LRSLAM: Low-rank Representation of Signed Distance Fields in Dense Visual SLAM System

Giljoo Nam; Hongbeen Park; Jinkyu Kim; Minjeong Park

REVIEW 4 major objections 5 minor 32 references

LRSLAM: Low-rank Representation of Signed Distance Fields in Dense Visual SLAM System

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 LRSLAM claims a hybrid CP and Six-axis tensor representation of signed distance fields gives dense RGB-D SLAM a linear-memory scene model, cutting parameters by 87.3–90.1% versus ESLAM while matching or improving localization and…

desk verdict A credible efficiency gain for dense SLAM via a genuinely lower-memory representation, but the accuracy claims are overstated and the rank selection is tuned on the same benchmark used for headlines. read the letter →

arxiv 2506.10567 v1 pith:L3W2Q3EU submitted 2025-06-12 cs.CV

Hongbeen Park , Minjeong Park , Giljoo Nam , Jinkyu Kim This is my paper

classification cs.CV MSC 68T4515A69

keywords densevisualSLAMRGB-Dsigneddistancefieldlow-ranktensordecompositionCPSix-axisscenerepresentationmemoryefficiency

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

This paper claims that dense RGB-D visual SLAM can be made dramatically lighter by representing the scene's signed distance field and appearance as sums of rank-one tensors instead of full feature grids or planes. The proposed system, LRSLAM, factorizes geometry with CP decomposition and appearance with a new Six-axis decomposition, giving a linear memory complexity $O(n)$ in the scene's side length. On three standard indoor benchmarks the authors report 87.3%–90.1% fewer scene parameters than ESLAM and 4.3%–73.2% faster frame processing, with matched or better localization and reconstruction accuracy. If this holds, the main obstacle to running dense neural SLAM on memory-limited robots and AR devices is relaxed.

What carries the argument

The central object is the hybrid scene representation built from two tensor factorizations. CP decomposition writes a 4D feature tensor for geometry as a sum of $k_g$ outer products of three axis-aligned feature tensors, reading out a feature vector at any point $p$ by summing the three interpolated rank-one terms. The new Six-axis decomposition factorizes the three tri-plane feature planes of appearance into six axis-aligned rank-one feature tensors, two per plane, so that querying color at $p$ sums three rank-one outer-product pairs. Both run at $O(n)$ storage, versus $O(n^2)$ for ESLAM's planes, and the two-layer MLP decoders then map the feature vectors to TSDF values and colors.

What would settle it

Run LRSLAM on a held-out set of indoor scenes that include thin structures, repeated textures, or high-frequency geometry (e.g., a cluttered office or a room with many chairs and monitors) and compare ATE RMSE and reconstruction accuracy against ESLAM. If the rank-2 CP geometry field produces visible reconstruction gaps or camera drift on those scenes, the universal-efficiency claim fails. A simpler check: re-run the ScanNet experiments with $k_g=1$ or $k_g=3$ and show that tracking accuracy changes materially, indicating sensitivity to the fixed rank.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hybrid low-rank tensor representation of a scene can replace ESLAM's twelve multi-resolution feature planes without sacrificing SLAM quality. Geometry is encoded by CP decomposition, a sum of $k_g$ outer products of three axis-aligned feature vectors. Appearance is encoded by the proposed Six-axis decomposition, where each of the three tri-planes is itself factored into $k_a$ outer products of two axis-aligned vectors, yielding six feature tensors total. With $k_g=2$ and $k_a=16$, the authors report 0.86M–1.38M total parameters depending on dataset, which is 87.3%–90.1% fewer than ESLAM, and frame processing times 4.3%–73.2% faster, while ATE RMSE on ScanNet improves from 7.4 cm to 7.0 cm and on TUM RGB-D improves on two of three scenes. The paper argues the hybrid works because geometry is lower-frequency and CP converges quickly, which then helps appearance optimization through the geometric weighting in color volume rendering.

Load-bearing premise

The system assumes that fixed tensor ranks, $k_g=2$ for geometry and $k_a=16$ for appearance, are expressive enough for every indoor scene; these ranks were selected by ablations on the same benchmark used for evaluation, so any scene needing richer geometry could be underfit and cause tracking drift.

