LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping

Dorian Cojocaru; Mohammad-Maher Nakshbandi; Sorin Grigorescu; Ziad Sharawy

REVIEW 4 major objections 5 minor 23 references

LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping

T0 review · 4 major / 5 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 LoopDB, a loop-closure benchmark whose five-image scene sequences come with ground-truth camera rotations and translations, enabling quantitative evaluation of loop-closure detection accuracy.

desk verdict LoopDB is a real but small dataset; the benchmarking claim fails because the ground truth is unverified visual alignment and the paper's own tables contradict each other by 40x. read the letter →

arxiv 2506.06771 v1 pith:3HWQRI6U submitted 2025-06-07 cs.CV cs.RO

Mohammad-Maher Nakshbandi , Ziad Sharawy , Dorian Cojocaru , Sorin Grigorescu This is my paper

classification cs.CVcs.RO

keywords loopclosuredetectionSLAMbenchmarkingvisualplacerecognitionmulti-viewdatasetgroundtruthposefeaturematchingmonocularsubmap

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

LoopDB is offered as a benchmark for loop closure detection in SLAM, built from over 1,000 images of parks, indoor scenes, parking areas, and individual objects. Each scene is arranged as a submap of five consecutive images that revisit a common focal point from different angles, and the dataset supplies the rotation and translation between every consecutive pair as ground truth. The authors' claim is that this structure provides guaranteed loop closures with known transformations, something trajectory-based datasets do not, so detection accuracy and geometric consistency can be measured quantitatively. The paper supports the claim with feature-matching experiments showing that established descriptors produce low mean reprojection errors on the dataset's images, and that the dataset yields higher feature density than prior benchmarks.

What carries the argument

The central mechanism is the five-image submap: a sequence of five photographs of one scene sharing a consistent focal point, with the camera moved to different viewpoints. For every consecutive image pair the dataset stores a homography $H = [R \mid t]$ as a $3\times3$ matrix, with rotation represented compactly as a quaternion $q = [q_x, q_y, q_z, q_w]$ and translation as $t = [t_x, t_y, t_z]$. This object carries the argument because it turns a loop-closure event into a known transformation rather than an incidental revisit, and the five-frame chain lets errors be studied as they accumulate across transitions.

What would settle it

For a sample of submaps, independently measure the true camera poses with a high-accuracy motion-capture system or a calibration object and compare the resulting rotations and translations with LoopDB's published values. If the disagreement translates to reprojection errors on the order of the reported 7 to 36 pixel spread across descriptors, the manual ground truth is too coarse to support the benchmark's accuracy comparisons.

Watch

Extended reading notes

Core claim

The central contribution is a dataset, not an algorithm: five-image submaps centered on a shared focal point, with inter-image rotation stored as quaternions and translation as a vector, published as ground truth for every consecutive image pair. With this metadata, every transition in a submap is a known loop-closure event, so a detector's output can be checked against a known relative pose rather than inferred from a trajectory. The authors demonstrate the dataset's utility by reporting mean reprojection errors for SIFT, ORB, DAISY, and BRIEF across the four transitions of each submap, where BRIEF achieves the lowest overall error, and by comparing feature density and error against five established benchmarks.

Load-bearing premise

The dataset's usefulness as a quantitative benchmark rests on the assumption that the manually aligned camera poses, set by visual alignment rather than external pose measurements, are accurate enough to serve as ground truth; the paper gives no error bound for that alignment.

Editorial extensions

If this is right

Loop-closure detectors can be scored directly: a match between two submap frames is correct only if its estimated transformation agrees with the published rotation and translation.
The five-frame chain supports measuring how pose error accumulates across sequential transformations, simulating drift in a SLAM trajectory.
Feature-based and learning-based methods can use the submap structure to build positive and negative training pairs with known relative poses.
The scene diversity across 200+ locations lets viewpoint invariance be tested over a broader appearance range than trajectory-centric benchmarks.
Cross-dataset comparisons of feature density and reprojection error place LoopDB alongside existing benchmarks for direct quantitative comparison.

