Pith. sign in

REVIEW 4 major objections 7 minor 68 references

Seeing in the Dark: Benchmarking Egocentric 3D Vision with the Oxford Day-and-Night Dataset

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper introduces Oxford Day-and-Night, a 30 km egocentric day-night dataset with laser-scanned ground truth, and shows that current novel view synthesis and visual relocalization methods fail on it under extreme illumination.

desk verdict A useful egocentric day/night benchmark with honest limitations, but the per-pose ground-truth validation needs to be reported before the strict-threshold relocalization numbers can be taken at face value. read the letter →

arxiv 2506.04224 v1 pith:FOHDGAW2 submitted 2025-06-04 cs.CV

Zirui Wang , Wenjing Bian , Xinghui Li , Yifu Tao , Jianeng Wang , Maurice Fallon , Victor Adrian Prisacariu This is my paper · ORCID
classification cs.CV
keywords OxfordDay-and-Nightdatasetegocentric3Dvisionday-nightilluminationnovelviewsynthesisvisualrelocalizationGaussiansplattingscenecoordinateregressionlow-lightperception
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Oxford Day-and-Night is a large-scale egocentric dataset built from smart-glasses recordings across five Oxford locations in daylight, dusk, and night, spanning 30 km of walking and 40,000 m². The paper's central claim is that this combination of natural head motion, full-color imagery, and extreme day-to-night illumination, backed by laser-scanned geometric ground truth, is missing from existing benchmarks and that current state-of-the-art novel view synthesis and visual relocalization methods fail on it. To support the claim, the authors align multi-session SLAM trajectories to millimeter-accurate terrestrial laser scans using AprilTags and ICP, then run standard NVS and relocalization pipelines on the resulting posed images and point clouds. The reported performance drops are the evidence that the dataset exposes real open problems in seeing at night.

What carries the argument

The load-bearing object is the dataset itself, built in four stages: smart-glasses capture with fisheye RGB and grayscale SLAM cameras plus dual IMUs; multi-session SLAM that fuses per-session trajectories and semi-dense point clouds into one world frame; alignment of that frame to terrestrial laser scans using AprilTag detections followed by ICP refinement, giving an average 6.7 cm point-to-point error; and spatial filtering that turns the posed images into a database/query split for relocalization (5,466 database images, 2,819 day queries, 7,179 night queries) and a temporally subsampled, point-filtered set for NVS. This machinery turns an otherwise unremarkable collection of head-mounted videos into a benchmark where ground-truth geometry and day/night pose pairs exist for two tasks at once.

What would settle it

Re-measure a subset of the dataset's camera poses independently — for example, with a handheld LiDAR-inertial rig walking the same routes — and recompute the night-time relocalization accuracies and the NVS point-to-point geometry errors using only poses that agree with the reference within a few centimeters. If the severe night-time failures persist, the central claim stands; if they vanish, the reported drops were largely pose-error artifacts.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a city-scale, egocentric, day-night benchmark with ground-truth geometry yields large and systematic failures in both novel view synthesis and visual relocalization. For relocalization, the best feature-matching methods correctly localize around 70–80% of night queries at the loosest threshold on the easiest scenes, dropping to near zero on harder ones, and scene-coordinate regression methods essentially fail at night. For NVS, two in-the-wild 3D Gaussian splatting methods produce renderings with LPIPS above 0.5 across all five scenes despite PSNR above 25, and on two scenes the trained model degenerates to fewer than 2,000 Gaussians. The paper also shows that image retrieval, not feature matching, is the main accuracy bottleneck at night: upgrading retrieval from NetVLAD to MegaLoc raises RoMA's strict-threshold accuracy from 25% to 70% on one scene, and ground-truth retrieval raises it further to about 80%.

Load-bearing premise

The multi-session SLAM trajectories supplied by the glasses' processing service are accurate enough to serve as ground-truth camera poses; the paper reports only a 6.7 cm average alignment error, without a per-pose error distribution, and every relocalization and NVS score inherits that uncertainty.

