Pith. sign in

REVIEW 5 major objections 6 minor 47 references

ThermalLoc: A Vision Transformer-Based Approach for Robust Thermal Camera Relocalization in Large-Scale Environments

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a hybrid CNN-Transformer network that regresses metric 6-DoF poses from single thermal images at city scale.

desk verdict First city-scale single-image thermal relocalization with a plausible method, but the headline IF-1 number is partly a product of model selection on that test set. read the letter →

arxiv 2506.18268 v1 pith:RRWDSLT2 submitted 2025-06-23 cs.CV

Yu Liu , Yangtao Meng , Xianfei Pan , Jie Jiang , Changhao Chen This is my paper
classification cs.CV
keywords thermalcamerarelocalizationabsoluteposeregressionvisiontransformerEfficientNet6-DoFestimationcity-scalelocalizationimagingdeeplearning
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

The paper proposes ThermalLoc, an end-to-end neural network that estimates a camera's six-degree-of-freedom pose from a single thermal image, and claims it is the first to do this at city scale. On five kilometer-scale scenarios, including the public STheReO Valley sequence, ThermalLoc reports an average mean position error of 4.23 meters, outperforming PoseNet, MapNet, AtLoc, and RobustLoc. The motivation is that thermal cameras, unlike visible-light cameras, measure heat emission and therefore keep working in darkness, fog, smoke, and glare. If the claim holds, reliable localization needs only one monocular 8-bit thermal camera plus a few milliseconds of compute per 10 Hz frame.

What carries the argument

The load-bearing object is the hybrid feature extractor: EfficientNet-B0 supplies local features through MBConv blocks with depthwise convolutions and squeeze-and-excitation units, and a six-layer Transformer — modified to compute queries, keys, and values from a single linear layer and to omit masked attention — associates those features globally. The 'Shape-First' bridging operation, which rearranges the CNN feature map to Transformer shape before embedding, is shown in ablations to outperform patch-first and ViNT-style masked connections. Two MLP heads regress 3-D position and a 4-D quaternion, trained with an L1 loss on position and log-quaternion weighted by learnable parameters $\beta$ and $\gamma$. The preprocessing equations $P' = aP + b$ and $T' = P' + h(P' - P' * G)$ stretch contrast and sharpen edges before the network sees the image.

What would settle it

Train ThermalLoc on the morning and evening Valley sequences and test on a newly collected midday sequence of the same route using the same equipment and preprocessing; if mean position error rises to the baseline level (about 5 m or more), the claimed all-day robustness would be contradicted. A cheaper check is to rerun the paper's exact splits with the sharpening step disabled and measure how much of the 4.23 m average depends on preprocessing.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid network — an EfficientNet-B0 convolutional backbone feeding a six-layer Vision Transformer, followed by two MLP heads — can regress metric position and orientation from single thermal images at kilometer scale with accuracy that beats existing learned relocalization models. The paper reports average mean position errors of 4.23 m and mean rotation errors of 7.84° across five scenarios, compared with 4.98 m and 8.92° for the strongest baseline, RobustLoc. The gain is attributed to the simultaneous extraction of local features by EfficientNet and global scene associations by the Transformer, enabled by a 'Shape-First' connectivity step in which the feature map is rearranged to Transformer dimensions before embedding. The paper also introduces a preprocessing chain, linear contrast and brightness stretch followed by Gaussian low-pass sharpening, which it says is necessary to make low-resolution, low-texture 8-bit thermal images usable for learned pose regression.

Load-bearing premise

The method assumes that a single thermal image carries enough stable information to determine its metric pose, and that the scene's thermal appearance at test time is close enough to the training trajectories for a learned regressor to generalize; the paper gives no analysis of diurnal thermal drift or non-uniformity-correction artifacts.

Editorial extensions

If this is right

  • If the reported accuracy holds, thermal-only relocalization becomes a practical fallback for autonomous vehicles and robots in darkness, fog, smoke, or glare, where RGB cameras and LiDAR degrade.
  • The 6 ms per-frame inference time with a 10 Hz thermal camera means the method can run in real time on commodity GPUs.
  • Because the method regresses absolute pose from a single image without temporal filtering, it can reinitialize localization after track loss caused by the periodic non-uniformity correction that thermal cameras require.
  • The claim of state-of-the-art performance implies that hybrid CNN-Transformer absolute pose regression transfers from RGB to the low-texture thermal modality without modality-specific geometric priors.

