Pith. sign in

REVIEW 5 major objections 5 minor 36 references

Robust Visual Localization via Semantic-Guided Multi-Scale Transformer

T0 review · 5 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 A hierarchical Transformer with cross-scale attention, guided during training by semantic NeRF supervision, achieves state-of-the-art absolute pose regression accuracy on TartanAir, with 0.22 m / 0.82 degrees median error for the semantic…

desk verdict Interesting integration, but the evaluation protocol and the semantic-supervision pathway are under-specified enough that the main claims are not auditable. read the letter →

arxiv 2506.08526 v1 pith:JUVQ2K4Z submitted 2025-06-10 cs.CV

Zhongtao Tian , Wenhao Huang , Zhidong Chen , Xiao Wei Sun This is my paper
classification cs.CV
keywords absoluteposeregressionvisuallocalizationmulti-scaleTransformercross-scaleattentionsemanticneuralradiancefieldssupervisionTartanAirrobustnesstoenvironmentalchange
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 an absolute-pose-regression method that combines a hierarchical Transformer with cross-scale attention and semantic supervision from a neural scene representation during training. It aims to show that this combination makes camera pose estimation more robust than existing APR methods under lighting changes, dynamic objects, and occlusions. On TartanAir, the pose-only model reports a median translation error of 0.23 m and rotation error of 0.88 degrees, improving on the DFNet baseline's 0.28 m and 2.07 degrees; adding semantic supervision gives 0.22 m and 0.82 degrees. The claim is that fusing features across scales and guiding them with view-invariant semantic cues is a workable route to robust localization at inference without extra cost.

What carries the argument

The load-bearing mechanism is the hierarchical Transformer with cross-scale attention: multi-scale feature maps are progressively upsampled and fused, with attention maps from coarser scales blended into finer levels through learnable weights, plus learnable sinusoidal position encodings to keep spatial precision. The second mechanism is SemanticMap, a semantic neural radiance field that outputs semantic logits as a function of 3D position only, renders them by volumetric integration, and supervises the pose net with cross-entropy and spectral angle mapping losses during training only. Together these are meant to let the network emphasize global structure when local textures are unreliable and to encode view-invariant scene structure rather than appearance.

What would settle it

Re-running PoseFormer and PoseFormer_sem on the standard TartanAir evaluation split, training DFNet, Direct-PN, and MS-T under the same protocol, and checking whether the median errors reproduce 0.23 m / 0.88 degrees and 0.22 m / 0.82 degrees against 0.28 m / 2.07 degrees; if the margin disappears or the reported numbers cannot be reproduced, the central claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a multi-scale Transformer pose regressor, trained jointly with a semantic neural rendering module, outperforms existing absolute pose regression methods on the TartanAir benchmark. PoseFormer uses a MobileNetV2 backbone to produce features at three resolutions, processes them with upsampling Transformer blocks and cross-scale attention, and regresses translation and rotation through separate heads. A SemanticMap module based on Semantic-NeRF renders view-invariant semantic logits along camera rays; during a third training stage, the frozen backbone is refined using semantic cross-entropy and spectral angle mapping losses instead of pose loss. The reported result is 0.23 m / 0.88 degrees for PoseFormer and 0.22 m / 0.82 degrees for PoseFormer_sem, versus 0.28 m / 2.07 degrees for DFNet, which the paper reads as evidence that semantic guidance helps the network keep static structure while suppressing environmental interference.

Load-bearing premise

The paper rests on an unstated evaluation protocol: it says only that ground-truth image-pose pairs are used for training, not which TartanAir frames or trajectories are held out for testing or how DFNet, Direct-PN, and MS-T were trained, so the reported improvements are meaningful only if the test set is disjoint from training and the baselines were trained fairly.