Editorial extensions

If this is right

  • Night-time relocalization is far from solved: the most accurate feature-matching method (RoMA) still fails on roughly half of night queries under a strict 0.25 m / 2° threshold on the harder outdoor scenes, and scene-coordinate regression methods drop to near zero at night.
  • Retrieval, not matching, is the current bottleneck for night localization: the same matcher with a stronger retriever (MegaLoc) nearly triples strict-threshold accuracy on the Bodleian scene, and with ground-truth retrieval reaches about 80%.
  • In-the-wild NVS methods do not yet render plausible images at city scale under extreme lighting: LPIPS stays above 0.5 everywhere even when PSNR looks acceptable, and two scenes collapse to degenerate point clouds.
  • The dataset's 7,197 night queries, 37 times the Aachen Day-Night night split, allow much tighter measurement of night-time failure rates, and its TLS geometry allows NVS geometry to be scored by point-to-point distance rather than image metrics alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the benchmark is fair, the results imply that illumination-invariant feature representations and retrieval objectives should be trained directly on day-night egocentric pairs; the roughly 1:1 day/night split makes this a natural training ground that the paper does not exploit.
  • The 6.7 cm average alignment error is probably not the whole story; a per-pose error analysis across the 30 km would tell whether some segments have drift large enough to bias the strictest (0.25 m) relocalization thresholds.
  • The point clouds aligned to TLS maps could be used to evaluate not just Gaussian center positions but also coverage and density of reconstructed scenes, so the benchmark can measure where NVS methods stop representing the scene at night.
  • A forward-looking extension would be to use the ground-truth TLS geometry as a training signal for exposure-aware or physics-based NVS models, testing whether explicit lighting models close the reported day-night gap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The manuscript introduces Oxford Day-and-Night, a large-scale egocentric dataset captured with Meta Aria glasses at five Oxford sites across day, dusk, and night conditions. The authors use multi-session SLAM (MPS) to estimate camera poses and point clouds, then align them to Oxford Spires terrestrial laser scans using AprilTag detections and ICP. The dataset supports two benchmarks: novel view synthesis (NVS), evaluated with Splatfacto-W and Gaussian-Wild, and visual relocalization, evaluated with feature-matching and scene-coordinate-regression methods. The central claim is that current state-of-the-art methods struggle on this dataset under extreme illumination, with large performance drops at night. The paper reports statistics of 30 km of trajectories, 40,000 m^2 covered area, and 7,197 night query images, and provides an analysis of retrieval as a bottleneck for relocalization.

Significance. If the ground-truth poses are validated at the claimed accuracy, this dataset fills a genuine gap: it combines egocentric color video, day-to-night illumination variation, full 6DoF motion, and TLS-aligned geometry at city scale. The night query set is substantially larger than Aachen Day-Night's, and the benchmark covers a broad range of modern methods. The paper is also honest in Section 3.6 about the dependence on MPS SLAM accuracy. However, the benchmark value rests on the per-pose accuracy of the MPS-derived ground truth, which is currently characterized only by a global cloud-to-cloud alignment error. This is the main risk to the paper's central contribution, and it affects both benchmarks.