Editorial extensions

If this is right

Dense SLAM runs with an order-of-magnitude smaller scene model, which directly translates to lower memory footprint on embedded and mobile platforms.
The linear memory growth in scene side length means larger indoor environments can be mapped at full resolution without the quadratic blow-up that limits ESLAM.
Faster convergence of the CP geometry field should let the system reach accurate camera tracking earlier in a sequence, reducing drift on long trajectories.
The robustness to synthetic depth noise reported on Replica suggests low-rank factorization acts as an implicit denoiser for sensor depth, which matters for real robot sensors.
Since ESLAM's loss functions and pipeline are reused unchanged, LRSLAM is a drop-in replacement for the scene representation inside existing ESLAM-based systems.

Reading between the lines

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

The hybrid design suggests a more general principle for neural scene representations: pair a low-frequency, fast-converging factorization for the geometric field with a higher-rank factorization for appearance, rather than seeking one universal decomposition.
The authors' choice of $k_g$ and $k_a$ was tuned on ScanNet and then evaluated on the same benchmark; a truly held-out experiment would be needed to confirm the fixed ranks generalize to unseen scene types.
The robustness to noise could be tested as an explicit denoising claim by injecting structured sensor noise (e.g., depth shadows or missing values) beyond zero-mean Gaussian, where low-rank filtering might fail.
Six-axis decomposition may be of independent interest for other tri-plane-based neural fields (NeRF, generative models), not just SLAM, since it preserves expressiveness while cutting storage from $O(n^2)$ to $O(n)$.

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

Read LRSLAM. The contribution is real: they factorize the tri-plane representation into a Six-axis low-rank form, pair it with CP for geometry, and get O(n) memory scaling for scene features in an ESLAM-style SLAM system. The parameter counts in Table 1 are dramatic (87–90% fewer than ESLAM) and the frame processing times are consistently faster, and on ScanNet and TUM the localization numbers are roughly matched or slightly better. That is a solid efficiency result, and the hybrid design — CP for smooth geometry, SA for appearance — is a sensible architectural choice that they back with an ablation.

The central caveat is the accuracy claim. The abstract says “superior performance” and “better convergence rates,” but the noiseless Replica results in Table 4 are clearly worse than ESLAM (Depth L1 1.58 vs 1.18, ATE RMSE 0.79 vs 0.63). That is not a small gap, and it aligns with the underfitting concern: a CP field with rank 2 is a very smooth representation. The noisy-depth result is genuinely interesting — LRSLAM holds up under sensor noise where ESLAM degrades — but that does not rescue the clean-scene comparison. The paper would be more honest if it framed the method as a compact representation that trades a bit of accuracy in clean synthetic scenes for robustness and efficiency.

The other issue is the rank selection. k_g=2 and k_a=16 are chosen via ablation on ScanNet scenes, and then the same six ScanNet scenes are the headline result. There is no held-out scene demonstrating that these ranks generalize. If the architecture underfits a new scene, the “retaining accuracy” half of the central claim fails. This is a selection-on-the-test-set problem, and it is correctable by evaluating on scenes not used for model selection or by showing rank sensitivity.

I would also note that code is withheld. That matters less for the complexity argument, which is straightforward, than for the accuracy claims, which depend on training details.

The math and complexity argument are sound. The citation pattern looks fine, though some concurrent linear-memory SLAM work is not discussed. This is a paper that deserves a serious referee: the efficiency result is real, the architecture is new, and the flaws are fixable rather than fatal. I would send it to review but with a clear request to address the Replica discrepancy and the selection issue.

Referee Report

4 major / 5 minor

Summary. The manuscript presents LRSLAM, a dense RGB-D visual SLAM system that replaces ESLAM's tri-plane feature grids with a hybrid low-rank scene representation: CP decomposition for geometry and the proposed Six-axis decomposition for appearance, both with O(n) memory complexity. The system is evaluated on ScanNet, TUM RGB-D, and Replica, with comparisons against NICE-SLAM and ESLAM in terms of parameter counts, frame processing time, ATE, and reconstruction metrics. The central claim is that LRSLAM retains localization and reconstruction accuracy while using 87.3% to 90.1% fewer scene parameters and 4.3% to 73.2% faster processing than ESLAM.