Reading between the lines

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

  • The paper does not analyze diurnal thermal drift: the Valley split trains on morning and evening sequences and tests on afternoon, but no analysis of solar-heating appearance shift or NUC artifacts is given, so the true day-to-day generalization remains untested.
  • A natural extension is to fuse ThermalLoc's single-image pose prior with a temporal filter or thermal-odometry constraint; the paper does not explore this, but such fusion could reduce outlier poses and smooth the trajectory.
  • The reported margin over RobustLoc is about 0.75 m on average; an independent re-implementation on the same splits, with the same preprocessing, would determine whether the gain is architectural or driven by the preprocessing choices.
  • If the preprocessing chain is the main driver, the same contrast-stretch and sharpening steps could improve RGB-based APR baselines; the paper does not test that cross-modal transfer, so part of the reported advantage may be preprocessing rather than the model.
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

5 major / 6 minor

Summary. The paper introduces ThermalLoc, an end-to-end absolute pose regression method for thermal camera relocalization. The architecture combines an EfficientNet-B0 backbone with a six-layer Transformer and two MLP heads that regress a 3-D position and a 4-D quaternion from a single 8-bit thermal image. The method also includes a preprocessing step that applies linear brightness/contrast stretching and Gaussian low-pass sharpening. The authors evaluate ThermalLoc on four self-collected urban thermal sequences (IF-1 to IF-4) and on the Valley sequence of the public STheReO dataset, comparing against PoseNet, AtLoc, MapNet (claimed but not reported), and RobustLoc. The reported average mean position error is 4.23 m for ThermalLoc versus 4.98 m for RobustLoc, and the paper claims state-of-the-art accuracy and robustness for monocular thermal camera relocalization in large-scale scenes.

Significance. If the empirical claims hold, the paper makes a useful contribution by demonstrating that a hybrid CNN-Transformer architecture with simple image preprocessing can regress metric 6-DoF poses from single thermal images at city scale, an area that is indeed underexplored compared with RGB-based relocalization. The ablation study on the Transformer depth, backbone choice, and bridging strategy provides useful engineering insight. However, the strength of the central claim is limited by the evaluation protocol: model selection is performed on the test set, MapNet results are missing, and no repeated-run statistics are provided. The paper also does not release code or data, which further impedes verification. The claimed 15.1% average position-error improvement over RobustLoc is modest, so the comparison protocol is load-bearing for the state-of-the-art claim.