major comments (4)
  1. [Sec. 3.5, Table 1] The validation of the ground-truth camera poses reports only the mean and median point-to-point distance between the aligned MPS point cloud and the TLS map (6.7 cm mean across sites in Table 1). This is a global cloud-to-cloud statistic, not a per-pose trajectory error, and it is not reported separately for night recordings. Since the relocalization benchmark in Table 2 uses thresholds as tight as (0.25 m, 2°), per-pose GT errors of comparable size could contaminate the reported success rates. The paper's own Section 3.6 acknowledges the dependence on MPS. Please provide a per-pose error distribution (e.g., a CDF of pose error against AprilTag-derived poses or a withheld TLS-aligned subset), separately for day, dusk, and night, and state how many poses are validated.
  2. [Sec. 4.1 and Supplementary C] The COLMAP model for the relocalization benchmark is constructed by projecting MPS point clouds using MPS poses, and the query ground-truth poses are also MPS poses. If MPS has systematic drift, that drift is shared by both the database model and the query reference, which could bias the HLoc evaluation in either direction. Because this is not a direct circularity in the derivation, the fix is empirical: report relocalization accuracy on a subset of query frames with independently obtained poses (e.g., from AprilTag detections or local TLS alignment) and compare it with the full-set numbers in Table 2.
  3. [Sec. 4.2, Table 4] The geometry metric in Table 4 is the point-to-point distance between 3DGS Gaussian centers and the TLS ground truth, but no matching or thresholding is specified, and entries marked '-' denote degenerated point clouds with fewer than 2,000 Gaussians. Comparing raw nearest distances for heavily culled point clouds is not a reliable geometry error. Specify the distance computation (e.g., CloudCompare nearest-neighbor mode), report thresholded chamfer distance or precision/recall at several distance thresholds, and discuss how culling affects the metric.
  4. [Sec. 4, Tables 2-4] All benchmark tables report a single run without repeated-run variance. Some differences are small (e.g., SP+SG vs. SP+LG on day queries in Table 2 differ by less than one percentage point in several scenes), so without variance estimates it is hard to know which differences are meaningful. Report standard deviations over at least three runs, or clearly state that only one run was performed and avoid fine-grained ranking claims.
minor comments (7)
  1. [Sec. 3.1] The text says sessions are recorded between 4-10pm, which conflicts with the day/dusk/night labels; clarify that 'day' includes late-afternoon recordings and define the time windows for each category.
  2. [Sec. 3.4] The section title contains a typo: 'Sprires' should be 'Spires'.
  3. [Sec. 3.5] The equation for Tmap,world uses Tmap,tag, Taria,tag, and Tworld,aria without defining the coordinate frames; spell out the frames for each transformation.
  4. [Table 1] State in the caption that values are in cm and that they are cloud-to-cloud point-to-point distances, not per-pose trajectory errors.
  5. [Figure 9] The y-axis label contains the typo 'Precentage'; also clarify that the y-axis is the percentage of nighttime queries having at least one database image within the given threshold.
  6. [Algorithm 1] Define the orientation difference ∠(R_i, R_j) precisely, for example as the geodesic rotation angle between the two rotation matrices.
  7. [Throughout] The naming 'Meta ARIA' vs. 'Aria' is inconsistent; standardize the capitalization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset and benchmarks are grounded in external TLS geometry and independent method evaluations, with no result that reduces to its inputs by construction.

full rationale

This is a dataset and benchmark paper rather than a derivation chain, so the circularity patterns do not apply. The central products are recorded egocentric sequences, poses and point clouds from Meta's MPS multi-session SLAM, and alignment to the Oxford Spires TLS ground truth. The paper does not derive the MPS poses from the evaluated algorithms or from the TLS map; the TLS map is an external, laser-scanned reference. Section 3.5 uses AprilTag observations and ICP to align MPS output to that map, so the ground-truth poses are anchored by independent measurements, and the reported 6.7 cm average point-to-point error is a validation statistic, not a fitted parameter renamed as a prediction. The NVS benchmark evaluates 3DGS methods against test images and TLS geometry, and the relocalization benchmark evaluates feature-matching and scene-coordinate-regression methods against query images with MPS-derived poses. Even in Table 3, 'RoMA + GT Pose 20' is presented as a retrieval upper bound using ground-truth pose proximity, not as the paper's own method. The Section 3.6 limitation that ground-truth pose accuracy ultimately depends on MPS is an honest accuracy caveat; if anything it weakens the strength of benchmark conclusions, but it does not make any result equivalent to its input by construction. The only self-citation is to Oxford Spires [7], co-authored by two of the present authors, but that citation supplies an external TLS dataset and is therefore independent support under the review rules. Because the paper is self-contained against external benchmarks and its conclusions are empirical measurements, no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated. The dataset and its split variants are artifacts, not invented entities. All free parameters are engineering choices in dataset construction; none are fitted to benchmark targets. The core assumption is that SLAM-derived trajectories, once aligned to TLS, can stand in for ground truth.