Editorial extensions

If this is right

  • If the reported errors hold, PoseFormer and PoseFormer_sem are the most accurate single-frame APR methods on TartanAir, with the semantic variant reducing median rotation error to below one degree.
  • The semantic supervision component can be removed at inference, so the robustness gain comes at no added runtime cost for pose estimation.
  • The design implies that APR accuracy under illumination changes, dynamic objects, and occlusions can be improved by architectural multi-scale fusion without explicit modeling of each disturbance type.
  • The three-stage training recipe, pose pretraining, semantic NeRF fitting, and semantic-guided refinement, becomes a reusable template for adding view-invariant supervision to other pose regressors.

Reading between the lines

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

  • A fair test of the headline numbers requires the exact TartanAir train/test split and baseline training budgets; the paper does not specify them, so the margins are conditional on that protocol.
  • Because semantic supervision needs a per-scene semantic NeRF trained with semantic labels, applying the method to open-world or real-world localization would require labels or a proxy view-invariant signal, a step the paper does not address.
  • The same cross-scale attention treatment could be plugged into other APR architectures; the ablation's progressive gains from unfreezing hierarchical layers suggest the benefit comes from scale fusion rather than the specific backbone.
  • A natural next experiment is transferring the trained models to real day-night benchmarks, since TartanAir is synthetic and the robustness claim is about environmental change in general.
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 / 5 minor

Summary. The paper proposes a visual localization framework called PoseFormer, a hierarchical multi-scale Transformer with cross-scale attention for absolute pose regression, and a variant PoseFormer_sem that adds semantic supervision through a semantic NeRF module. The authors claim state-of-the-art results on the TartanAir benchmark, reporting average median translation/rotation errors of 0.23 m/0.88° for PoseFormer and 0.22 m/0.82° for PoseFormer_sem, versus 0.28 m/2.07° for the previous best method DFNet. The paper also presents ablations on backbone choice and multi-scale feature fusion. The central claim is that multi-scale feature learning combined with semantic consistency supervision improves robustness to lighting changes, dynamic objects, and occlusions.

Significance. If the reported results are reproducible and the comparison is fair, the method would be a meaningful advance in absolute pose regression under challenging environmental conditions, and the combination of semantic NeRF supervision with APR is a novel idea worth exploring. The paper addresses a practically important problem and the experimental scenarios are diverse. However, as written, the evidence for the central claim is not verifiable from the manuscript: the evaluation protocol is incompletely specified, baseline training details are absent, and the semantic supervision pathway is not clearly described. These gaps prevent an assessment of whether the claimed improvements reflect a genuine methodological advance or an artifact of the experimental setup.