Reading between the lines

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

A natural extension the authors do not explore is using LoopDB to benchmark learned relative-pose regression, since each submap provides dense $[R,t]$ supervision rather than just a match/no-match label.
The five-image submaps could be chained into longer synthetic trajectories by composing their transformations, allowing controlled testing of full SLAM loop closure under accumulated error.
The reported cross-dataset advantages are only interpretable if the manual alignment error is itself smaller than the quoted reprojection-error differences; measuring that alignment error would sharpen every comparison the dataset enables.
The guaranteed revisits could also serve as a source of hard negatives for training place-recognition networks, since images within a submap share a focal point but differ markedly in viewpoint.

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.

Desk Editor's Note

read the letter

Colleague,

Short version: LoopDB is a real, downloadable dataset of 1,000 images organized as 200+ five-image submaps, with per-pair quaternion/translation metadata. That is new, and the multi-angle controlled capture design is a sensible niche for loop closure work. But the paper's central claim that this lets you quantitatively benchmark loop closure is not supported. The ground truth poses come from visual alignment, the paper admits that, and there is no independent measurement, no calibration report, no error bound. Since the same kind of feature matching is what your algorithms will do, the reference is not independent of the thing being tested. That is a circularity problem, not a minor one.

The experimental section makes it worse. Table II reports a mean reprojection error for SIFT around 35.95 px with max 163.5 px; Table III reports the mean SIFT error on the same dataset as 0.78 px. Both tables are labeled mean reprojection errors over LoopDB. The text gives no explanation for the factor of 40 discrepancy. One of these numbers is wrong, or the metric/subset changed silently. Either way, the validity experiment cannot be trusted.

What is good: the dataset structure is clean, the hosting links are plausible, and the authors are unusually honest in Section V-B about relying on visual alignment rather than external pose measurements. The related work is standard, and the comparison table with KITTI, EuRoC, etc. is informative, even though it makes the "large scale" claim in the title look overstated at 1,000 images.

What is missing: loop closure detection metrics. The evaluation only reports reprojection errors of feature correspondences, which says little about whether the dataset actually tests loop closure detection accuracy. And no attempt is made to verify the [R,t] against any external source, or even against a standard bundle adjustment with known baseline.

My take: this is a salvageable dataset paper, but not a benchmark paper in its current form. The fixes are concrete: validate poses with an independent method (even a calibration pattern), explain or correct the tables, and report actual loop closure detection metrics. If the metadata holds up, the dataset would be a useful small resource for viewpoint-invariance studies.

I'd send it to review rather than desk-reject, because the asset is real and the flaws are fixable, but I would not cite it or rely on the numbers until those issues are addressed.

Referee Report

4 major / 5 minor

Summary. The paper introduces LoopDB, a loop-closure benchmarking dataset of over 1000 high-resolution images organized into five-image submaps, with inter-image rotations and translations provided as ground truth. The dataset is publicly released on GitHub and Zenodo. The authors validate the dataset by evaluating SIFT, ORB, DAISY, and BRIEF descriptors in terms of feature reprojection error, both across transitions of the five-image chains and in comparison with five existing datasets. The central claim is that LoopDB enables quantitative evaluation of loop closure detection accuracy and geometric consistency.

Significance. If the ground-truth poses were accurate and the evaluation well posed, LoopDB would fill a real gap: it is specifically designed for loop closure rather than general SLAM, it is small but scene-diverse, and it is publicly released with scripts and metadata. The paper also makes a useful move by separating each scene into a controlled five-view submap, which could support studies of viewpoint-invariant place recognition. However, the central benchmarking claim is not supported in the submitted manuscript: the ground-truth poses are derived from manual visual alignment without any independent measurement or error bound, and the validation experiments are internally inconsistent. As a result, the dataset cannot currently serve as a trustworthy quantitative benchmark, and the reported comparisons do not demonstrate the claimed utility.

major comments (4)