free parameters (5)
  • Spatial filtering position threshold theta_pos = 1.5 m outdoor, 0.5 m indoor
    Chosen by hand in Section 3.3 to control redundancy; directly controls database/query size and night query difficulty.
  • Spatial filtering orientation threshold theta_ori = 20 degrees
    Chosen by hand in Section 3.3; used with theta_pos and affects which images count as near-duplicates.
  • Point cloud uncertainty filter thresholds = depth std < 0.4 m; inverse depth std < 0.005 m^-1
    Chosen in Section 3.2 to clean MPS point clouds; no validation beyond visual quality.
  • NVS voxel downsampling size = 0.1 m (0.2 m for Bodleian)
    Chosen in Section 4.2 to keep GPU memory under 80 GB; changes the geometry initialization.
  • NVS temporal subsampling and image cap = 5x subsampling; about 2500 images per scene
    Chosen in Sections 3.2 and 4.2 for storage and CPU feasibility, not from task performance optimization.
assumptions (5)
  • domain assumption MPS multi-session SLAM poses and point clouds are accurate enough to serve as ground truth.
    Invoked in Sections 3.1 and 3.4; the paper's own Section 3.6 warns that accuracy depends on the SLAM service.
  • domain assumption AprilTag detections and ICP registration correctly transfer ARIA trajectories into the TLS map frame.
    Section 3.5 reports only mean and median point-to-point error; no per-trajectory error analysis is provided.
  • domain assumption Oxford Spires TLS maps are millimeter-accurate external ground truth.
    Section 3.4 imports Reference [7]; the map was produced by overlapping authors, though published separately.
  • domain assumption Point-to-point distance between Gaussian centers and TLS map is a meaningful geometry metric for NVS.
    Section 4.2 and Supplementary D use it; the authors state it is only a rough indication and can fail when Gaussians are culled.
  • domain assumption The HLoc pipeline with NetVLAD retrieval is a fair representative baseline for structure-based localization.
    Used throughout Section 4.1; standard in the field, but retrieval choices affect rankings, as Table 3 shows.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seeing in the Dark: Benchmarking Egocentric 3D Vision with the Oxford Day-and-Night Dataset." pith.science (2026). https://pith.science/paper/FOHDGAW2

@misc{pith2026250604224,
  author       = {Pith},
  title        = {Pith review of: Seeing in the Dark: Benchmarking Egocentric 3D Vision with the Oxford Day-and-Night Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FOHDGAW2}},
  note         = {Machine review of arXiv:2506.04224}
}
abstract

We introduce Oxford Day-and-Night, a large-scale, egocentric dataset for novel view synthesis (NVS) and visual relocalisation under challenging lighting conditions. Existing datasets often lack crucial combinations of features such as ground-truth 3D geometry, wide-ranging lighting variation, and full 6DoF motion. Oxford Day-and-Night addresses these gaps by leveraging Meta ARIA glasses to capture egocentric video and applying multi-session SLAM to estimate camera poses, reconstruct 3D point clouds, and align sequences captured under varying lighting conditions, including both day and night. The dataset spans over 30 $\mathrm{km}$ of recorded trajectories and covers an area of 40,000 $\mathrm{m}^2$, offering a rich foundation for egocentric 3D vision research. It supports two core benchmarks, NVS and relocalisation, providing a unique platform for evaluating models in realistic and diverse environments.

Figures

Figures reproduced from arXiv: 2506.04224 by the authors.