Significance. If the efficiency and accuracy claims hold, LRSLAM is a meaningful step toward practical dense SLAM: linear memory growth in the scene side length is well motivated, and the complexity analysis in Sec. 3 is straightforward and sound. The paper's strengths include multi-run reporting on ScanNet and Replica with standard deviations, a clear ablation over representation combinations, and detailed parameter/complexity comparisons. The main risks are that the key hyperparameters are tuned on the same benchmark used for headline results, no held-out scene demonstrates generalization, and the noiseless Replica numbers are worse than ESLAM, so the 'retaining accuracy' part of the claim is not fully established. The paper would be strengthened by releasing code and adding held-out or sensitivity experiments.

major comments (4)

[§5.1, §5.4, Table 2] The fixed ranks kg=2 for CP geometry and ka=16 for Six-axis appearance are selected through the ablation in Sec. 5.4 on ScanNet scenes, and the same six ScanNet scenes are then reported as the main localization evidence in Table 2. Because there is no held-out model selection, the expressive sufficiency of kg=2 is untested on scenes outside the tuning subset. Since the SDF in Eq. 6 is used both for rendering depth and for computing tracking weights, an underfit geometry field would cause tracking drift, directly undermining the 'retaining reconstruction and localization accuracy' claim. Please add held-out scenes or a sensitivity analysis over kg and ka, with error bars, to show that the chosen configuration generalizes.
[Abstract, Table 4] The abstract and introduction claim 'superior' accuracy and 'better reconstruction/localization quality than existing state-of-the-art approaches', but the noiseless Replica results in Table 4 show LRSLAM is worse than ESLAM on Depth L1 (1.58 vs 1.18), Accuracy (1.00 vs 0.97), Completion (1.07 vs 1.05), ATE Mean (0.61 vs 0.52), and ATE RMSE (0.79 vs 0.63). The evidence supports a more modest claim of matched or slightly worse accuracy on clean synthetic depth, with improved robustness under added depth noise. The paper should revise these overclaims and explicitly discuss whether the Replica degradation is consistent with underfitting of the rank-2 CP geometry field.
[§4.2, Fig. 6] The rationale for the hybrid representation asserts that CP decomposition converges faster for geometry and that early geometry convergence helps appearance optimization through the geometric dependency in Eq. 6, but no convergence curves or controlled experiments isolating this mechanism are provided. Fig. 6 compares only final ATE, not convergence behavior, so the 'better convergence rates' claim in the abstract is unsupported by the evidence shown. Adding per-iteration tracking or depth error curves for the CP-CP, SA-SA, and hybrid variants would directly test this load-bearing justification.
[§5.3, Table 4] The noise-robustness conclusion is based on a single scene (room0) with one noise level and no repeated trials or standard deviations for the noisy rows. The statement that low-rank representations 'have the ability to remove or filter sensor noise' is therefore not supported by the reported evidence. Please report results over multiple Replica scenes and noise levels, with multiple seeds, before drawing this conclusion.

minor comments (5)

[Table 1] The dataset name 'TUMRBG-D' in the table header is a typo for 'TUM RGB-D'.
[§3, Eq. (5)] The notation fSDF = MLP(f(p)) in Sec. 3 is inconsistent with the later use of phi_g(p) in Eq. (5) and Eq. (8); please align the notation throughout.
[Table 3] Unlike the ScanNet and Replica tables, Table 3 reports only single ATE RMSE values without standard deviations or repeated runs, so the claimed 0.8% to 13.5% improvements may not be statistically meaningful.
[§5.3, Fig. 5] The qualitative reconstruction figures for TUM RGB-D and Replica would be more informative with quantitative depth or mesh error maps, since visual comparison alone cannot distinguish the methods.
[§1, Contribution list] The paper states code will be publicly available, but no code or supplementary material is provided in the manuscript; given that hyperparameters and implementation details are central to the efficiency claims, releasing code or providing a detailed implementation appendix would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: LRSLAM's efficiency and accuracy claims rest on explicit scene-representation equations and external benchmark measurements, not on self-referential fits.

full rationale

Walking the claimed derivation chain, no step reduces to its own inputs. The scene representation is defined explicitly in Eqs. 7-12 (CP for geometry, Six-axis for appearance) and is not defined in terms of the evaluation metrics (ATE, depth L1, reconstruction error). The efficiency gains in Table 1 follow from parameter counting for the O(n) representation, and the accuracy claims are measured against external baselines (ESLAM, NICE-SLAM) on public datasets. ESLAM is used as the system framework, but this is disclosed as 'All the experiments shown in this paper base on the ESLAM [14] framework,' and it is not a load-bearing self-citation or an imported uniqueness theorem. The hyperparameters k_g=2 and k_a=16 are selected through the ablation in Sec. 5.4 on ScanNet scenes and then evaluated on the same benchmark; this is a model-selection concern that could affect external validity, but it is not a fitted parameter being relabeled as a prediction, and no equation is circular. The possible overstatement of the novelty of the Six-axis decomposition is a prior-art and contribution-scope question, not a circularity of the derivation. No self-citation chain, ansatz-smuggling, or renaming of a known result as a derived prediction appears. The central claims are self-contained empirical results, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims rely on hand-chosen hyperparameters (ranks, resolutions, channels), domain assumptions about indoor scene frequency content, and the assumption that ESLAM's optimization pipeline transfers to the new representation. No new physical entities are introduced; the Six-axis decomposition is a representational construct.