[Section V-B] The ground-truth poses are established by 'precise manual control over the camera position' and 'based on visual alignment rather than external pose measurements'. No calibration report, uncertainty analysis, or independent verification is provided. Because the benchmark reference is generated by the same kind of visual correspondence operation that loop-closure algorithms perform, errors in the reference are unquantified and may be correlated with the errors of the algorithms being benchmarked. This directly undermines the claim, stated in Section I, that LoopDB provides 'guaranteed loop closures with known transformations' enabling 'quantitative evaluation of detection accuracy and geometric consistency'. A concrete remedy would be to validate a substantial subset of poses against an independent measurement modality (e.g., a rigid AprilTag rig with bundle adjustment, a total station, or a calibrated external motion-capture system) and to report pose error statistics per submap.
[Section VI, Tables II and III] The reported validation numbers are internally inconsistent. Table II reports LoopDB's overall mean reprojection error for SIFT as 35.95 px (with a maximum of 163.5 px), while Table III reports LoopDB's mean SIFT error as 0.78 px. The same evaluation protocol cannot produce both numbers unless the metric, image subset, or implementation changed, and the text provides no explanation. In addition, the per-transition values in Table II do not average to the stated overall means under either a simple mean or a match-weighted mean (e.g., for SIFT, the simple mean of 4.16, 163.5, 8.35, and 3.71 is 44.93, not 35.9; the match-weighted mean is approximately 30.4). This makes the validity experiment impossible to reproduce and means the quantitative benchmarking claim is not supported by the presented evidence.
[Section III-B, Eqs. (1)-(3)] The mathematical formulation of the published transformation is internally inconsistent. Equation (1) defines a 2D homography acting on homogeneous 2D points X1 = [X, Y, 1]^T, but Equation (3) defines H as a 4x4 rigid-body matrix [R t; 0 1] acting on 3D points. Applied to a 3-vector X1, the 4x4 matrix is dimensionally incompatible. Since the dataset stores quaternions and translation vectors, the paper should state clearly whether the provided ground truth is a set of relative camera poses (3D rigid transformations) or 2D planar homographies, and Equations (1)-(3) should be corrected to a consistent notation. This matters for users who may apply the metadata directly in pose-graph or feature-reprojection pipelines.
[Section VI] The validation experiments measure feature reprojection error, not loop-closure detection performance. The paper's central promise, repeated in Section I and Section IV, is a benchmark for 'detection accuracy' of loop closure algorithms. However, no precision-recall curves, recall-at-top-N metrics, or localization accuracy metrics are reported, and no loop-closure detection method is actually evaluated on the dataset. Even if the pose reference were trustworthy, the experiments as presented would only show that the images support feature matching under viewpoint change; they would not demonstrate that the dataset is suitable for quantitatively comparing loop-closure detectors. The evaluation should include at least one standard loop-closure/place-recognition metric and one representative baseline detector.

minor comments (5)

[Abstract and Section I] The abstract contains grammatical issues: 'comprising of over 1000 images' should be 'comprising over 1000 images', and 'between each consecutive images' should be 'between each pair of consecutive images'. Section I begins with 'Automonous robots', which should be 'Autonomous robots'.
[Section III-E] The sentence 'the he dataset images are available on Zenodo' contains a duplicated article and should read 'the dataset images are available on Zenodo'.
[Figure 3 caption] The caption reads 'LoobDB folder structure' and should be 'LoopDB folder structure'.
[Table I and Section II] The dataset name 'EuRoC MAV' is inconsistently typeset as 'EuRoC MA V' in Table I and in the related-work discussion. Please use a consistent spelling, e.g., 'EuRoC MAV'.
[Table II] The descriptor names are inconsistently capitalized: 'Daisy' should be 'DAISY' to match the text and Table III. Additionally, the table does not define what the numbers under 'Transition Pairs' mean precisely (e.g., 'Error/Match' with two numbers per cell is not explained in the text).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LoopDB's derivation chain does not reduce to its own inputs; the ground-truth accuracy and table inconsistencies are validity limitations, not circular reasoning.