major comments (5)
  1. [§4.1, Table 1] The train/test split for the TartanAir evaluation is never specified. The sentence 'For all studies, we use the ground-truth image-pose pairs for training' does not indicate which of the fourteen scenes, trajectories, or frames are used for training versus testing, nor whether evaluation frames overlap the training set. Without this information, every number in Table 1 is uninterpretable because the reader cannot determine whether the comparison is held-out or memorized. This is load-bearing for the paper's central SOTA claim and must be corrected by explicitly stating the split (e.g., per-scene train/test trajectories, frame intervals, or a reference to a standard TartanAir APR protocol).
  2. [§4.1, Table 1] No training details are provided for the baselines (PoseNet, Direct-PN, MS-T, DFNet, Direct-PN+U). The reader cannot assess whether these methods were trained on the same split, with the same input resolution (320×240), the same optimization budget, or the same loss weighting. If baselines were under-trained or given less capacity, the reported improvements would be misleading. The paper should specify the baseline training protocol or cite the exact source and configuration used to produce each baseline number.
  3. [§3.3.2, Eq. (12)] The spectral angle mapping loss L_sam in Eq. (12) is never defined. The text merely states that it 'evaluates structural similarity between predicted and ground-truth semantic maps,' but no formula, distance metric, or implementation details are given. Since L_sam is a component of the semantic supervision loss that is essential to the method, its absence makes the loss function non-reproducible. The authors should provide the explicit definition of L_sam (e.g., the cosine distance between flattened semantic maps or spectral vectors).
  4. [§3.2, §3.3, and Stage 3 in §4.1] The mechanism by which semantic supervision influences the PoseFormer network is not clearly described. The SemanticMap module appears to be a separate NeRF-like network that renders semantic logits using ground-truth poses (Eq. 6-7), but the text does not state whether it shares parameters with PoseFormer, whether its output is compared against ground-truth semantic labels, or how the resulting loss in Eq. (12) backpropagates into PoseFormer. Stage 3 says the PoseFormer backbone is frozen and only BatchNorm layers are unfrozen, and that 'we disable direct pose supervision and instead rely solely on the semantic consistency loss'—this is confusing because the semantic loss appears to operate on rendered semantic maps, not on PoseFormer's pose output. Without a clear description of the gradient path from L_semantic to PoseFormer's feature extractor, the claim that semantic guidance improves pose estimation is not supported by the described architecture.
  5. [§4.4, Tables 2 and 3] The ablation studies do not specify whether they are performed on PoseFormer or PoseFormer_sem, nor do they state whether they use the same train/test split and evaluation protocol as Table 1. The reported values in Tables 2 and 3 (e.g., MobileNetV2 0.25/0.90) differ from the PoseFormer row in Table 1 (0.23/0.88), indicating that the ablated model is not identical to the final model, but the relationship is not explained. This makes the ablations difficult to interpret and should be clarified.
minor comments (5)
  1. [Throughout] There are several typographical and grammatical errors, e.g., 'we combines' in the introduction, 'an hierarchical' in Section 3.1, and '1×1convolutions' missing a space in Section 3.1.1. The paper should be proofread.
  2. [Table 1, scene names] Scene names in Table 1 are not consistently formatted (e.g., 'Japanesealley' should be 'Japanese Alley', 'Soulcity' as 'Soul City', 'Gascola' as 'Gascola' or 'Gas Cola'). This is cosmetic but should be fixed for readability.
  3. [§4.2] The claim that 'PoseFormer achieves 18% improvement in translation error (0.23 m vs. 0.28 m)' is arithmetically inconsistent: (0.28-0.23)/0.28 ≈ 17.9%, so '18%' is acceptable, but the relative improvement in rotation is (2.07-0.88)/2.07 ≈ 57.5%, so '57%' is fine. The text should be internally consistent in the number of significant digits.
  4. [§4.4] The ablation descriptions for backbone choice rely on qualitative claims about 'overfitting tendencies' and 'spatially discriminative feature encoding' without quantitative support or references. These claims should either be substantiated or softened.
  5. [§4.2, Figure 4] The caption says the color bar represents the distribution of rotation errors, but the color bar in the figure is not described in the text; it would be helpful to state the units or the mapping from color to error magnitude.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claims are external benchmark comparisons, not derivations from fitted inputs or self-citations.

full rationale

This is an empirical benchmark paper rather than a derivation chain, so the circularity patterns do not apply. The pose objective in Eq. (11) is direct regression against ground-truth poses from the external TartanAir dataset, and the semantic objective in Eq. (12) is an auxiliary supervision loss using external semantic labels and NeRF rendering, not a quantity derived from the method's own outputs or fitted to the test set. The reported headline numbers are comparisons against published APR baselines on TartanAir, and the paper does not fit any parameter to the reported test errors, invoke a uniqueness theorem, or rely on self-citations. The missing description of the train/test split and baseline training details is a reproducibility and evaluation-protocol concern, not a circularity concern, because nothing in the paper's equations reduces its claimed result to its own inputs. Under the hard rules requiring a quoted specific reduction, no circular step can be identified.

