0.6	5, 0.3	19, 0.6	54.68
SP+SG	7, 0.1	12, 0.1	9, 0.2	2, 0.1	4, 0.1	89.4
SFD2+IMP	7, 0.1	11, 0.1	10, 0.2	2, 0.1	4, 0.1	89.1
VRS-NeRF	9, 0.1	-	11, 0.2	2, 0.1	5, 0.2	89.3

Aachen dataset (percentage of poses within error of 0.25m, 2deg / 0.5m, 5deg / 5m, 10deg)

	Day	Night
ESAC	42.6 / 59.6 / 75.5	3.1 / 9.2 / 11.2
HSCNet	71.1 / 81.9 / 91.7	32.7 / 43.9 / 65.3
SP+SPG	89.6 / 95.4 / 98.8	86.7 / 93.9 / 100.0
SFD2+IMP	89.7 / 96.5 / 98.9	84.7 / 94.9 / 100.0
VRS-NeRF (15)	60.8 / 67.8 / 73.1	19.4 / 22.4 / 25.5
VRS-NeRF (31)	70.1 / 76.9 / 80.9	44.9 / 51.0 / 62.2

Imperfect rendering

Code and pretrained models will come soon

Citation

 @article{xue2024vrs,
          author    = {Fei Xue and Ignas Budvytis and Daniel Olmeda Reino and Roberto Cipolla},
          title     = {VRS-NeRF: Visual Relocalization with Sparse Neural Radiance Field},
          booktitle = {ECCVW},
          year = {2024}
 }

@inproceedings{sfd22023,
        title={{SFD2: Semantic-guided Feature Detection and Description}},
        author={Xue, Fei and Budvytis, Ignas and Cipolla, Roberto},
        booktitle={CVPR},
        year={2023}
}

@inproceedings{imp2023,
        title={IMP: Iterative Matching and Pose Estimation with Adaptive Pooling},
        author={Xue, Fei and Budvytis, Ignas and Cipolla, Roberto},
        booktitle={CVPR},
        year={2023}
}

@inproceedings{barron2023zipnerf,
      title={Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields}, 
      author={Jonathan T. Barron and Ben Mildenhall and Dor Verbin and Pratul P. Srinivasan and Peter Hedman},
       booktitle={ICCV},
       year={2023}
}