Pith. sign in

REVIEW 3 major objections 6 minor 48 references

LEGO-SLAM is the first framework to achieve real-time, open-vocabulary mapping inside Gaussian Splatting SLAM, by compressing language features into a 16-dimensional space.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 21:11 UTC pith:TIXWVB6K

load-bearing objection A credible 3DGS-SLAM system with 16-dim language features and integrated pruning/loop closure; the main weakness is unspecified offline pretraining data that may overlap the evaluation scenes. the 3 major comments →

arxiv 2511.16144 v2 pith:TIXWVB6K submitted 2025-11-20 cs.CV cs.RO

LEGO-SLAM: Language-Embedded Gaussian Optimization SLAM

Sibaek Lee , Seongbo Ha , Kyeongsu Kang , Joonyeol Choi , Seungjun Tak , Hyeonwoo Yu This is my paper
classification cs.CV cs.RO
keywords Open-Vocabulary SLAM3D Gaussian Splattingscene-adaptive autoencoderlanguage embedding compressionlanguage-guided pruninglanguage-based loop closurereal-time RGB-D SLAMfeature distillation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to prove that open-vocabulary semantic maps can be built online, in real time, inside a 3D Gaussian Splatting SLAM system. Existing language-embedded 3D reconstructions store high-dimensional features per Gaussian, which makes them too slow and memory-heavy for SLAM, and their static models do not adapt to new scenes. LEGO-SLAM answers this with a scene-adaptive autoencoder that distills high-dimensional language embeddings into a compact 16-dimensional feature per Gaussian, enabling direct text-query comparison in that compact space. The same features are reused for semantic pruning of redundant Gaussians and for language-based loop detection, so no separate models or extra feature extraction are needed. If correct, this would make language-annotated maps an online byproduct of SLAM rather than an offline processing step, at speeds of about 15 frames per second.

Core claim

LEGO-SLAM extends each 3D Gaussian with a 16-dimensional language feature, initialized by passing the keyframe's high-dimensional language map through a pretrained, then scene-adaptive, encoder. The map is optimized by joint RGB, depth, and feature distillation losses: the compact feature map is rendered, decoded by a lightweight convolutional decoder, and compared to the full 512-dimensional features of a frozen guidance model. After the Gaussian features converge, the encoder is periodically adapted online by freezing the map and decoder and minimizing the gap between encoder output and the rendered features. This yields a compact, scene-specific language space in which text queries can be

What carries the argument

The central mechanism is the scene-adaptive autoencoder, which compresses a high-dimensional language feature (e.g., 512-dim) into a 16-dim feature per Gaussian. The decoder is used only for distillation; queries are projected into the compact space by the encoder. This compact space carries the argument: it makes real-time rendering possible, enables the language-guided pruning rule (a neighbor Gaussian is pruned if within a distance threshold and above a cosine-similarity threshold), and provides the histogram signatures for loop detection via an offline language codebook.

Load-bearing premise

The load-bearing premise is that the offline language codebook used for loop detection is built from data representative of the environments the system will see; the paper does not say which scenes produced the millions of feature vectors, so if the codebook came from the same scenes used for evaluation, the loop-detection advantage over a position-based baseline would not necessarily generalize to novel environments.

What would settle it

Take two disjoint sets of indoor RGB-D sequences. Build the language codebook from set A only, run LEGO-SLAM on set B with all other settings unchanged, and compare tracking error and loop-candidate recall against the reported numbers. If the results degrade significantly, the loop-detection advantage came from codebook overlap with the test data, not from generalizable language-based place recognition.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Open-vocabulary 3D maps become an online capability, available during the SLAM run instead of after a separate reconstruction step.
  • Gaussian memory and rendering costs drop enough to sustain roughly 15 FPS, making language-embedded maps practical for real-time robotic systems.
  • Language-guided pruning reduces map size by over half while preserving rendering quality, compared with aggressive geometric pruning that degrades sharply.
  • Loop closure can reuse the mapping features, so place recognition does not require a separate visual descriptor network.
  • Tracking accuracy remains competitive with state-of-the-art RGB-D SLAM systems that do not build semantic maps, on synthetic and real-world benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The compact feature space could double as a general query interface: any text projected through the encoder can localize objects directly in the map, so the same framework could support language-conditioned robotic tasks such as 'pick up the red mug' without extra modules.
  • The scene-adaptive encoder points to a broader recipe: any dense per-pixel embedding (affordances, object states, material properties) could be compressed online into a low-dimensional map using a frozen guidance model, extending open-vocabulary SLAM beyond language alone.
  • A natural next step is to make the loop-closure codebook online-updatable, removing the offline k-means step; the paper leaves this extension unstated, but the histogram mechanism would likely tolerate incremental cluster refinement.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. LEGO-SLAM proposes a 3D Gaussian Splatting-based RGB-D SLAM system in which each Gaussian carries a compact 16-dimensional language feature. A scene-adaptive encoder-decoder distills LSeg features into this low-dimensional space, with a pretrained encoder used for initialization and online adaptation. The compact features are also used for language-guided pruning of redundant Gaussians and for histogram-based loop detection using a k-means codebook. Experiments on Replica, TUM-RGBD, and ScanNet report tracking accuracy of 0.20 cm, 2.30 cm, and 8.68 cm ATE; mapping PSNR of 36.38, 23.86, and 19.44 dB; open-vocabulary segmentation accuracy of 0.882, 0.834, and 0.791; Gaussian-count reductions up to roughly 71%; and 15 FPS operation.