major comments (5)
  1. [Section IV-C and Tables III-V] The model configuration is selected using the IF-1 scenario: Table III chooses the Transformer depth and mask design, Table IV chooses EfficientNet over ResNet, and Table V chooses the Shape-First bridging module, all on IF-1. The same IF-1 test trajectory is then used in Table II to report the headline 4.18 m result. Because the reported advantage over RobustLoc is only 1.16 m on IF-1 and 0.75 m on average, the 'significantly outperforming' claim is partly a selection artifact. Please re-run the evaluation with a held-out validation split (or nested cross-validation) and report Table II using only configurations chosen without access to the corresponding test trajectory.
  2. [Abstract and Section IV-B] The abstract and the comparison list in Section IV-B state that ThermalLoc outperforms MapNet, but Table II contains no MapNet rows and no MapNet results appear anywhere in the paper. This part of the claim is therefore unverifiable. Either add MapNet results obtained with the same training and testing protocol, or remove MapNet from the list of compared methods and adjust the abstract accordingly.
  3. [Section IV-A and Table II] Only a single run is reported for each method, with no error bars, no seed variance, and no statistical significance test. APR training is stochastic, and some reported margins are very small (e.g., IF-2 mean position error is 3.50 m for RobustLoc versus 3.48 m for ThermalLoc). The paper should report mean and standard deviation over at least three to five seeds, and ideally a paired significance test, before claiming that ThermalLoc is significantly more accurate than RobustLoc.
  4. [Equations (1)-(2) and Section III-A] The preprocessing method depends on parameters a, b, h, and sigma, but the manuscript never states their values or analyzes their sensitivity. Since preprocessing is presented as a contribution and all experimental results depend on it, the paper needs to specify the exact parameter settings and provide an ablation or sensitivity study showing how the reported errors vary with these parameters.
  5. [Section IV-B] The paper states that all algorithms are trained and tested on the same dataset, but it does not describe how PoseNet, AtLoc, and RobustLoc were instantiated, which hyperparameters were used, or whether public implementations were employed. Without this information, the comparison could be biased against the baselines. Please provide the full training recipe for each baseline, including architecture details, optimizer settings, and any per-scene hyperparameter tuning.
minor comments (6)
  1. [Figure 3 caption] The caption says the figure shows four scenarios, but it actually lists five (IF-1, IF-2, IF-3, IF-4, and Valley). Please correct the wording.
  2. [References] Reference [48] duplicates reference [14]; both cite the same MapNet paper. Please remove the duplicate.
  3. [Equation (1)] The sentence introducing Equation (1) is grammatically incomplete ('define raw thermal image T'). Please clarify the notation for P, P', T, and T'.
  4. [Section III-C] The paper states that quaternions are constrained to a single hemisphere to ensure a unique representation, but it does not explain how this constraint is enforced during training. Please specify the implementation.
  5. [Table II and Section IV-B] The discussion focuses almost entirely on position error, but Table II also reports orientation errors. Some orientation errors are large (e.g., roughly 14-16 degrees on IF-2 and IF-4 for all methods), and the paper should discuss whether the orientation results actually support the claimed robustness advantage.
  6. [Section IV-A and Table I] Table I lists the Valley sequence as 'All-day' while Section IV-A says the morning and evening sequences are used for training and the afternoon sequence for testing. Please make the season/lighting description consistent and specify which part of Valley is the test split.

Circularity Check

1 steps flagged · score 6.0 of 10

The IF-1 SOTA result is selected on the IF-1 test trajectory: the reported 4.18 m is the argmin of the ablations, so the headline margin over RobustLoc is partly fitted.

  1. fitted input called prediction [Section IV-C (Ablation Study), Tables III-V and Table II / Section IV-B]
    "In this part, we conducted an ablation study to evaluate the impact of different architectural components on the ThermalLoc model primarily using the IF-1 scenario from our thermal dataset. ... The results indicate that the Transformer module without masking performs better in thermal image relocalization, and a 6-layer Transformer strikes an optimal balance ... [Table III row:] IF-1 ... Depth=6 (ThermalLoc) ... 4.18m,1.93° [Table II row:] IF-1 ... ThermalLoc ... 4.18m,1.93°"

    The paper selects the Transformer depth (Table III), the backbone (Table IV), and the bridging module (Table V) by comparing mean errors on the IF-1 scenario, then reports the winner's IF-1 error (4.18 m, 1.93°) in Table II as the final IF-1 result. The same number used to choose the model is the number presented as the outcome, so the IF-1 entry in the SOTA comparison is the argmin of a model search over that very test trajectory rather than an independent evaluation. Because the 'Average' row and the 15.1% improvement over RobustLoc include this fitted IF-1 point, part of the headline accuracy claim is manufactured by test-set selection. The other scenarios and Valley remain held out, so the circularity is partial, not total.

full rationale

The clearest circular step is the IF-1 ablation protocol: Transformer depth, EfficientNet-vs-ResNet choice, and Shape-First bridging are all chosen using IF-1 test errors, and the same IF-1 test errors are then reported as the method's result. This fits the 'fitted input called prediction' pattern for the IF-1 scenario and, by extension, for the five-scenario average that drives the '15.1% improvement over RobustLoc' claim. I do not score this higher because the held-out scenes and the public STheReO Valley sequence still favor ThermalLoc, so the central claim has independent content; the issue is a partially contaminated benchmark rather than a derivation that is circular by definition. No load-bearing self-citation, uniqueness theorem, or ansatz-smuggling-by-citation is present. MapNet being named in the abstract but absent from Table II is an inconsistency, but it is not a circularity and does not affect this score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities, forces, or conserved quantities. Its load-bearing assumptions are standard APR domain assumptions: single-image metric pose recoverability and train/test distribution similarity. The free parameters are preprocessing settings, Transformer depth, and loss weights, none of which are justified independently of the target data.

