Pith. sign in

REVIEW 4 major objections 5 minor 39 references

DINO-SLAM: DINO-informed RGB-D SLAM for Neural Implicit and Explicit Representations

T0 review · 4 major / 5 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 Self-supervised DINO features, made geometry-aware by a Scene Structure Encoder, give NeRF- and 3DGS-based dense RGB-D SLAM better reconstruction, rendering, and tracking on Replica, ScanNet, and TUM.

desk verdict A useful empirical result on DINO features for neural SLAM, but the central SSE component's training is unspecified and the key claim is not reproducible as written. read the letter →

arxiv 2507.19474 v2 pith:3Q5M4OJK submitted 2025-07-25 cs.CV

Ziren Gong , Xiaohan Li , Fabio Tosi , Youmin Zhang , Stefano Mattoccia , Jun Wu , Matteo Poggi This is my paper
classification cs.CV
keywords DINO-SLAMdenseRGB-DSLAMneuralimplicit3DGaussianSplattingself-supervisedvisiontransformersscenestructureencoderEDINOfeaturestri-planeencoding
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

DINO-SLAM sets out to show that semantic features from a self-supervised vision transformer (DINO) can improve dense RGB-D SLAM if they are first made geometry-aware. The authors argue that raw DINO features alone give only marginal gains, because they lack 3D spatial understanding, so they add a Scene Structure Encoder (SSE) that mixes DINO features with depth and appearance features into enhanced EDINO encodings. These EDINO features supervise a tri-plane field in a NeRF-style pipeline and augment each Gaussian in a 3D Gaussian Splatting pipeline. On Replica, ScanNet, and TUM, the paper reports better reconstruction completeness, rendering quality, and pose accuracy than the compared methods, with ablations showing the gains require both DINO and the geometry enrichment. If correct, the result is a plug-in recipe: any radiance-field SLAM representation can carry semantic-structure supervision without labels.

What carries the argument

The load-bearing component is the Scene Structure Encoder (SSE), a small learned module $\Phi_{\mathrm{SSE}}$ that maps color and depth to two outputs: raw DINO features $f_d$ and enhanced EDINO features $f_{ed}$. Inside it, an MLP encodes DINO features into an appearance stream $f_a$ and another MLP encodes depth into a geometric stream $f_g$; two attention steps re-weight the DINO features using geometry-appearance similarity and then re-weight the appearance stream using the refined DINO features. The final EDINO feature is the concatenation of the refined appearance, refined DINO, and geometric streams. What this module does for the argument is transform a semantic prior that knows about objects and parts into a representation that also knows where surfaces are, so the same features can supervise geometry-carrying quantities such as tri-planes and Gaussian parameters.

What would settle it

Run the NeRF variant on three Replica scenes with the SSE's attention modules replaced by fixed random projections of the same inputs, keeping all other losses and hyperparameters identical; if tracking error, completion rate, and L1 depth do not clearly worsen, the claimed gain does not come from learned geometry enrichment.

Watch

Extended reading notes

Core claim

The central claim is that incorporating DINO-derived continuous features into neural implicit and explicit scene representations improves SLAM across mapping, rendering, and tracking, provided they are enriched with geometry. Vanilla DINO features encode object- and scene-level structure but not metric 3D layout, so the SSE processes color and depth into appearance and geometric feature streams, refines them with two attention operations against the DINO features, and concatenates them into EDINO features. In the NeRF variant, EDINO features directly supervise a dedicated tri-plane $\tau_d$ through an $\ell^1$ loss while DINO features supervise the rendered feature map; in the 3DGS variant, each Gaussian carries an EDINO feature channel and the rasterized feature map is supervised with DINO features. Reported experiments on Replica, ScanNet, and TUM show the full system outperforms the compared NeRF- and 3DGS-based SLAM baselines, and the ablations show that removing either DINO or EDINO degrades results, with raw DINO alone performing worse than the complete SSE.

Load-bearing premise

The load-bearing premise is that the scene-structure encoder's internal parameters can actually be learned well enough during the SLAM process, because the paper does not say how those parameters are trained; if they are not, the reported gains reduce to whatever raw DINO features provide on their own.