free parameters (5)

CP rank for geometry (k_g) = 2
Hand-chosen in implementation details; controls the capacity of the geometry field and is central to the claim that CP is sufficient for indoor geometry.
SA rank for appearance (k_a) = 16
Hand-chosen; controls appearance capacity and the trade-off between parameter count and color fidelity.
Coarse and fine feature resolutions = 24 cm coarse; 6 cm fine geometry; 3 cm fine appearance
Chosen by hand; affects stored resolution and reconstruction detail.
Feature channels C = 32
Fixed for all axis-aligned tensors; a design choice that sets the feature vector length.
Sampling counts N_s and N_t = N_s=32 or 48, N_t=8 per ray
Chosen per dataset; affects runtime and rendering quality.

assumptions (4)

domain assumption The geometry of indoor scenes is sufficiently low-frequency that a CP decomposition of rank 2 can encode it for tracking and reconstruction.
Invoked in Section 4.2 as the first reason for using CP for geometry; if false, mapping accuracy degrades.
domain assumption The Six-axis decomposition is sufficiently expressive to represent scene appearance at the given resolutions.
Used throughout Section 4.2 for the appearance field; no proof that rank 16 captures arbitrary color fields.
ad hoc to paper ESLAM's sampling, loss, and optimization pipeline transfers unchanged to the new representation.
Section 4.2 says all experiments are based on the ESLAM framework with a different scene representation; the compatibility of low-rank gradients with the SLAM losses is assumed.
ad hoc to paper Early convergence of CP geometry helps appearance optimization through the geometric dependency in Eq. 6.
Stated as a reason for the hybrid choice in Section 4.2 and supported only by the ablation, not by an independent mechanism.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LRSLAM: Low-rank Representation of Signed Distance Fields in Dense Visual SLAM System." pith.science (2026). https://pith.science/paper/L3W2Q3EU

@misc{pith2026250610567,
  author       = {Pith},
  title        = {Pith review of: LRSLAM: Low-rank Representation of Signed Distance Fields in Dense Visual SLAM System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3W2Q3EU}},
  note         = {Machine review of arXiv:2506.10567}
}

read the original abstract

Simultaneous Localization and Mapping (SLAM) has been crucial across various domains, including autonomous driving, mobile robotics, and mixed reality. Dense visual SLAM, leveraging RGB-D camera systems, offers advantages but faces challenges in achieving real-time performance, robustness, and scalability for large-scale scenes. Recent approaches utilizing neural implicit scene representations show promise but suffer from high computational costs and memory requirements. ESLAM introduced a plane-based tensor decomposition but still struggled with memory growth. Addressing these challenges, we propose a more efficient visual SLAM model, called LRSLAM, utilizing low-rank tensor decomposition methods. Our approach, leveraging the Six-axis and CP decompositions, achieves better convergence rates, memory efficiency, and reconstruction/localization quality than existing state-of-the-art approaches. Evaluation across diverse indoor RGB-D datasets demonstrates LRSLAM's superior performance in terms of parameter efficiency, processing time, and accuracy, retaining reconstruction and localization quality. Our code will be publicly available upon publication.

Figures

Figures reproduced from arXiv: 2506.10567 by the authors.