free parameters (3)
  • Preprocessing parameters (a, b, h, sigma) = Not reported
    Equations (1) and (2) define a linear brightness/contrast stretch and a sharpening step with h and sigma as manual settings. The values are never given, so exact preprocessing is unreproducible.
  • Transformer depth = 6
    Selected as best on IF-1 test error in Table III; depth is a hyperparameter that directly affects the reported accuracy.
  • Loss balancing weights (beta, gamma) = beta0=-3.0, gamma0=0.0, learned
    Learnable parameters in the L1 pose loss (Eq. 9); initialized as stated and fitted during training.
assumptions (4)
  • domain assumption A single monocular thermal image contains sufficient metric information to regress an absolute 6-DoF pose in a known environment.
    The entire APR pipeline assumes this; no temporal or multi-view constraints are used.
  • domain assumption The test trajectory's thermal appearance is similar enough to the training trajectories for a learned regressor to generalize.
    Train/test splits differ by time of day (e.g., Valley morning/evening vs afternoon), and thermal signatures shift with heating; no domain adaptation is applied.
  • domain assumption RTK ground truth is accurate to about 0.1 m and serves as an unbiased regression target.
    Section IV.A.2 reports RTK accuracy; label noise this small is acceptable but not independently verified.
  • domain assumption ImageNet-pretrained EfficientNet features transfer usefully to 8-bit grayscale thermal images.
    The network is initialized with ImageNet weights despite the RGB-to-thermal domain gap; the paper does not test alternative initializations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ThermalLoc: A Vision Transformer-Based Approach for Robust Thermal Camera Relocalization in Large-Scale Environments." pith.science (2026). https://pith.science/paper/RRWDSLT2

@misc{pith2026250618268,
  author       = {Pith},
  title        = {Pith review of: ThermalLoc: A Vision Transformer-Based Approach for Robust Thermal Camera Relocalization in Large-Scale Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRWDSLT2}},
  note         = {Machine review of arXiv:2506.18268}
}
read the original abstract

Thermal cameras capture environmental data through heat emission, a fundamentally different mechanism compared to visible light cameras, which rely on pinhole imaging. As a result, traditional visual relocalization methods designed for visible light images are not directly applicable to thermal images. Despite significant advancements in deep learning for camera relocalization, approaches specifically tailored for thermal camera-based relocalization remain underexplored. To address this gap, we introduce ThermalLoc, a novel end-to-end deep learning method for thermal image relocalization. ThermalLoc effectively extracts both local and global features from thermal images by integrating EfficientNet with Transformers, and performs absolute pose regression using two MLP networks. We evaluated ThermalLoc on both the publicly available thermal-odometry dataset and our own dataset. The results demonstrate that ThermalLoc outperforms existing representative methods employed for thermal camera relocalization, including AtLoc, MapNet, PoseNet, and RobustLoc, achieving superior accuracy and robustness.

Figures

Figures reproduced from arXiv: 2506.18268 by the authors.

Figure 1
Figure 1. In our ThermalLoc model, we trained an efficient Transformer [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Thermal image preprocessing: the processed thermal image (below) is compared with the original thermal image (above). The processed [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. A comparison of the trajectories generated by PoseNet, MapNet, AtLoc, RobustLoc, and ThermalLoc across four scenarios (from top to bottom: [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [1]

    Monocular vision for mobile robot localization and autonomous navigation,

    E. Royer, M. Lhuillier, M. Dhome, and J.-M. Lavest, “Monocular vision for mobile robot localization and autonomous navigation,” International Journal of Computer Vision, vol. 74, no. 3, pp. 237– 260, 2007

    work page 2007

  2. [2]