Editorial extensions

If this is right

  • Any NeRF- or 3DGS-based SLAM system that can render an auxiliary feature map can absorb the SSE by adding two losses, one tying the rendered feature to DINO and one tying the encoded geometry to EDINO.
  • Semantic supervision for SLAM no longer requires discrete labels or a predefined category set, since DINO provides continuous, self-supervised features.
  • The reported gains are not representation-specific: the paper shows the SSE improves hash-grid, tri-plane, neural-point, and 3D Gaussian baselines on Replica.
  • If the reported tracking gains hold on the long ScanNet sequences, feature-level supervision plus global bundle adjustment is a workable recipe for drift control in large scenes.

Reading between the lines

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

  • The paper leaves monocular input untested; since the SSE consumes depth only as a geometric stream, a depth-prediction network could supply that stream, which would extend the method to monocular SLAM if the added supervision remains stable.
  • The attention equations in the SSE are a generic cross-modal fusion pattern, so the same EDINO loss could plausibly be applied to other radiance-field backbones beyond those tested, such as low-rank or sparse-grid encodings.
  • The paper does not state how the SSE's MLP and attention parameters are trained; the most informative follow-up would compare jointly optimized SSE weights against frozen, pretrained ones, since the claim's generality depends on that choice.
  • Because EDINO features define a continuous 3D field, one could query them at arbitrary points for downstream tasks like open-vocabulary segmentation or object retrieval without retraining the map.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DINO-SLAM, a general strategy for injecting DINO self-supervised visual features into dense RGB-D SLAM systems, implemented for both neural implicit (NeRF-based) and explicit (3D Gaussian Splatting) scene representations. A Scene Structure Encoder (SSE) takes an RGB-D frame and raw DINO features and outputs enriched EDINO features, which are then used to supervise tri-plane features in the NeRF pipeline and to augment per-Gaussian attributes in the 3DGS pipeline. The authors report experiments on Replica, ScanNet, and TUM, showing improvements over several existing dense SLAM baselines, and ablations indicate that both raw DINO and EDINO components contribute to the gains.

Significance. If fully specified and reproducible, this would be a useful contribution to dense neural SLAM: it proposes a relatively light-weight way to leverage a self-supervised vision transformer for geometry- and appearance-aware scene encoding, and it demonstrates the recipe on two representation families and multiple base encodings. The generality experiments in Table 4 and the dual NeRF/3DGS implementation are notable strengths. The central claim is plausible, but it is currently undercut by the absence of a training protocol for the SSE and by an overbroad abstract claim; once these are addressed, the method could be of interest to the SLAM and scene representation community.