Figure 1
Figure 1. Overview of the Oxford-Day-and-Night Dataset at Example Scene Bodleian. Our dataset captures egocentric sequences across five locations in Oxford under diverse lighting conditions using Meta ARIA glasses. Top-left: Sample fisheye camera views across day and night recordings. Bottom-left: multi-session SLAM points aligned with high-quality laser ground truth. Right: Multi￾session SLAM trajectories visualized on a sat… view at source ↗
Figure 2
Figure 2. Example Frames Captured at Different Lighting Conditions. The severe degradation in visual quality from day to night highlights the difficulty of consistent scene understanding, posing significant challenges for both novel view synthesis (NVS) and visual relocalization methods. 2 Related Work 3D Reconstruction Datasets. Evaluating 3D reconstruction algorithms relies on accurate ground truth 3D models, which are typi… view at source ↗
Figure 3
Figure 3. Data Collection and Processing Pipeline. At a collection site, our pipeline starts with a) capturing 2–10 minute videos using ARIA glasses under varying lighting conditions. These multi-session recordings are processed using b) the MPS SLAM system to generate point clouds and camera trajectories in a unified coordinate frame. The colors of the points and trajectories represent different recording sessions; c) Levera… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: ARIA MPS Quality Assessment. We leverage Frontier and AprilTag to align ARIA recordings to TLS ground truth map. a) The Frontier handheld perception unit, equipped with three wide FoV cameras and a 64-channel LiDAR; b) A snapshot of an AprilTag; c) ARIA trajectories al…
Figure 5
Figure 5. Figure 5: NVS In-the-Wild Results in the H.B. Allen Centre (top) and Bodleian Library (bottom). Compared to Gaussian-Wild, Splatfacto-W performs better at the H.B. Allen Centre but fails at the Bodleian Library. Although Gaussian-Wild produces some renderings with recognizable c…
Figure 6
Figure 6. Figure 6: Our dataset covers 40,000 m2 area. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Our dataset spans 30 kilometers of walking trajectory. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: We provide three image types: the original fisheye, a [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: The percentage of nighttime queries that have a database image given a spatial and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Camera poses for visual relocalization in each scene. The cameras of database images are in black; the cameras of day query images are in orange and the cameras of night query images are in blue. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Visualization of 3D Geometry. In c) and d), less than 2000 Gaussian primitives remain after the culling process during training. This may be due to limited capability in handling large-scale scenes and dramatic light variations, resulting in a degenerated case for 3DG…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 63 canonical work pages

  1. [1]

    1 year, 1000 km: The oxford robotcar dataset.IJRR, 2017

    Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, 1000 km: The oxford robotcar dataset.IJRR, 2017. 1, 4

    work page 2017

  2. [2]

    Visual topometric localization

    Hernán Badino, Daniel Huber, and Takeo Kanade. Visual topometric localization. InIEEE Intelligent V ehicles Symposium, 2011. 1, 4

    work page 2011

  3. [3]

    Posenet: A convolutional network for real-time 6-dof camera relocalization

    Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. InICCV, 2015. 1, 4

    work page 2015

  4. [4]

    Inloc: Indoor visual localization with dense matching and view synthesis

    Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In CVPR, 2018. 1, 3, 5

    work page 2018

  5. [5]

    Benchmarking 6dof outdoor visual localization in changing conditions

    Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. InCVPR, 2018. 1, 4, 5, 8, 15

    work page 2018

  6. [6]

    Schönberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys

    Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schönberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking localization and mapping for augmented reality. InECCV, 2022. 1, 4

    work page 2022

  7. [7]

    The oxford spires dataset: Benchmarking large-scale lidar-visual localisation, reconstruction and radiance field methods.IJRR, 2025

    Yifu Tao, Miguel Ángel Muñoz-Bañón, Lintong Zhang, Jiahao Wang, Lanke Frank Tarimo Fu, and Maurice Fallon. The oxford spires dataset: Benchmarking large-scale lidar-visual localisation, reconstruction and radiance field methods.IJRR, 2025. 1, 2, 3, 6, 8

    work page 2025

  8. [8]