    Video-rate localization in multiple maps for wearable augmented reality,

    R. Castle, G. Klein, and D. W. Murray, “Video-rate localization in multiple maps for wearable augmented reality,” in2008 12th IEEE International Symposium on Wearable Computers. IEEE, 2008, pp. 15–22

    work page 2008

  3. [3]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE Transactions on Robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  4. [4]

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

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “Orb-slam: a versatile and accurate monocular slam system,”IEEE transactions on robotics, vol. 31, no. 5, pp. 1147–1163, 2015

    work page 2015

  5. [5]

    Lsd-slam: Large-scale di- rect monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale di- rect monocular slam,” inEuropean conference on computer vision. Springer, 2014, pp. 834–849

    2014

  6. [6]

    Dsac-differentiable ransac for camera localization,

    E. Brachmann, A. Krull, S. Nowozin, J. Shotton, F. Michel, S. Gumhold, and C. Rother, “Dsac-differentiable ransac for camera localization,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 6684–6692

    work page 2017

  7. [7]

    Loam: Lidar odometry and mapping in real-time

    J. Zhang, S. Singhet al., “Loam: Lidar odometry and mapping in real-time.” inRobotics: Science and systems, vol. 2, no. 9. Berkeley, CA, 2014, pp. 1–9

    2014

  8. [8]

    Pointloc: Deep pose regressor for lidar point cloud localization,

    W. Wang, B. Wang, P. Zhao, C. Chen, R. Clark, B. Yang, A. Markham, and N. Trigoni, “Pointloc: Deep pose regressor for lidar point cloud localization,”IEEE Sensors Journal, vol. 22, no. 1, pp. 959–968, 2021

    2021

Show all 47 references
  1. [9]

    Hypliloc: Towards effective lidar pose regression with hyperbolic fusion,

    S. Wang, Q. Kang, R. She, W. Wang, K. Zhao, Y . Song, and W. P. Tay, “Hypliloc: Towards effective lidar pose regression with hyperbolic fusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 5176–5185

    2023

  2. [10]

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

    A. Kendall, M. Grimes, and R. Cipolla, “Posenet: A convolutional network for real-time 6-dof camera relocalization,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 2938–2946

    2015

  3. [11]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 1–9

    2015

  4. [12]

    Modelling uncertainty in deep learning for camera relocalization,

    A. Kendall and R. Cipolla, “Modelling uncertainty in deep learning for camera relocalization,” in2016 IEEE international conference on Robotics and Automation (ICRA). IEEE, 2016, pp. 4762–4769

    2016

  5. [13]

    Image-based localization using lstms for structured fea- ture correlation,

    F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers, “Image-based localization using lstms for structured fea- ture correlation,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 627–637

    2017

  6. [15]

    Learning multi-scene absolute pose regression with transformers,

    Y . Shavit, R. Ferens, and Y . Keller, “Learning multi-scene absolute pose regression with transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2733–2742

    2021

  7. [16]

    Darkloc+: Thermal image-based indoor localization for dark environments with relative geometry constraints,

    B. Zhou, Y . Xiao, Q. Li, C. Sun, B. Wang, L. Pan, D. Zhang, J. Zhu, and Q. Li, “Darkloc+: Thermal image-based indoor localization for dark environments with relative geometry constraints,”IEEE Transac- tions on Geoscience and Remote Sensing, 2024

    2024

  8. [17]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  9. [18]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  10. [19]

    Sthereo: Stereo thermal dataset for research in odometry and mapping,

    S. Yun, M. Jung, J. Kim, S. Jung, Y . Cho, M.-H. Jeon, G. Kim, and A. Kim, “Sthereo: Stereo thermal dataset for research in odometry and mapping,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 3857–3864

    2022

  11. [20]

    Eil-slam: Depth-enhanced edge-based infrared-lidar slam,

    W. Chen, Y . Wang, H. Chen, and Y . Liu, “Eil-slam: Depth-enhanced edge-based infrared-lidar slam,”Journal of Field Robotics, vol. 39, no. 2, pp. 117–130, 2022

    2022

  12. [21]