full rationale

The paper makes no first-principles derivation; it presents a dataset. The central claim—that LoopDB provides known transformations for quantitative loop-closure benchmarking—rests on the accuracy of the [R,t] metadata. Section V-B discloses the key limitation: 'Although this ground truth is based on visual alignment rather than external pose measurements...' The paper then validates the dataset in Section VI by computing reprojection errors of SIFT, ORB, DAISY, and BRIEF against those same transformations. This is a validation-quality concern: without an independent pose reference, the reported errors measure agreement between automatic visual-feature matching and a manually/visually established alignment, and any reference error is unquantified. However, this is not a by-construction circularity under the criteria: the manual visual alignment is not defined in terms of the evaluated descriptors, and the paper does not fit a parameter to the evaluation data and then call it a prediction. A separate internal-consistency problem appears between Table II (LoopDB SIFT overall mean 35.9 px, max 163.5 px) and Table III (LoopDB SIFT mean 0.78 px) with no protocol explanation; that inconsistency undermines reproducibility but is not a circular reduction. No self-citation chain is load-bearing, and no uniqueness or ansatz is imported from the authors' prior work. Accordingly, no circular step can be exhibited at the level of Eq. = Eq. by construction, and the honest finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No fitted mathematical model or invented physical entity is claimed. The key ledger item is the manually aligned pose annotation treated as ground truth; the supporting axioms are that visual alignment is accurate, calibration is correct, reprojection error measures loop closure quality, and five re-photographed views of a scene form a loop closure.

free parameters (1)

Inter-image pose annotations (quaternion and translation) = Stored in datastream_2 metadata; exact values not listed
The pose numbers are the dataset's ground truth and are obtained by manual visual alignment, not by an independent sensor; their accuracy is an unverified input to every benchmark result on LoopDB.

assumptions (4)

ad hoc to paper Manual visual alignment produces ground-truth poses (Section V-B)
The paper relies on this to call the pose metadata ground truth, but no external measurement or error analysis is supplied.
domain assumption Camera calibration (focal length, principal point, lens distortion) is accurate (Section V)
The authors state calibration was performed but do not report calibration parameters, residuals, or validation.
domain assumption Reprojection error is a valid proxy for loop closure detection quality (Section VI)
Loop closure detection is normally evaluated with retrieval metrics such as precision and recall; the paper provides no such evaluation.
ad hoc to paper Each five-image submap constitutes a meaningful loop closure event (Section III-B)
A five-image multi-view sequence of one scene is called a loop closure, but there is no trajectory revisit or global consistency check, so the mapping to loop closure is asserted rather than demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping." pith.science (2026). https://pith.science/paper/3HWQRI6U

@misc{pith2026250606771,
  author       = {Pith},
  title        = {Pith review of: LoopDB: A Loop Closure Dataset for Large Scale Simultaneous Localization and Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3HWQRI6U}},
  note         = {Machine review of arXiv:2506.06771}
}

read the original abstract

In this study, we introduce LoopDB, which is a challenging loop closure dataset comprising over 1000 images captured across diverse environments, including parks, indoor scenes, parking spaces, as well as centered around individual objects. Each scene is represented by a sequence of five consecutive images. The dataset was collected using a high resolution camera, providing suitable imagery for benchmarking the accuracy of loop closure algorithms, typically used in simultaneous localization and mapping. As ground truth information, we provide computed rotations and translations between each consecutive images. Additional to its benchmarking goal, the dataset can be used to train and fine-tune loop closure methods based on deep neural networks. LoopDB is publicly available at https://github.com/RovisLab/LoopDB.

Figures

Figures reproduced from arXiv: 2506.06771 by the authors.

**Figure 1.** LoopDB samples (top) and data acquisition procedure (bottom). Each LoopDB sequence consists of five consecutive images, having their rotations and translations given as ground truth. II. RELATED WORK A comparison between LoopDB and the notable datasets used in computer vision research is presented in Table I. Although LoopDB is smaller (1000 images), it contains a much larger variety of scenes, as opposed to KITTI… view at source ↗

**Figure 2.** Snapshots from the LoopDB dataset (ours). Each row corresponds to a submap sequence. The rotation and translation [R, t] between two consecutive images is given as ground truth. avoid gimbal lock issues. A quaternion q is represented as q = [qx, qy, qz, qw], where qx, qy, qz are the imaginary components of the quaternion and qw is the scalar part. Quaternions are used to represent the 3D orientation of the camera a… view at source ↗