    Matterport3D: Learning from RGB-D data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In3DV, 2017. 3

    work page 2017

Show all 68 references
  1. [9]

    Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. InCVPR, 2017. 3

    2017

  2. [10]

    The EuRoC micro aerial vehicle datasets.IJRR, 2016

    Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets.IJRR, 2016. 3

    2016

  3. [11]

    Large-scale data for multiple-view stereopsis.IJCV, 2016

    Henrik Aanæs, Rasmus Ramsbøl Jensen, George V ogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis.IJCV, 2016. 3

    2016

  4. [12]

    A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM

    Ankur Handa, Thomas Whelan, John McDonald, and Andrew J Davison. A benchmark for RGB-D visual odometry, 3D reconstruction and SLAM. InICRA, 2014. 3

    2014

  5. [13]

    A comparison and evaluation of multi-view stereo reconstruction algorithms

    Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. InCVPR, volume 1, pages 519–528. IEEE,

  6. [14]

    A multi-view stereo benchmark with high-resolution images and multi- camera videos

    Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi- camera videos. InCVPR, 2017. 3

    2017

  7. [15]

    Tanks and Temples: Benchmarking large-scale scene reconstruction.ToG, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction.ToG, 2017. 3

    2017

  8. [16]

    ScanNet++: A high-fidelity dataset of 3D indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. InICCV, pages 12–22, 2023. 3

    2023

  9. [17]

    The Newer College dataset: Handheld lidar, inertial and vision with ground truth

    Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The Newer College dataset: Handheld lidar, inertial and vision with ground truth. InIROS, 2020. 3, 6

    2020

  10. [18]

    Hilti-Oxford dataset: A millimeter-accurate benchmark for simultaneous localization and mapping.RAL, 2022

    Lintong Zhang, Michael Helmberger, Lanke Frank Tarimo Fu, David Wisth, Marco Camurri, Davide Scaramuzza, and Maurice Fallon. Hilti-Oxford dataset: A millimeter-accurate benchmark for simultaneous localization and mapping.RAL, 2022. 3

    2022

  11. [19]

    FusionPortableV2: A unified multi-sensor dataset for generalized SLAM across diverse platforms and scalable environments.IJRR, 2024

    Hexiang Wei, Jianhao Jiao, Xiangcheng Hu, Jingwen Yu, Xupeng Xie, Jin Wu, Yilong Zhu, Yuxuan Liu, Lujia Wang, and Ming Liu. FusionPortableV2: A unified multi-sensor dataset for generalized SLAM across diverse platforms and scalable environments.IJRR, 2024. 3 10

    2024

  12. [20]

    Sun, Hongwei Yu, Chun Liu, Long Chen, Wei Tao, and Hui Zhao

    Yuanzhi Liu, Yujia Fu, Minghui Qin, Yufeng Xu, Baoxin Xu, Fengdong Chen, Bart Goossens, Poly Z.H. Sun, Hongwei Yu, Chun Liu, Long Chen, Wei Tao, and Hui Zhao. BotanicGarden: A high-quality dataset for robot navigation in unstructured natural environments.RAL, 2024. 3

    2024

  13. [21]

    MCD: Diverse large-scale multi-campus dataset for robot perception

    Thien-Minh Nguyen, Shenghai Yuan, Thien Hoang Nguyen, Pengyu Yin, Haozhi Cao, Lihua Xie, Maciej Wozniak, Patric Jensfelt, Marko Thiel, Justin Ziegenbein, and Noel Blunder. MCD: Diverse large-scale multi-campus dataset for robot perception. InCVPR, 2024. 3

    2024

  14. [22]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3

    2015 arXiv

  15. [23]

    Stereo magnification: learning view synthesis using multiplane images.ToG, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images.ToG, 2018. 3

    2018

  16. [24]

    Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar

    Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ToG, 2019. 3

    2019

  17. [25]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InECCV, 2020. 3

    2020

  18. [26]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields.CVPR, 2022. 3

    2022

  19. [27]

    Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. InICCV, 2021. 3

    2021

  20. [28]

    Objaverse-xl: A universe of 10m+ 3d objects

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram V oleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects. NeurIPS, 2023. 3

    2023

  21. [29]

    Image matching across wide baselines: From paper to practice.IJCV, 2021

    Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice.IJCV, 2021. 3

    2021

  22. [30]

    Megascenes: Scene-level view synthesis at scale

    Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. InECCV, 2024. 3

    2024

  23. [31]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InCVPR, 2024. 3

    2024

  24. [32]

    Scene coordinate regression forests for camera relocalization in rgb-d images

    Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. InCVPR, 2013. 3

    2013

  25. [33]

    Learning to navigate the energy landscape

    Julien Valentin, Angela Dai, Matthias Nießner, Pushmeet Kohli, Philip Torr, Shahram Izadi, and Cem Keskin. Learning to navigate the energy landscape. In3DV, 2016. 3

    2016

  26. [34]

    Learning to detect scene landmarks for camera localization

    Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. InCVPR, 2022. 3

    2022

  27. [35]

    Large-scale localization datasets in crowded indoor spaces

    Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guérin, Gabriela Csurka, et al. Large-scale localization datasets in crowded indoor spaces. InCVPR, 2021. 3

    2021

  28. [36]

    Vision meets robotics: The kitti dataset.IJRR, 2013

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.IJRR, 2013. 4

    2013

  29. [37]

    The epic-kitchens dataset: Collection, challenges and baselines.TPAMI, 2020

    Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. The epic-kitchens dataset: Collection, challenges and baselines.TPAMI, 2020. 4

    2020

  30. [38]

    In the eye of beholder: Joint learning of gaze and actions in first person video

    Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. InECCV, 2018. 4

    2018

  31. [39]

    Introducing hot3d: An egocentric dataset for 3d hand and object tracking

    Prithviraj Banerjee, Sindi Shkodrani, Pierre Moulon, Shreyas Hampali, Fan Zhang, Jade Fountain, Edward Miller, Selen Basol, Richard Newcombe, Robert Wang, et al. Introducing hot3d: An egocentric dataset for 3d hand and object tracking. InCVPR, 2025. 4 11

    2025

  32. [40]

    Ego4d: Around the world in 3,000 hours of egocentric video

    Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. InCVPR, 2022. 4

    2022

  33. [41]

    Egovid-5m: A large-scale video-action dataset for egocentric video generation

    Xiaofeng Wang, Kang Zhao, Feng Liu, Jiayu Wang, Guosheng Zhao, Xiaoyi Bao, Zheng Zhu, Yingya Zhang, and Xingang Wang. Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380, 2024. 4

    2024 arXiv

  34. [42]

    Aria digital twin: A new benchmark dataset for egocentric 3d machine perception

    Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. InICCV, 2023. 4

    2023

  35. [43]

    Aria everyday activities dataset.arXiv preprint arXiv:2402.13349, 2024

    Zhaoyang Lv, Nicholas Charron, Pierre Moulon, Alexander Gamino, Cheng Peng, Chris Sweeney, Edward Miller, Huixuan Tang, Jeff Meissner, Jing Dong, et al. Aria everyday activities dataset.arXiv preprint arXiv:2402.13349, 2024. 4

    2024 arXiv

  36. [44]

    Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives

    Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. InCVPR, 2024. 4

    2024

  37. [45]

    3d gaussian splatting for real-time radiance field rendering.ToG, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ToG, 2023. 5

    2023

  38. [46]

    Splatfacto-w: A nerfstudio implementation of gaussian splatting for unconstrained photo collections.arXiv preprint arXiv:2407.12306, 2024