major comments (4)
  1. [Section 3.1 and Sections 3.2-3.3 (Mapping and Tracking)] The training protocol for the Scene Structure Encoder (SSE) is never specified. The contribution list states the method works 'without ... offline training', yet the mapping/tracking optimization lists in §3.2 and §3.3 include only the tri-planes, decoder MLPs, and camera poses (NeRF) or the Gaussian parameters and poses (3DGS); the SSE parameters Φ_SSE (the MLPs F_ed, F_eg, F_ea and the two attention modules) appear in no optimization list. The losses L_ef (Eq. 7) and L_df (Eq. 14) use f_ed as a target or embedding, so they do not by themselves train the SSE. The authors must state whether the SSE is (a) jointly optimized online as part of SLAM, in which case its initialization, learning rate, and loss terms must be given; (b) pre-trained offline on some dataset, which contradicts the no-offline-training claim; or (c) frozen at a random or heuristic initialization, in which case the claim that the attention/MLP machinery 'captures structural relationships' is unsupported. This is load-bearing because the ablations in Table 5 attribute the performance gains to EDINO features, and without a training specification the reader cannot tell whether those gains come from a learned encoder, an untrained feature transform, or an online overfit.
  2. [Eq. (7) in Section 3.2] The encoding feature loss L_ef = ||τ_d - f_ed||_1 is not well-defined as written. τ_d is a tri-plane feature grid (three 2D planes at the resolutions given in Implementation Details), while f_ed is a per-pixel 2D feature map produced by the SSE for the current RGB-D frame. The paper does not explain how a per-pixel 2D target is compared to a 3D tri-plane representation: is f_ed unprojected into 3D and sampled at the tri-plane coordinates? Is the tri-plane rendered into 2D via the ray-casting process before the loss is applied? Or is f_ed duplicated across some canonical volume? This is not a minor notation issue: the central NeRF pipeline trains τ_d with this loss, and the loss as written has incompatible domains.
  3. [Abstract and Tables 1-3] The abstract's claim of 'superior performance on the Replica, ScanNet, and TUM compared to state-of-the-art methods' is not supported on all metrics and in all comparisons. On Replica (Table 1, left), the NeRF variant's accuracy (1.94 cm) is worse than Point-SLAM (1.41 cm) and its depth L1 (0.66 cm) is worse than Point-SLAM (0.44 cm). On ScanNet (Table 2, Avg column), PLGSLAM (6.8) outperforms DINO-SLAM (NeRF) (7.1) in ATE RMSE. On TUM (Table 3), MonoGS performs better than DINO-SLAM (3DGS) on fr1 desk and fr3 office. The paragraph in §4.1 discussing PLGSLAM partially acknowledges the ScanNet case, but the abstract and the conclusion remain unqualified. The authors should either soften the claim to 'competitive or superior on most metrics' or explicitly quantify where the method is not best.
  4. [Section 3.3, Eqns. (10) and (12)] The description of how EDINO features are attached to Gaussians is inconsistent. The text says 'we incorporate the EDINO features f_ed obtained from our SSE into each Gaussian's parameters', but Eq. (10), which defines the Gaussian, contains no f_ed term; f_ed appears only in the rendering equation for f'_d (Eq. 12). Is f_ed a per-Gaussian feature vector that is stored as an attribute and optimized via the DINO-feature loss L_df (Eq. 14)? If so, please state how such a feature is initialized and updated. Additionally, L_df = ||f'_d - f_d||_1 pushes the EDINO-rendered feature map toward the raw DINO features f_d, which appears to conflict with the stated goal that EDINO is an enriched representation beyond raw DINO. Please explain why this supervision direction is appropriate and how the method avoids collapsing EDINO back to DINO.
minor comments (5)
  1. [Tables 1-5] No standard deviations or repeated-run statistics are reported anywhere; this is especially important for the small differences in Table 5 (e.g., PSNR 21.43 vs. 21.50) and for the general claim that EDINO helps. Please report the number of runs and error bars, or at least state explicitly that each result is from a single run.
  2. [Section 3.2, Eqns. (5)-(6)] The notation in Eqns. (5) and (6) is inconsistent: the rendered color, depth, and DINO features are written with the subscript n (as if per-sample) but are defined as sums over n; they should be written without the subscript (e.g., ĉ, d̂, f'_d).
  3. [Table 1 caption] The caption says 'The best results are highlighted as first, second, and third', but in the table as presented there is no visible bolding or color. Please make the highlighting explicit (e.g., bold, underline, or colored entries).
  4. [Section 4, Implementation Details] For the 3DGS pipeline only the DINO-feature loss weight (1) is given; the weights for the color and depth losses and any additional terms in the total loss are not specified. Please state the full loss composition and weights for both pipelines.
  5. [Throughout] There are several typos and formatting artifacts: 'componentes' in the Conclusion, 'represenation' in the PLGSLAM reference, and unrendered LaTeX hash symbols in Eqs. (5), (8), and (14). These should be corrected before publication.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: benchmark results rest on external datasets and standard metrics; self-citations are non-load-bearing, and the unspecified SSE training protocol is a reproducibility gap, not a circular reduction.