**Figure 1.** (a) Comparison with SOTA Approaches. Our model is clearly positioned as an efficient yet effective dense visual SLAM method. Our proposed LRSLAM requires fewer parameters, faster frame processing time (FPT), and better reconstruction accuracy (regarding ATE mean and RMSE). This is further validated by our (b) Qualitative Comparison of Scene Reconstruction and Localization between our proposed LRSLAM and the state-… view at source ↗

**Figure 2.** (a) Tri-plane representation factorizes a 4D tensor into three feature planes (Eq. 1). (b) CP decomposition factorizes a 4D tensor into a sum of outer products of three axis-aligned low-rank feature tensors (Eq. 2). (c) Our proposed Six-axis decomposition factorizes a 4D tensor into a set of six axis-aligned low-rank feature tensors (Eq. 3). scale 3D reconstructions by utilizing neural implicit representations. Rec… view at source ↗

**Figure 3.** An overview of our proposed model, called LRSLAM. Our model utilizes a combination of low-rank tensor decomposition methods to provide a better convergence rate, memory efficiency, and reconstruction quality. Specifically, we utilize the CP decomposition to represent the geometry of a scene (see top, fg) and use our Sixaxis decomposition for reconstructing color (see bottom, fa). where f (i) yz , f(i) zy , f(i) zx… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Reconstruction and Localization on ScanNet [7]. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Reconstruction on (a) TUM RGB-D [24] and (b) Replica [23]. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Ablation Study. Localization and reconstruction accuracy comparison between variants of our model with different combinations of scene geometry and appearance representations, i.e., CP decomposition and Six-axis (SA) decomposition. For example, the model [A] uses CP …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages

[1]

Bresson,G.,Alsayed,Z.,Yu,L.,Glaser,S.:Simultaneouslocalizationandmapping: Asurveyofcurrenttrendsinautonomousdriving.IEEETransactionsonIntelligent Vehicles2(3), 194–220 (2017)

work page 2017
[2]

eckart-young

Carroll, J.D., Chang, J.J.: Analysis of individual differences in multidimensional scaling via an n-way generalization of “eckart-young” decomposition. Psychome- trika35(3), 283–319 (1970)

work page 1970
[3]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123–16133 (2022)

2022
[4]

In: European Conference on Computer Vision

Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333–350. Springer (2022)

2022
[5]

In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Chen, X., Milioto, A., Palazzolo, E., Giguere, P., Behley, J., Stachniss, C.: Suma++: Efficient lidar-based semantic slam. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4530–4537. IEEE (2019)

work page 2019
[6]

In: 2020 22nd Symposium on Virtual and Augmented Reality (SVR)

Covolan, J.P.M., Sementille, A.C., Sanches, S.R.R.: A mapping of visual slam al- gorithms and their applications in augmented reality. In: 2020 22nd Symposium on Virtual and Augmented Reality (SVR). pp. 20–29. IEEE (2020)

work page 2020
[7]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828–5839 (2017)

2017
[8]

Robotics10(4), 110 (2021)

Dworakowski, D., Thompson, C., Pham-Hung, M., Nejat, G.: A robot architec- ture using contextslam to find products in unknown crowded retail environments. Robotics10(4), 110 (2021)

work page 2021

Show all 32 references

[9]

Artificial intelligence review43, 55–81 (2015)

Fuentes-Pacheco,J.,Ruiz-Ascencio,J.,Rendón-Mancha,J.M.:Visualsimultaneous localization and mapping: a survey. Artificial intelligence review43, 55–81 (2015)

2015
[10]

Advanced Robotics29(20), 1289–1301 (2015)

Gui,J.,Gu,D.,Wang,S.,Hu,H.:Areviewofvisualinertialodometryfromfiltering and optimisation perspectives. Advanced Robotics29(20), 1289–1301 (2015)

2015
[11]

In: 2019 international con- ference on robotics and automation (ICRA)

Huang, G.: Visual-inertial navigation: A concise review. In: 2019 international con- ference on robotics and automation (ICRA). pp. 9572–9582. IEEE (2019)

2019
[12]

arXiv preprint arXiv:2311.16728 (2023)

Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular, stereo, and rgb-d cameras. arXiv preprint arXiv:2311.16728 (2023)

2023 arXiv
[13]

Virtual Reality & Intelligent Hardware1(4), 386–410 (2019)

Jinyu, L., Bangbang, Y., Danpeng, C., Nan, W., Guofeng, Z., Hujun, B.: Survey and evaluation of monocular visual-inertial slam algorithms for augmented reality. Virtual Reality & Intelligent Hardware1(4), 386–410 (2019)

2019
[14]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Johari, M.M., Carta, C., Fleuret, F.: Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17408–17419 (2023)

2023
[15]

arXiv preprint arXiv:2312.02126 (2023)

Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)

2023 arXiv
[16]

ACM Transactions on Graphics42(4) (2023)

Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (2023)

2023
[17]

arXiv preprint arXiv:2402.03246 (2024)