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

The paper introduces no new physical or conceptual entities such as particles, forces, or dimensions. PoseFormer and SemanticMap are network components, not postulated entities with independent empirical handles. The main load-bearing assumptions are about the evaluation protocol and the view-invariance of semantic labels.

free parameters (3)
  • Semantic loss weights = [0.7, 0.3] for cross-entropy and spectral angle mapping
    Chosen by hand in Stage 3, Eq. 12 and Section 4.1; no ablation is shown for these values.
  • lambda_p and lambda_s in Eq. 8 = not reported
    Global pose and semantic loss weights are introduced but their values are not given.
  • Semantic-NeRF semantic loss weight = 4e-2
    Set by hand for Stage 2 training of the semantic scene representation.
assumptions (4)
  • domain assumption TartanAir has a meaningful train/test protocol and the paper uses it correctly.
    Section 4.1 only says ground-truth image-pose pairs are used for training; no held-out split is described, so the benchmark comparison depends on this unstated assumption.
  • domain assumption Semantic labels are view-invariant and resistant to lighting and weather changes.
    This property is asserted in Sections 1 and 3.2 to justify semantic supervision, but no experiment isolates or validates it.
  • standard math Volumetric semantic rendering, Eq. 7, and its gradients are valid for training.
    Borrowed from NeRF and Semantic-NeRF without formal proof, which is reasonable given the prior literature.
  • ad hoc to paper The baselines DFNet, Direct-PN, and MS-T are reimplemented or configured fairly and comparably.
    No training details, hyperparameters, or code are given for these baselines; Table 1 depends on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Visual Localization via Semantic-Guided Multi-Scale Transformer." pith.science (2026). https://pith.science/paper/JUVQ2K4Z

@misc{pith2026250608526,
  author       = {Pith},
  title        = {Pith review of: Robust Visual Localization via Semantic-Guided Multi-Scale Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUVQ2K4Z}},
  note         = {Machine review of arXiv:2506.08526}
}
read the original abstract

Visual localization remains challenging in dynamic environments where fluctuating lighting, adverse weather, and moving objects disrupt appearance cues. Despite advances in feature representation, current absolute pose regression methods struggle to maintain consistency under varying conditions. To address this challenge, we propose a framework that synergistically combines multi-scale feature learning with semantic scene understanding. Our approach employs a hierarchical Transformer with cross-scale attention to fuse geometric details and contextual cues, preserving spatial precision while adapting to environmental changes. We improve the performance of this architecture with semantic supervision via neural scene representation during training, guiding the network to learn view-invariant features that encode persistent structural information while suppressing complex environmental interference. Experiments on TartanAir demonstrate that our approach outperforms existing pose regression methods in challenging scenarios with dynamic objects, illumination changes, and occlusions. Our findings show that integrating multi-scale processing with semantic guidance offers a promising strategy for robust visual localization in real-world dynamic environments.

Figures

Figures reproduced from arXiv: 2506.08526 by the authors.

Figure 1
Figure 1. Overview of our framework. The network consists of a feature extraction backbone followed by PoseFormer for pose estimation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of our PoseFormer. Multi-scale features [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the Cross-scale Attention mechanism. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Overlaid attention maps for 12 different scenes, encom [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 29 canonical work pages

  1. [1]

    Nerfels: renderable neural codes for improved camera pose estima- tion

    Gil Avraham, Julian Straub, Tianwei Shen, Tsun-Yi Yang, Hugo Germain, Chris Sweeney, Vasileios Balntas, David Novotny, Daniel DeTone, and Richard Newcombe. Nerfels: renderable neural codes for improved camera pose estima- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 5061–5070,

    work page

  2. [2]

    Reloc- net: Continuous metric learning relocalisation using neural nets

    Vassileios Balntas, Shuda Li, and Victor Prisacariu. Reloc- net: Continuous metric learning relocalisation using neural nets. InProceedings of the European conference on com- puter vision (ECCV), pages 751–767, 2018. 2

    2018

  3. [3]

    Depth camera based indoor mobile robot localization and navigation

    Joydeep Biswas and Manuela Veloso. Depth camera based indoor mobile robot localization and navigation. In2012 IEEE International Conference on Robotics and Automation, pages 1697–1702. IEEE, 2012. 1

    work page 2012

  4. [4]

    Geometry-aware learning of maps for camera localization

    Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2616–2625, 2018. 1

    work page 2018

  5. [5]