    Sparse depth enhanced direct thermal-infrared slam beyond the visible spectrum,

    Y .-S. Shin and A. Kim, “Sparse depth enhanced direct thermal-infrared slam beyond the visible spectrum,”IEEE Robotics and Automation Letters, vol. 4, no. 3, pp. 2918–2925, 2019

    2019

  13. [22]

    Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions,

    C. Doer and G. F. Trommer, “Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 331–338

    2021

  14. [23]

    Rgb-t slam: A flexible slam framework by combining appearance and thermal information,

    L. Chen, L. Sun, T. Yang, L. Fan, K. Huang, and Z. Xuanyuan, “Rgb-t slam: A flexible slam framework by combining appearance and thermal information,” in2017 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2017, pp. 5682–5687

    2017

  15. [24]

    Keyframe-based di- rect thermal–inertial odometry,

    S. Khattak, C. Papachristos, and K. Alexis, “Keyframe-based di- rect thermal–inertial odometry,” in2019 International conference on robotics and automation (ICRA). IEEE, 2019, pp. 3563–3569

    2019

  16. [25]

    Thermal stereo odometry for uavs,

    T. Mouats, N. Aouf, L. Chermak, and M. A. Richardson, “Thermal stereo odometry for uavs,”IEEE Sensors Journal, vol. 15, no. 11, pp. 6335–6347, 2015

    2015

  17. [26]

    Practical infrared visual odometry,

    P. V . K. Borges and S. Vidas, “Practical infrared visual odometry,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 8, pp. 2205–2213, 2016

    2016

  18. [27]

    Thermal-depth odometry in challenging illumination conditions,

    X. Chen, W. Dai, J. Jiang, B. He, and Y . Zhang, “Thermal-depth odometry in challenging illumination conditions,”IEEE Robotics and Automation Letters, vol. 8, no. 7, pp. 3988–3995, 2023

    2023

  19. [28]

    Tp-tio: A robust thermal-inertial odometry with deep thermalpoint,

    S. Zhao, P. Wang, H. Zhang, Z. Fang, and S. Scherer, “Tp-tio: A robust thermal-inertial odometry with deep thermalpoint,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 4505–4512

    2020

  20. [29]

    Deeptio: A deep thermal-inertial odometry with visual hallucination,

    M. R. U. Saputra, P. P. De Gusmao, C. X. Lu, Y . Almalioglu, S. Rosa, C. Chen, J. Wahlstr ¨om, W. Wang, A. Markham, and N. Trigoni, “Deeptio: A deep thermal-inertial odometry with visual hallucination,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1672–1679, 2020

    2020

  21. [30]

    Graph-based thermal–inertial slam with probabilistic neural networks,

    M. R. U. Saputra, C. X. Lu, P. P. B. de Gusmao, B. Wang, A. Markham, and N. Trigoni, “Graph-based thermal–inertial slam with probabilistic neural networks,”IEEE Transactions on Robotics, vol. 38, no. 3, pp. 1875–1893, 2021

    2021

  22. [31]

    Keyframe-based thermal– inertial odometry,

    S. Khattak, C. Papachristos, and K. Alexis, “Keyframe-based thermal– inertial odometry,”Journal of Field Robotics, vol. 37, no. 4, pp. 552– 579, 2020

    2020

  23. [32]

    Thermal-inertial slam for the environments with challenging illumination,

    J. Jiang, X. Chen, W. Dai, Z. Gao, and Y . Zhang, “Thermal-inertial slam for the environments with challenging illumination,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 8767–8774, 2022

    2022

  24. [33]

    Leveraging equivariant features for absolute pose regression,

    M. A. Musallam, V . Gaudilliere, M. O. Del Castillo, K. Al Ismaeil, and D. Aouada, “Leveraging equivariant features for absolute pose regression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 6876–6886

    2022

  25. [34]

    Atloc: Attention guided camera localization,

    B. Wang, C. Chen, C. X. Lu, P. Zhao, N. Trigoni, and A. Markham, “Atloc: Attention guided camera localization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 06, 2020, pp. 10 393–10 401