**Figure 3.** LoobDB folder structure. The data is organized as datastreams of images and their corresponding metadata. D. Metadata Information LoopDB is organized using two large metadata files in CSV format for the detailed annotation of each image, as mentioned above. The CSV files provide rotational and transformational information for each image-pair structure. All the rows are mapped to an image corresponding to the image … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 21 canonical work pages

[1]

Orb- slam: A versatile and accurate monocular slam system,

R. Mur-Artal, J. M. M. Montiel, and J. D. Tard ´os, “Orb- slam: A versatile and accurate monocular slam system,” IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147– 1163, 2015

work page 2015
[2]

The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,

K. A. Tsintotas, L. Bampis, and A. Gasteratos, “The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 11, pp. 19 929–19 953, nov 2022. [Online]. Available: http://dx.doi.org/10.1109/TITS.2022.3175656

arXiv 2022
[3]

Fab-map: Probabilistic localization and mapping in the space of appearance,

M. Cummins and P. Newman, “Fab-map: Probabilistic localization and mapping in the space of appearance,” Int. J. Robot. Res., vol. 27, no. 6, pp. 647–665, 2008

work page 2008
[4]

Visual odom- etry for ground vehicle applications,

D. Nister, O. Naroditsky, and J. Bergen, “Visual odom- etry for ground vehicle applications,”Journal of Field Robotics, vol. 23, no. 1, pp. 3–20, 2006

work page 2006
[5]

Pedes- trian detection: A benchmark,

P. Dollar, C. Wojek, B. Schiele, and P. Perona, “Pedes- trian detection: A benchmark,” inComputer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Confer- ence on. IEEE, 2009, pp. 304–311

work page 2009
[6]

Semantic object classes in video: A high-definition ground truth database,

G. J. Brostow, J. Fauqueur, and R. Cipolla, “Semantic object classes in video: A high-definition ground truth database,”Pattern Recognition Letters, vol. 30, no. 2, pp. 88–97, 2009

work page 2009
[7]

Ford campus vision and lidar data set,

G. Pandey, J. R. McBride, and R. M. Eustice, “Ford campus vision and lidar data set,”The International Journal of Robotics Research, vol. 30, no. 13, pp. 1543– 1552, 2011

work page 2011
[8]

Exploiting the power of stereo confidences,

D. Pfeiffer, S. Gehrig, and N. Schneider, “Exploiting the power of stereo confidences,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 297–304

work page 2013

Show all 23 references

[9]

The malaga urban dataset: High- rate stereo and lidar in a realistic urban scenario,

J.-L. Blanco-Claraco, F.-A. Moreno-Due ˜nas, and J. Gonz ´alez-Jim´enez, “The malaga urban dataset: High- rate stereo and lidar in a realistic urban scenario,”The International Journal of Robotics Research, vol. 33, no. 2, pp. 207–214, 2014

2014
[10]

Are we ready for autonomous driving? the kitti vision benchmark suite,

A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2012, pp. 3354–3361

2012
[11]

The euroc micro aerial vehicle datasets,

M. Burri, J. Nikolic, P. Gohl, T. Schneider, J. Rehder, S. Omari, M. Achtelik, and R. Siegwart, “The euroc micro aerial vehicle datasets,”The International Journal of Robotics Research, vol. 35, 01 2016

2016
[12]

1 year, 1000 km: The oxford robotcar dataset,

W. Maddern, G. Pascoe, C. Linegar, and P. Newman, “1 year, 1000 km: The oxford robotcar dataset,”Int. J. Robot. Res., vol. 36, no. 1, pp. 3–15, 2017

2017
[13]

4seasons: A cross- season dataset for multi-weather slam in autonomous driving,

P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers, “4seasons: A cross- season dataset for multi-weather slam in autonomous driving,” inGerman Conference on Pattern Recognition. Springer, 2020, pp. 404–417

2020
[14]

Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d,

Y . Liao, J. Xie, and A. Geiger, “Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3292–3310, 2022

2022
[15]

Distinctive image features from scale- invariant keypoints,

D. G. Lowe, “Distinctive image features from scale- invariant keypoints,”International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004

2004
[16]