full rationale

The central claim — that EDINO features improve reconstruction, rendering, and tracking over state-of-the-art methods — is tested against external benchmarks only. The reported numbers (Tables 1-3) are standard metrics on Replica, ScanNet, and TUM (ATE RMSE, accuracy, completion, completion rate, depth L1, PSNR, SSIM, LPIPS) defined by each benchmark's ground truth, not by any quantity the method optimizes, so no predicted quantity is equivalent by construction to the method's inputs. Equations (7), (8), and (14) are training losses that supervise the map's feature grids (τ_d, f'_d) to match DINO/EDINO features extracted from the input frames; this is a self-supervised feature-fitting loop during mapping, and none of these quantities is reported as an evaluation metric. The self-citations are minor and non-load-bearing: (Tosi et al. 2024), a survey with overlapping authorship, is cited for 'standard SLAM metrics' and for 'more detailed literature', but the metrics are defined by the benchmark papers, not by that survey; (Gong et al. 2025) and (Zhang et al. 2023) appear only as related-work context. No uniqueness theorem from the authors' prior work is invoked, and no ansatz is smuggled in via citation. Per the missing-support review rule, one gap is flagged but is not circular: the Scene Structure Encoder of Section 3.1 contains trainable MLPs (F_ed, F_eg, F_ea) and attention modules, yet Section 3.2 states that mapping 'jointly optimizes three components: the tri-planes..., the decoder MLPs, and camera poses' (SSE parameters excluded), while the contributions claim operation 'without requiring semantic annotations, pre-defined categories, or offline training'; the paper never specifies whether the SSE is randomly initialized, pre-trained, or optimized online, so the attribution of the measured gains to learned structural relationships is unverifiable. This is a completeness and reproducibility defect, not a circular derivation: the ablation outcomes in Tables 4 and 5 are empirical and would stand or fall independently of how the SSE is trained. The abstract's SGE/geoDINO terminology, which never reappears in the body, is an internal inconsistency rather than a circular step.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The central empirical claim rests on DINO as a pretrained feature source, an unspecified training protocol for the SSE, and several hand-set loss weights and thresholds. No new physical entities are introduced, but the SSE is an invented component whose independent evidence is limited to the paper's own experiments.

free parameters (5)
  • DINO-feature loss weight (3DGS) = 1
    Set in Implementation Details; selected by hand to balance feature supervision.
  • Encoding feature loss weight (NeRF) = 5.0
    Set for the EDINO-to-triplane loss in Implementation Details.
  • DINO-feature loss weight (NeRF) = 0.01
    Set for the rendered DINO feature loss in Implementation Details.
  • Keyframe motion thresholds = d_l=0.7, r_l=15
    Used to select keyframes in the 3DGS pipeline, following but not derived from MonoGS.
  • Tri-plane resolutions = 24cm/6cm geometry, 24cm/3cm appearance/DINO
    Multi-resolution grid sizes chosen for the NeRF pipeline in Implementation Details.
assumptions (3)
  • domain assumption DINO features encode semantic and structural priors useful for scene representation.
    The entire method relies on DINO pretrained features carrying information beyond color and depth; Section 3.1.
  • ad hoc to paper The SSE attention mechanism effectively fuses DINO, appearance, and geometry features.
    The attention equations in Section 3.1 are proposed specifically for this paper and no evidence is given that they generalize beyond the tested datasets.
  • domain assumption Existing ESLAM and MonoGS pipelines are valid backbones for measuring gains.
    The NeRF and 3DGS variants build directly on ESLAM and MonoGS and their losses and rendering equations.
invented entities (1)
  • Scene Structure Encoder with EDINO features
    purpose: To enrich raw DINO features with geometric and appearance relationships for SLAM supervision.
    The SSE is a newly introduced learned module; the paper provides no external validation or training details beyond its own benchmark results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DINO-SLAM: DINO-informed RGB-D SLAM for Neural Implicit and Explicit Representations." pith.science (2026). https://pith.science/paper/3Q5M4OJK

@misc{pith2026250719474,
  author       = {Pith},
  title        = {Pith review of: DINO-SLAM: DINO-informed RGB-D SLAM for Neural Implicit and Explicit Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Q5M4OJK}},
  note         = {Machine review of arXiv:2507.19474}
}
read the original abstract

This paper presents DINO-SLAM, a DINO-informed design strategy to enhance implicit (Neural Radiance Field -- NeRF) and explicit representations (Gaussian Splatting -- GS) in SLAM systems through the more comprehensive semantics understanding enabled by DINO. This latter alone, however, lacks proper 3D geometry understanding, allowing only for marginal improvements. Therefore, we rely on a Scene Geometry Encoder (SGE) to enrich DINO features into geometry-aware DINO features (geoDINO), to better understand those geometric relationships that vanilla DINO features fail to capture. Building upon it, we propose two foundational paradigms for NeRF and GS SLAM systems integrating geoDINO features. Compared to state-of-the-art methods, our DINO-informed pipelines achieve superior performance on the Replica, ScanNet, and TUM datasets.

Figures

Figures reproduced from arXiv: 2507.19474 by the authors.