    Geometry-aware learning of maps for cam- era localization

    Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for cam- era localization. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2616–2625,

  6. [6]

    Let’s take this online: Adapting scene coordinate regression network predictions for online rgb-d camera relocalisation

    Tommaso Cavallari, Luca Bertinetto, Jishnu Mukhoti, Philip Torr, and Stuart Golodetz. Let’s take this online: Adapting scene coordinate regression network predictions for online rgb-d camera relocalisation. In2019 International Confer- ence on 3D Vision (3DV), pages 564–573, 2019. 1

    work page 2019

  7. [7]

    Crossvit: Cross-attention multi-scale vision transformer for image classification

    Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. InProceedings of the IEEE/CVF in- ternational conference on computer vision, pages 357–366,

  8. [8]

    Direct- posenet: Absolute pose regression with photometric consis- tency

    Shuai Chen, Zirui Wang, and Victor Prisacariu. Direct- posenet: Absolute pose regression with photometric consis- tency. In2021 International Conference on 3D Vision (3DV), pages 1175–1185. IEEE, 2021. 2

    work page 2021

Show all 36 references
  1. [9]

    Dfnet: Enhance absolute pose regression with direct feature matching

    Shuai Chen, Xinghui Li, Zirui Wang, and Victor A Prisacariu. Dfnet: Enhance absolute pose regression with direct feature matching. InEuropean Conference on Com- puter Vision, pages 1–17. Springer, 2022. 1, 3

    2022

  2. [10]

    Camnet: Coarse-to-fine retrieval for camera re- localization

    Mingyu Ding, Zhe Wang, Jiankai Sun, Jianping Shi, and Ping Luo. Camnet: Coarse-to-fine retrieval for camera re- localization. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2871–2880, 2019. 2

    2019

  3. [11]

    Modelling uncertainty in deep learning for camera relocalization

    Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In2016 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 4762–4769, 2016. 2

    2016

  4. [12]

    Geometric loss func- tions for camera pose regression with deep learning

    Alex Kendall and Roberto Cipolla. Geometric loss func- tions for camera pose regression with deep learning. In2017 IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 6555–6564, 2017. 1, 2, 5

    2017

  5. [13]

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

    Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof cam- era relocalization. In2015 IEEE International Conference on Computer Vision (ICCV), pages 2938–2946, 2015. 1, 2

    2015

  6. [14]

    Vela, and Stan Birchfield

    Yunzhi Lin, Thomas M ¨uller, Jonathan Tremblay, Bowen Wen, Stephen Tyree, Alex Evans, Patricio A. Vela, and Stan Birchfield. Parallel inversion of neural radiance fields for robust pose estimation. In2023 IEEE International Confer- ence on Robotics and Automation (ICRA), pages...

  7. [15]

    Image-based localization using hourglass networks

    Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and Esa Rahtu. Image-based localization using hourglass networks. In2017 IEEE International Conference on Computer Vision Workshops (ICCVW), pages 870–877, 2017. 2

    2017

  8. [16]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthe- sis.Commun. ACM, 65(1):99–106, 2021. 1

    2021

  9. [17]

    Lens: Localization enhanced by nerf synthesis

    Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. InConference on Robot Learn- ing, pages 1347–1356. PMLR, 2022. 1, 3

    2022

  10. [18]

    Crossfire: Camera relocalization on self-supervised features from an implicit representation

    Arthur Moreau, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Crossfire: Camera relocalization on self-supervised features from an implicit representation. InProceedings of the IEEE/CVF International Conference on Computer Vi...

    2023

  11. [19]

    Leveraging equivariant features for absolute pose regression

    Mohamed Adel Musallam, Vincent Gaudilliere, Miguel Or- tiz Del Castillo, Kassem Al Ismaeil, and Djamila Aouada. Leveraging equivariant features for absolute pose regression. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6876–6886, 2022. 2

    2022

  12. [20]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 4510–4520, 2018. 3, 6

    2018

  13. [21]

    Back to the feature: Learning robust camera localization from pixels to pose

    Paul-Edouard Sarlin, Ajaykumar Unagar, M ˚ans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the feature: Learning robust camera localization from pixels to pose. In2021 IEEE/CVF ...