    Congrong Xu, Justin Kerr, and Angjoo Kanazawa. Splatfacto-w: A nerfstudio implementation of gaussian splatting for unconstrained photo collections.arXiv preprint arXiv:2407.12306, 2024. 5, 8, 9, 18

    2024 arXiv

  39. [47]

    Gaussian in the wild: 3d gaussian splatting for unconstrained image collections

    Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. InECCV, 2024. 5, 8, 9, 18

    2024

  40. [48]

    From coarse to fine: Robust hierarchical localization at large scale

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. InCVPR, 2019. 5, 7, 16

    2019

  41. [49]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InCVPR, 2020. 5, 7

    2020

  42. [50]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. InCVPR, 2021. 5, 7

    2021

  43. [51]

    Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses

    Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. InCVPR, 2023. 5, 7

    2023

  44. [52]

    Glace: Global local accelerated coordinate encoding

    Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph V ogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. InCVPR, 2024. 5, 7

    2024

  45. [53]

    AprilTag: A robust and flexible visual fiducial system

    Edwin Olson. AprilTag: A robust and flexible visual fiducial system. InICRA, 2011. 6

    2011

  46. [54]

    AprilTag 2: Efficient and robust fiducial detection

    John Wang and Edwin Olson. AprilTag 2: Efficient and robust fiducial detection. InIROS, 2016. 6

    2016

  47. [55]

    Extrinsic calibration of camera to lidar using a differentiable checkerboard model

    Lanke Frank Tarimo Fu, Nived Chebrolu, and Maurice Fallon. Extrinsic calibration of camera to lidar using a differentiable checkerboard model. InIROS, 2023. 6

    2023

  48. [56]

    Besl and Neil D

    P.J. Besl and Neil D. McKay. A method for registration of 3-d shapes.TPAMI, 1992. 6

    1992

  49. [57]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InCVPR, 2016. 7, 8

    2016

  50. [58]

    Object recognition from local scale-invariant features

    David G Lowe. Object recognition from local scale-invariant features. InICCV, 1999. 7

    1999

  51. [59]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. InCVPR, 2023. 7

    2023

  52. [60]

    Disk: Learning local features with policy gradient

    Michał Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. In NeurIPS, 2020. 7

    2020

  53. [61]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. InCVPR, 2024. 7

    2024

  54. [62]

    Grounding image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In ECCV, 2024. 7 12

    2024

  55. [63]

    R-score: Revisiting scene coordinate regression for robust large-scale visual localization

    Xudong Jiang, Fangjinhua Wang, Silvano Galliani, Christoph V ogel, and Marc Pollefeys. R-score: Revisiting scene coordinate regression for robust large-scale visual localization. InCVPR, 2025. 7

    2025

  56. [64]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. InCVPRW, 2018. 7

    2018

  57. [65]

    Gordo, J

    A. Gordo, J. Almazan, J. Revaud, and D. Larlus. End-to-end learning of deep visual representations for image retrieval.IJCV, 2017. 8

    2017

  58. [66]

    Revaud, J

    J. Revaud, J. Almazan, R.S. Rezende, and C.R. de Souza. Learning with average precision: Training image retrieval with a listwise loss. InICCV, 2019. 8

    2019

  59. [67]

    Self-supervising fine-grained region similarities for large-scale image localization

    Yixiao Ge, Haibo Wang, Feng Zhu, Rui Zhao, and Hongsheng Li. Self-supervising fine-grained region similarities for large-scale image localization. InECCV, 2020. 8

    2020

  60. [68]

    Megaloc: One retrieval to place them all.arXiv preprint arXiv:2502.17237, 2025

    Gabriele Berton and Carlo Masone. Megaloc: One retrieval to place them all.arXiv preprint arXiv:2502.17237, 2025. 8 13 Seeing in the Dark: Benchmarking Egocentric 3D Vision with the Oxford Day-and-Night Dataset (Supplementary) https://oxdan.active.vision/ A Full Dataset Statis...

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.