HistReNeRF: Historic Image Relocalisation within Contemporary Neural Radiance Field Reconstructions

Benjamin T. Hughes; Stuart James

cs.CVAug 15, 2026

HistReNeRF: Historic Image Relocalisation within Contemporary Neural Radiance Field Reconstructions

Authors: Benjamin T. Hughes, Stuart James

Organizations: Durham University, Durham, UK

Abstract

Relocalising archival photographs within a contemporary scene model is challenging because historic and modern views can differ in photographic appearance, visible objects, and spatial layout. Therefore, we present HistReNeRF, a framework that estimates the 6-DoF pose of a historic photograph by matching adapted DINOv2 patch features to candidate rays sampled from a contemporary Neural Radiance Field (NeRF) reconstruction. The continuous representation of a NeRF provides a queryable scene interface from which candidate rays can be sampled and matched, enabling domain adaptation between historic photography and contemporary images directly in the feature representation used for localisation. We evaluate embedding-space-based domain adaptation against pixel-space methods on a new cross-temporal dataset comprising 10,545 contemporary street-level images and 230 archival photographs from three European landmarks. Embedding-space adaptation reduces translation and rotation errors by an average of 11% and 16%, respectively, across the three scenes. These results show that neural scene relocalisation provides a natural interface for feature-space adaptation, reducing cross-temporal appearance shift without modifying the query image. Code and dataset at https://github.com/ARTUROLab/HistReNeRF.

Figures & tables

Figure 1: Overview of HistReNeRF . Given a historic query image (i), it is encoded by frozen DINOv2 (ii) into patch embeddings (iii), which are mapped towards the contemporary embedding distribution by embedding-space adaptation (iv) to produce adapted query features (v). In parallel, a contemporary TensoRF reconstruction (vi) is used to sample candidate surface points and generate normal-oriented candidate rays, and encode them as ray features (viii). Cross-attention matching (viiii) scores the compatibility between adapted query features and candidate rays. The weighted pose solver (x) aggregates the highest-scoring rays to estimate the 6-DoF camera pose of the historic photograph within the contemporary neural scene.

Scene	Contemporary	Historic	Pose-validated (Test)	Training
Arc de Triomphe	2,371	92	25	67
Brandenburg Gate	5,000	96	18	78
Piazza del Duomo	3,174	42	15	27
Total	10,545	230	58	172

Table 1: HistScene statistics. Pose-validated images form the historic evaluation set. Remaining historic photographs are used as unpaired samples for domain adaptation training.

	Gradient NCC		Edge F1		SSIM (gray)
Scene	Modern	Historic	Modern	Historic	Modern	Historic
Arc de Triomphe	0.124	0.047	0.240	0.637	0.499	0.268
Brandenburg Gate	0.044	-0.001	0.116	0.217	0.375	0.351
Piazza del Duomo	0.089	-0.004	0.092	0.099	0.413	0.418

Table 2: Reference-pose validation metrics averaged per scene. Gradient NCC and Edge F1 assess structural alignment between TensoRF renders and query images; higher is better. Grayscale SSIM is a photometric sanity check and is expected to be lower for historic images.

Panel of Figure 2 — Figure 2: Example images from the proposed HistScene dataset facing the monument with left archival photography and right Mapillary street view.

Figure 3: Qualitative results across scenes. Each row shows a historic query image (col. 1), the domain-transferred output from the best pixel-space translation method (col. 2), the novel view rendered by TensoRF at the HistReNeRF estimated pose (col. 3), the render at the HistReNeRF + iNeRF refined pose (col. 4), and the ground-truth render at the registered pose (col. 5). Misalignment between cols. 3–4 and col. 5 reflects localisation error. The renders are visually imprecise across all methods, consistent with the high MTE and MAE values reported in Table 3 .

		Arc		Brandenburg		Duomo
Method	Setting	MTE	MAE	MTE	MAE	MTE	MAE
IFFNeRF	Historic — no adaptation	6.624	71.7°	4.712	23.4°	1.783	85.5°
IFFNeRF + iNeRF	Historic — no adaptation	6.818	81.5°	5.279	35.7°	2.234	71.0°
HistReNeRF	Historic — domain transfer	6.608	69.6°	3.912	23.4°	1.589	58.6°
HistReNeRF + iNeRF	Historic — domain transfer	6.608	76.4°	4.708	29.2°	1.902	68.9°
Img. CycleGAN	Historic — domain transfer	7.179	88.6°	4.809	25.6°	1.799	72.3°

Table 3: HistScene relocalisation MTE is reported in metres and MAE in degrees (lower is better). Best result per scene and metric is shown in bold.