    2021

  14. [22]

    Benchmarking 6dof outdoor visual localiza- tion in changing conditions

    Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6dof outdoor visual localiza- tion in changing conditions. In2018 IEEE/CVF Con...

    2018

  15. [23]

    Sch ¨onberger and Jan-Michael Frahm

    Johannes L. Sch ¨onberger and Jan-Michael Frahm. Structure- from-motion revisited. In2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 4104– 4113, 2016. 1

    2016

  16. [24]

    Do we really need scene- specific pose encoders? In2020 25th International Con- ference on Pattern Recognition (ICPR), pages 3186–3192

    Yoli Shavit and Ron Ferens. Do we really need scene- specific pose encoders? In2020 25th International Con- ference on Pattern Recognition (ICPR), pages 3186–3192. IEEE, 2021. 2

    2021

  17. [25]

    Camera pose auto-encoders for improving pose regression

    Yoli Shavit and Yosi Keller. Camera pose auto-encoders for improving pose regression. InEuropean Conference on Computer Vision, pages 140–157. Springer, 2022. 2

    2022

  18. [26]

    Sparsepose: Sparse- view camera pose regression and refinement

    Samarth Sinha, Jason Y Zhang, Andrea Tagliasacchi, Igor Gilitschenski, and David B Lindell. Sparsepose: Sparse- view camera pose regression and refinement. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21349–21359, 2023. 1

    2023

  19. [27]

    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 Ak- ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogn...

    2018

  20. [28]

    Global localization from monocular slam on a mobile phone.IEEE Transactions on Visualization and Computer Graphics, 20(4):531–539, 2014

    Jonathan Ventura, Clemens Arth, Gerhard Reitmayr, and Di- eter Schmalstieg. Global localization from monocular slam on a mobile phone.IEEE Transactions on Visualization and Computer Graphics, 20(4):531–539, 2014. 1

    2014

  21. [29]

    Walch, C

    F. Walch, C. Hazirbas, L. Leal-Taix ´e, T. Sattler, S. Hilsen- beck, and D. Cremers. Image-based localization using lstms for structured feature correlation. In2017 IEEE Interna- tional Conference on Computer Vision (ICCV), pages 627– 637, 2017. 2

    2017

  22. [30]

    Atloc: Atten- tion guided camera localization

    Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Atten- tion guided camera localization. InProceedings of the AAAI Conference on Artificial Intelligence, pages 10393–10401,

  23. [31]

    Tartanair: A dataset to push the limits of visual slam

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Se- bastian Scherer. Tartanair: A dataset to push the limits of visual slam. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909–491...

    2020

  24. [32]

    Crossformer++: A versatile vision transformer hinging on cross-scale attention

    Wenxiao Wang, Wei Chen, Qibo Qiu, Long Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wei Liu. Crossformer++: A versatile vision transformer hinging on cross-scale attention. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 46(5):3123–3136, 2023. 4

    2023

  25. [33]

    Inverted pyramid multi-task trans- former for dense scene understanding

    Hanrong Ye and Dan Xu. Inverted pyramid multi-task trans- former for dense scene understanding. InEuropean Confer- ence on Computer Vision, pages 514–530. Springer, 2022. 4

    2022

  26. [34]

    Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin

    Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1323–1330, 2021. 1

    2021

  27. [35]

    Pnerfloc: Visual localization with point- based neural radiance fields

    Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, and Zhaopeng Cui. Pnerfloc: Visual localization with point- based neural radiance fields. InProceedings of the AAAI Conference on Artificial Intelligence, pages 7450–7459,

  28. [36]

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and An- drew J. Davison. In-place scene labelling and understanding with implicit scene representation. In2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 15818–15827, 2021. 1, 6

    2021

Pith tools

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