Orb: An efficient alternative to sift or surf,

E. Rublee, V . Rabaud, K. Konolige, and G. Bradski, “Orb: An efficient alternative to sift or surf,” in2011 International Conference on Computer Vision. IEEE, 2011, pp. 2564–2571

2011
[17]

Brief: Binary robust independent elementary features,

M. Calonder, V . Lepetit, C. Strecha, and P. Fua, “Brief: Binary robust independent elementary features,” inEuro- pean Conference on Computer Vision. Springer, 2010, pp. 778–792

2010
[18]

Daisy: An efficient dense descriptor applied to wide-baseline stereo,

E. Tola, V . Lepetit, and P. Fua, “Daisy: An efficient dense descriptor applied to wide-baseline stereo,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 32, no. 5, pp. 815–830, 2010

2010
[19]

Mixvpr: Fea- ture mixing for visual place recognition,

A. Ali-bey, B. Chaib-draa, and P. Gigu `ere, “Mixvpr: Fea- ture mixing for visual place recognition,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 2998–3008

2023
[20]

Transvpr: Transformer-based place recognition with multi-level attention aggregation,

F. Wang, H. Cai, H. Zhou, L. Zhang, H. Ji, X. Ji, C. Gong, F. Zhu, and X. Zhang, “Transvpr: Transformer-based place recognition with multi-level attention aggregation,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2022, pp. 11 231– 11 240

2022
[21]

Superglue: Learning feature matching with graph neural networks,

P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Ra- binovich, “Superglue: Learning feature matching with graph neural networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4938–4947

2020
[22]

Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,” inAdvances in Neural Information Processing Systems, vol. 34. Curran Associates, Inc., 2021, pp. 9872–9883

2021
[23]

Deep- factors: Real-time probabilistic dense monocular slam,

J. Czarnowski, S. Leutenegger, and A. J. Davison, “Deep- factors: Real-time probabilistic dense monocular slam,” inIEEE Robotics and Automation Letters, vol. 5, no. 2. IEEE, 2020, pp. 721–728

2020

Pith tools

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

[1] [1]

Orb- slam: A versatile and accurate monocular slam system,

R. Mur-Artal, J. M. M. Montiel, and J. D. Tard ´os, “Orb- slam: A versatile and accurate monocular slam system,” IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147– 1163, 2015

work page 2015

[2] [2]

The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,

K. A. Tsintotas, L. Bampis, and A. Gasteratos, “The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 11, pp. 19 929–19 953, nov 2022. [Online]. Available: http://dx.doi.org/10.1109/TITS.2022.3175656

arXiv 2022

[3] [3]

Fab-map: Probabilistic localization and mapping in the space of appearance,

M. Cummins and P. Newman, “Fab-map: Probabilistic localization and mapping in the space of appearance,” Int. J. Robot. Res., vol. 27, no. 6, pp. 647–665, 2008

work page 2008

[4] [4]

Visual odom- etry for ground vehicle applications,

D. Nister, O. Naroditsky, and J. Bergen, “Visual odom- etry for ground vehicle applications,”Journal of Field Robotics, vol. 23, no. 1, pp. 3–20, 2006

work page 2006

[5] [5]

Pedes- trian detection: A benchmark,

P. Dollar, C. Wojek, B. Schiele, and P. Perona, “Pedes- trian detection: A benchmark,” inComputer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Confer- ence on. IEEE, 2009, pp. 304–311

work page 2009

[6] [6]

Semantic object classes in video: A high-definition ground truth database,

G. J. Brostow, J. Fauqueur, and R. Cipolla, “Semantic object classes in video: A high-definition ground truth database,”Pattern Recognition Letters, vol. 30, no. 2, pp. 88–97, 2009

work page 2009

[7] [7]

Ford campus vision and lidar data set,

G. Pandey, J. R. McBride, and R. M. Eustice, “Ford campus vision and lidar data set,”The International Journal of Robotics Research, vol. 30, no. 13, pp. 1543– 1552, 2011

work page 2011

[8] [8]

Exploiting the power of stereo confidences,

D. Pfeiffer, S. Gehrig, and N. Schneider, “Exploiting the power of stereo confidences,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 297–304

work page 2013