Figure 1
Figure 1. Mapping, rendering, and tracking results on Replica room1 (Straub et al. 2019) and TUM fr2 xyz (Sturm et al. 2012). Our method produces the accurate 3D reconstruction shown in (a), along with more precise camera tracking in (c). Compared to existing methods such as PLGSLAM, ESLAM, SplaTAM, and MonoGS, DINO-SLAM better preserves completeness and detail (see red boxes and blue boxes), achieving superior reconstruction… view at source ↗
Figure 2
Figure 2. Overview of DINO-SLAM for neural implicit or explicit representations. Both neural representations share the same structure of Scene Structure Encoder (SSE) to capture enhanced-DINO (EDINO) and DINO features. In our neural implicit pipeline (NeRF), EDINO features provide supervision for the tri-plane optimization (red double arrow) while DINO features serve roles in guiding the optimization of the estimated DINO fea… view at source ↗
Figure 3
Figure 3. The architecture of Scene Structure Encoder (SSE). Our SSE has two outputs (in yellow): 1) DINO features fd: Given the RGB frames, the self-supervised vision transformer (DINO) extracts self-attention features as one of the outputs. 2) EDINO features fed: We separately encode DINO features fd and depth information into appearance fa and geometric features fg. These features, along with the extracted DINO features, u… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Architectures of DINO-informed neural implicit (NeRF) and explicit (3DGS) SLAM pipelines. In the NeRF￾based pipeline (left), we leverage EDINO features fed to supervise the tri-plane τd (red double arrow) while DINO features fd guide the optimization of the estimated D…
Figure 5
Figure 5. Figure 5: Visualization of DINO-SLAM and baselines on the Replica. We present details of reconstruction and rendering quality with red boxes and blue boxes. The two left columns present the reconstruction performance of our DINO-SLAM (NeRF). Our method yields superior mesh resul…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Campos, C.; Elvira, R.; Rodr \' guez, J. J. G.; Montiel, J. M.; and Tard \'o s, J. D. 2021. Orb-slam3: An accurate open-source library for visual, visual--inertial, and multimap slam. IEEE transactions on robotics, 37(6): 1874--1890

    work page 2021

  4. [4]

    Caron, M.; Touvron, H.; Misra, I.; J \'e gou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 9650--9660

    2021

  5. [5]

    X.; Savva, M.; Halber, M.; Funkhouser, T.; and Nie ner, M

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

    work page 2017

  6. [6]

    Deng, T.; Shen, G.; Qin, T.; Wang, J.; Zhao, W.; Wang, J.; Wang, D.; and Chen, W. 2024. Plgslam: Progressive neural scene represenation with local to global bundle adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19657--19666

    work page 2024

  7. [7]

    Gong, Z.; Tosi, F.; Zhang, Y.; Mattoccia, S.; and Poggi, M. 2025. HS-SLAM: Hybrid Representation with Structural Supervision for Improved Dense SLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)

    work page 2025

  8. [8]

    Ha, S.; Yeon, J.; and Yu, H. 2024. Rgbd gs-icp slam. In European Conference on Computer Vision, 180--197. Springer

    work page 2024

Show all 39 references
  1. [9]

    Hu, J.; Chen, X.; Feng, B.; Li, G.; Yang, L.; Bao, H.; Zhang, G.; and Cui, Z. 2024. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. In European Conference on Computer Vision, 93--112. Springer

    2024

  2. [10]

    Hu, J.; Mao, M.; Bao, H.; Zhang, G.; and Cui, Z. 2023. Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems, 36: 39429--39442

    2023

  3. [11]

    Huang, H.; Li, L.; Cheng, H.; and Yeung, S.-K. 2024. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21584--21593

    2024

  4. [12]

    M.; Carta, C.; and Fleuret, F

    Johari, M. M.; Carta, C.; and Fleuret, F. 2023. 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, 17408--17419

    2023

  5. [13]

    M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J

    Keetha, N.; Karhade, J.; Jatavallabhula, K. M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J. 2024. SplaTAM: Splat Track & Map 3D Gaussians for Dense RGB-D SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21357--21366

    2024

  6. [14]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  7. [15]

    Li, G.; Chen, Q.; Yan, Y.; and Pu, J. 2024. EC-SLAM: Effectively Constrained Neural RGB-D SLAM with Sparse TSDF Encoding and Global Bundle Adjustment. arXiv preprint arXiv:2404.13346

    2024 arXiv

  8. [16]