		Angular (%)						Translation (%)
Scene	Method	5°	15°	30°	45°	60°	90°	0.5	1.0	2.5	4.0	5.0	7.5	10.0
Arc	No DT	3.8	34.6	50.0	57.7	57.7	57.7	0.0	0.0	0.0	15.4	23.1	73.1	80.8
	Img. CycleGAN	0.0	19.2	26.9	34.6	38.5	46.2	0.0	0.0	0.0	11.5	26.9	65.4	80.8
	Img. CUT	3.8	15.4	26.9	34.6	34.6	42.3	0.0	0.0	0.0	0.0	23.1	61.5	92.3
	Img. UNIT	0.0	15.4	15.4	15.4	23.1	26.9	0.0	0.0	0.0	3.8	15.4	53.8	92.3
	Img. MUNIT	0.0	11.5	23.1	23.1	23.1	34.6	0.0	0.0	7.7	11.5	11.5	34.6	61.5

Table 4: Cumulative localisation recall (%) at angular in degrees and translation thresholds. Cell colour indicates recall level: darker green indicates higher recall.

Camera Pose Estimation

Camera pose estimation recovers a camera’s position and orientation from images, visual features, scene geometry, and sometimes sensor measurements.

See all 237 papers

RehearsalNeRF: Decoupling Intrinsic Neural Fields of Dynamic Illuminations for Scene Editing
Mar 30, 2026Changyeon Won, Hyunjun Jung, Jungu Cho +3Neural Radiance Fields 3D Scene Editing
Department of AI Convergence, GIST, Gwangju, Korea. · AI R&D Division, CJ Corporation, Seoul, Korea. · Department of Artificial Intelligence, Yonsei University, Seoul, Korea.
CLEAR-NeRF: Collinearity and Local-region Enhanced Accurate 3D Reconstruction in Unbounded Scenes
May 27, 2026Vladislav Polianskii, Elijs Dima, Isabel Salmerón Marazuela +33D Reconstruction Multi-View 3D Reconstruction
Ericsson Research, Stockholm, Sweden
NaCR: Visual Localization via NeRF-aided Camera Ray Regression
Sep 22, 2026Yesheng Zhang, Xiang Dai, Xu Zhao +1Camera Pose Estimation
Shanghai Jiao Tong University

Mar 30, 2026J/K move · Enter open · S save

RehearsalNeRF: Decoupling Intrinsic Neural Fields of Dynamic Illuminations for Scene Editing

Changyeon Won, Hyunjun Jung, Jungu Cho, Seonmi Park+2

Department of AI Convergence, GIST, Gwangju, Korea. · AI R&D Division, CJ Corporation, Seoul, Korea. · Department of Artificial Intelligence, Yonsei University, Seoul, Korea.

Neural Radiance Fields 3D Scene Editing Dynamic Novel View Synthesis

Open paper AskPDF arXiv

Although there has been significant progress in neural radiance fields, an issue on dynamic illumination changes still remains unsolved. Different from relevant works that parameterize time-variant/-invariant components in scenes, subjects' radiance is highly entangled with their own emitted radiance and lighting colors in spatio-temporal domain. In this paper, we present a new effective method to learn disentangled neural fields under the severe illumination changes, named RehearsalNeRF. Our key idea is to leverage scenes captured under stable lighting like rehearsal stages, easily taken before dynamic illumination occurs, to enforce geometric consistency between the different lighting conditions. In particular, RehearsalNeRF employs a learnable vector for lighting effects which represents illumination colors in a temporal dimension and is used to disentangle projected light colors from scene radiance. Furthermore, our RehearsalNeRF is also able to reconstruct the neural fields of dynamic objects by simply adopting off-the-shelf interactive masks. To decouple the dynamic objects, we propose a new regularization leveraging optical flow, which provides coarse supervision for the color disentanglement. We demonstrate the effectiveness of RehearsalNeRF by showing robust performances on novel view synthesis and scene editing under dynamic illumination conditions. Our source code and video datasets will be publicly available.

Abstract

Camera Pose Estimation

Explore similar work

RehearsalNeRF: Decoupling Intrinsic Neural Fields of Dynamic Illuminations for Scene Editing

CLEAR-NeRF: Collinearity and Local-region Enhanced Accurate 3D Reconstruction in Unbounded Scenes

NaCR: Visual Localization via NeRF-aided Camera Ray Regression