    2020

  26. [35]

    Effloc: Lightweight vision transformer for efficient 6-dof camera relocalization,

    Z. Xiao, C. Chen, S. Yang, and W. Wei, “Effloc: Lightweight vision transformer for efficient 6-dof camera relocalization,”arXiv preprint arXiv:2402.13537, 2024

    2024 arXiv

  27. [36]

    Ct-loc: Cross-domain visual localization with a channel-wise transformer,

    D. Kim and J. Kim, “Ct-loc: Cross-domain visual localization with a channel-wise transformer,”Neural Networks, vol. 158, pp. 369–383, 2023

    2023

  28. [37]

    Transbonet: Learning camera localization with transformer bottleneck and attention,

    X. Song, H. Li, L. Liang, W. Shi, G. Xie, X. Lu, and X. Hei, “Transbonet: Learning camera localization with transformer bottleneck and attention,”Pattern Recognition, vol. 146, p. 109975, 2024

    2024

  29. [38]

    Deep 6-dof camera relocalization in variable and dynamic scenes by multitask learning,

    J. Wang and Y . Qi, “Deep 6-dof camera relocalization in variable and dynamic scenes by multitask learning,”Machine Vision and Applications, vol. 34, no. 3, p. 37, 2023

    2023

  30. [39]

    Learn then match: A fast coarse-to-fine depth image-based indoor localization framework for dark environments via deep learning and keypoint-based geometry alignment,

    Q. Li, R. Cao, J. Zhu, H. Fu, B. Zhou, X. Fang, S. Jia, S. Zhang, K. Liu, and Q. Li, “Learn then match: A fast coarse-to-fine depth image-based indoor localization framework for dark environments via deep learning and keypoint-based geometry alignment,”ISPRS Journal of Photogr...

    2023

  31. [40]

    Featloc: Absolute pose regres- sor for indoor 2d sparse features with simplistic view synthesizing,

    T. B. Bach, T. T. Dinh, and J.-H. Lee, “Featloc: Absolute pose regres- sor for indoor 2d sparse features with simplistic view synthesizing,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 189, pp. 50–62, 2022

    2022

  32. [41]

    Radarloc: Learning to relocalize in fmcw radar,

    W. Wang, P. P. de Gusm ˜ao, B. Yang, A. Markham, and N. Trigoni, “Radarloc: Learning to relocalize in fmcw radar,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 5809–5815

    2021

  33. [42]

    Sgloc: Scene ge- ometry encoding for outdoor lidar localization,

    W. Li, S. Yu, C. Wang, G. Hu, S. Shen, and C. Wen, “Sgloc: Scene ge- ometry encoding for outdoor lidar localization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9286–9295

    2023

  34. [43]

    Nidaloc: Neurobiologically inspired deep lidar localization,

    S. Yu, X. Sun, W. Li, C. Wen, Y . Yang, B. Si, G. Hu, and C. Wang, “Nidaloc: Neurobiologically inspired deep lidar localization,”IEEE Transactions on Intelligent Transportation Systems, 2023

    2023

  35. [44]

    Stcloc: Deep lidar localization with spatio-temporal constraints,

    S. Yu, C. Wang, Y . Lin, C. Wen, M. Cheng, and G. Hu, “Stcloc: Deep lidar localization with spatio-temporal constraints,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 1, pp. 489–500, 2022

    2022

  36. [45]

    Vint: A foundation model for visual navigation,

    D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine, “Vint: A foundation model for visual navigation,”arXiv preprint arXiv:2306.14846, 2023

    2023 arXiv

  37. [46]

    Rt-1: Robotics transformer for real-world control at scale,

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsuet al., “Rt-1: Robotics transformer for real-world control at scale,”arXiv preprint arXiv:2212.06817, 2022

    2022 arXiv

  38. [47]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”arXiv preprint arXiv:1706.03762, 2017

    2017 arXiv

  39. [48]

    Geometry- aware learning of maps for camera localization,

    S. Brahmbhatt, J. Gu, K. Kim, J. Hays, and J. Kautz, “Geometry- aware learning of maps for camera localization,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2616–2625

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.