Li, M., Liu, S., Zhou, H.: Sgs-slam: Semantic gaussian splatting for neural dense slam. arXiv preprint arXiv:2402.03246 (2024)

2024 arXiv
[18]

Robotics9(1), 19 (2020) 16 H

Liu, C., Zhou, C., Cao, W., Li, F., Jia, P.: A novel design and implementation of autonomous robotic car based on ros in indoor scenario. Robotics9(1), 19 (2020) 16 H. Park et al

2020
[19]

arXiv preprint arXiv:2312.06741 (2023)

Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)

2023 arXiv
[20]

Commu- nications of the ACM65(1), 99–106 (2021)

Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

2021
[21]

Robotics10(3), 87 (2021)

Ruan, K., Wu, Z., Xu, Q.: Smart cleaner: A new autonomous indoor disinfection robot for combating the covid-19 pandemic. Robotics10(3), 87 (2021)

2021
[22]

In: 2019 third IEEE international conference on robotic com- puting (IRC)

Singandhupe, A., La, H.M.: A review of slam techniques and security in au- tonomous driving. In: 2019 third IEEE international conference on robotic com- puting (IRC). pp. 602–607. IEEE (2019)

2019
[23]

arXiv preprint arXiv:1906.05797 (2019)

Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)

2019 arXiv
[24]

In: 2012 IEEE/RSJ international conference on intelligent robots and systems

Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573–580. IEEE (2012)

2012
[25]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and position- ing in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229–6238 (2021)

2021
[26]

IPSJ Transactions on Computer Vision and Applications9(1), 1–11 (2017)

Taketomi, T., Uchiyama, H., Ikeda, S.: Visual slam algorithms: A survey from 2010 to 2016. IPSJ Transactions on Computer Vision and Applications9(1), 1–11 (2017)

2017
[27]

arXiv preprint arXiv:2402.13255 (2024)

Tosi, F., Zhang, Y., Gong, Z., Sandström, E., Mattoccia, S., Oswald, M.R., Poggi, M.: How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255 (2024)

2024 arXiv
[28]

arXiv preprint arXiv:2311.11700 (2023)

Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)

2023 arXiv
[29]

Intelligent Industrial Systems 1(4), 289–311 (2015)

Yousif, K., Bab-Hadiashar, A., Hoseinnezhad, R.: An overview to visual odometry and visual slam: Applications to mobile robotics. Intelligent Industrial Systems 1(4), 289–311 (2015)

2015
[30]

arXiv preprint arXiv:2312.10070 (2023)

Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)

2023 arXiv
[31]

IEEE Access 9, 21367–21387 (2021)

Zhang, S., Zheng, L., Tao, W.: Survey and evaluation of rgb-d slam. IEEE Access 9, 21367–21387 (2021)

2021
[32]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786– 12796 (2022)

2022

Pith tools

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

[1] [1]

Bresson,G.,Alsayed,Z.,Yu,L.,Glaser,S.:Simultaneouslocalizationandmapping: Asurveyofcurrenttrendsinautonomousdriving.IEEETransactionsonIntelligent Vehicles2(3), 194–220 (2017)

work page 2017

[2] [2]

eckart-young

Carroll, J.D., Chang, J.J.: Analysis of individual differences in multidimensional scaling via an n-way generalization of “eckart-young” decomposition. Psychome- trika35(3), 283–319 (1970)

work page 1970

[3] [3]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123–16133 (2022)

2022

[4] [4]

In: European Conference on Computer Vision

Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333–350. Springer (2022)

2022

[5] [5]

In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

Chen, X., Milioto, A., Palazzolo, E., Giguere, P., Behley, J., Stachniss, C.: Suma++: Efficient lidar-based semantic slam. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4530–4537. IEEE (2019)

work page 2019

[6] [6]

In: 2020 22nd Symposium on Virtual and Augmented Reality (SVR)

Covolan, J.P.M., Sementille, A.C., Sanches, S.R.R.: A mapping of visual slam al- gorithms and their applications in augmented reality. In: 2020 22nd Symposium on Virtual and Augmented Reality (SVR). pp. 20–29. IEEE (2020)

work page 2020

[7] [7]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828–5839 (2017)

2017

[8] [8]

Robotics10(4), 110 (2021)

Dworakowski, D., Thompson, C., Pham-Hung, M., Nejat, G.: A robot architec- ture using contextslam to find products in unknown crowded retail environments. Robotics10(4), 110 (2021)

work page 2021