    Li, Y.; Brasch, N.; Wang, Y.; Navab, N.; and Tombari, F. 2020. Structure-slam: Low-drift monocular slam in indoor environments. IEEE Robotics and Automation Letters, 5(4): 6583--6590

    2020

  9. [17]

    Liso, L.; Sandstr \"o m, E.; Yugay, V.; Van Gool, L.; and Oswald, M. R. 2024. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20363--20373

    2024

  10. [18]

    H.; and Davison, A

    Matsuki, H.; Murai, R.; Kelly, P. H.; and Davison, A. J. 2024. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18039--18048

    2024

  11. [19]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  12. [20]

    Mur-Artal, R.; Montiel, J. M. M.; and Tardós, J. D. 2015. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics, 31(5): 1147--1163

    2015

  13. [21]

    Mur-Artal, R.; and Tard \'o s, J. D. 2017. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5): 1255--1262

    2017

  14. [22]

    A.; Lovegrove, S

    Newcombe, R. A.; Lovegrove, S. J.; and Davison, A. J. 2011. DTAM: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, 2320--2327. IEEE

    2011

  15. [23]

    Park, H.; Park, M.; Nam, G.; and Kim, J. 2024. LRSLAM: Low-Rank Representation of Signed Distance Fields in Dense Visual SLAM System. In European Conference on Computer Vision, 225--240. Springer

    2024

  16. [24]

    Peng, Z.; Shao, T.; Liu, Y.; Zhou, J.; Yang, Y.; Wang, J.; and Zhou, K. 2024. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, 1--11

    2024

  17. [25]

    F.; Newcombe, R

    Salas-Moreno, R. F.; Newcombe, R. A.; Strasdat, H.; Kelly, P. H.; and Davison, A. J. 2013. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1352--1359

    2013

  18. [26]

    Sandstr \"o m, E.; Li, Y.; Van Gool, L.; and Oswald, M. R. 2023. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 18433--18444

    2023

  19. [27]

    J.; Mur-Artal, R.; Ren, C.; Verma, S.; et al

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

    2019 arXiv

  20. [28]

    Sturm, J.; Engelhard, N.; Endres, F.; Burgard, W.; and Cremers, D. 2012. A benchmark for the evaluation of RGB-D SLAM systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, 573--580. IEEE

    2012

  21. [29]

    Sucar, E.; Liu, S.; Ortiz, J.; and Davison, A. J. 2021. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, 6229--6238

    2021

  22. [30]

    Tateno, K.; Tombari, F.; Laina, I.; and Navab, N. 2017. Cnn-slam: Real-time dense monocular slam with learned depth prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, 6243--6252

    2017

  23. [31]

    Teed, Z.; and Deng, J. 2021. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras . In Advances in neural information processing systems

    2021

  24. [32]

    R.; and Poggi, M

    Tosi, F.; Zhang, Y.; Gong, Z.; Sandstr \"o m, E.; Mattoccia, S.; Oswald, M. R.; and Poggi, M. 2024. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4: 1

    2024 arXiv

  25. [33]

    Wang, H.; Wang, J.; and Agapito, L. 2023. Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13293--13302

    2023

  26. [34]

    Xin, Z.; Yue, Y.; Zhang, L.; and Wu, C. 2024. Hero-slam: Hybrid enhanced robust optimization of neural slam. In 2024 IEEE International Conference on Robotics and Automation (ICRA), 8610--8616. IEEE

    2024

  27. [35]

    Yan, C.; Qu, D.; Xu, D.; Zhao, B.; Wang, Z.; Wang, D.; and Li, X. 2024. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19595--19604

    2024

  28. [36]

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

    2023 arXiv

  29. [37]

    Zhang, Y.; Tosi, F.; Mattoccia, S.; and Poggi, M. 2023. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3727--3737

    2023

  30. [38]

    Zheng, J.; Zhu, Z.; Bieri, V.; Pollefeys, M.; Peng, S.; and Armeni, I. 2025. Wildgs-slam: Monocular gaussian splatting slam in dynamic environments. In Proceedings of the Computer Vision and Pattern Recognition Conference, 11461--11471

    2025

  31. [39]

    R.; and Pollefeys, M

    Zhu, Z.; Peng, S.; Larsson, V.; Xu, W.; Bao, H.; Cui, Z.; Oswald, M. R.; and Pollefeys, M. 2022. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12786--12796

    2022

Pith tools

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