Significance. If reproducible, the contribution is significant: it is among the first demonstrations of open-vocabulary language features embedded directly in a real-time 3DGS SLAM system. The ablations target the right questions (initialization, feature dimension, pruning, loop detection), and the evaluation is honest in noting that LEGO-SLAM uses estimated poses while the semantic baselines use ground-truth poses. The compact 16-d representation also enables scaling to ScanNet, where the 512-d Feature 3DGS baseline fails. Two caveats temper the significance: the pretraining data for the autoencoder and codebook are unspecified, which bears on the claimed generalization to unseen scenes, and all results are single-run, so the small performance differences reported may not be stable. The open-vocabulary metric is defined with respect to LSeg features, so it measures preservation of a fixed embedding space rather than generalization to arbitrary text outside LSeg's vocabulary; this is a standard protocol but should be stated as a limitation.

major comments (3)
  1. [Sec. 4.1 / Sec. 3.4] The pretraining data for the autoencoder are not stated (Sec. 4.1 only says it is 'pretrained as an autoencoder'), and the language codebook is 'generated offline by applying k-means clustering to high-dimensional language features' on 'millions of feature vectors' without specifying their source. If the same Replica/TUM/ScanNet scenes used in Tabs. 1-9 supplied these vectors, then the encoder prior and the loop-detection vocabulary are fitted to the test distribution, and the central claim of adaptation to unseen scenes (Sec. 1, Sec. 3.2) plus the Tab. 9 advantage over the position-based baseline are not established. Please specify the exact pretraining corpus and demonstrate generalization by pretraining on scenes disjoint from all evaluation scenes, or by evaluating on a held-out environment not used for codebook construction.
  2. [Tabs. 1-9] All quantitative results are single runs with no error bars or repeated-seed statistics. Many headline differences are small (e.g., Tab. 2 average ATE of 0.20 cm vs. LoopSplat's 0.25 cm; Tab. 8 PSNR of 35.21 vs. 36.14 without pruning), so it is unclear whether the reported improvements or 'maintaining quality' claims are significant. Please report mean and standard deviation over multiple runs, or explicitly state that the pipeline is deterministic and fix all seeds, and provide the resulting values.
  3. [Tab. 5] FPS is reported as 15.0 on all three datasets despite very different map sizes (Tab. 8: about 645K, 761K, and 2063K Gaussians before pruning). This pattern suggests the frame rate is capped, synchronized, or measured with a methodology that differs from end-to-end online SLAM throughput. Since real-time operation is a headline contribution, specify how FPS was measured (whether tracking, mapping, pruning, and loop detection are all included) and explain why the value is identical across datasets with different map sizes.
minor comments (6)
  1. [Abstract vs. Tab. 8] The arXiv abstract states 'up to 58%' Gaussian reduction, the full-text abstract says 'over 60%', and Tab. 8 implies reductions of about 40.8%, 71.4%, and 55.0% (computed as 1 - #GS after / #GS before). Please align these numbers and state the range accurately.
  2. [Sec. 2] Typo: 'H0owever' should be 'However'.
  3. [Sec. 3.4] The phrase 'spatially proximate past keyframes' is ambiguous. Is proximity based on current estimated pose, map position, or covisibility? This matters for defining the loop-candidate set.
  4. [Sec. 4.1] The encoder-decoder is described only as 'lightweight 1x1 convolutional layers'. Please provide the number of layers, channels, the pretraining objective, and optimizer/hyperparameters for reproducibility.
  5. [Sec. 4.2] The open-vocabulary evaluation uses LSeg features as ground truth, so the metric measures how well the map reconstructs the LSeg embedding space, not generalization to arbitrary text categories outside LSeg's vocabulary. Please state this as a limitation or include a qualitative experiment with free-form queries.
  6. [Eq. (4) and Sec. 3.4] The symbol τ_sim is used both for the pruning similarity threshold (Eq. 4) and for the loop-candidate histogram similarity threshold (Sec. 3.4). Clarify whether these are the same threshold or different quantities.

Circularity Check

0 steps flagged

No significant circularity; core pipeline is a self-contained distillation/loop-closure system benchmarked externally, with one missing-data caveat.

full rationale

LEGO-SLAM's core derivation is not circular. Map features f are optimized by Eq. (2) to reconstruct frozen LSeg features F_gt, and the scene-adaptive encoder E_phi is trained by Eq. (3) to map F_gt to the learned compact F_render. This is a conventional distillation/autoencoder consistency loop whose semantic content originates from the external LSeg model, not from the paper's own predictions. The open-vocabulary evaluation follows the Feature 3DGS protocol and reports mIoU/Accuracy, which test how well the LSeg-aligned feature space transfers to semantic labels rather than re-measuring the training loss. Tracking (G-ICP against the Gaussian map), pruning (geometric + cosine-similarity redundancy), and loop detection (k-means codebook histograms + G-ICP verification) are each compared against external baselines (Tabs. 2-9) and do not reduce to fitted parameters renamed as predictions. The references to the authors' prior RGBD GS-ICP SLAM [7] and LAMP [15] are contextual citations, not load-bearing uniqueness or ansatz-importation arguments. The main caveat is experimental rather than circular: Sec. 4.1 does not disclose the data used to pretrain the autoencoder or to build the offline language codebook from 'millions of feature vectors'. If those data included the evaluation scenes, the claimed generalization to unseen scenes and the Tab. 9 loop-detection advantage would be weakened. That is a missing experimental condition, not a derivation that is equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on the expressiveness of the 16-dim space (tuned post hoc on the test sets), the validity of LSeg/CLIP features as the language ground truth, and the unstated source domain of the k-means codebook. No new physical entities are introduced; the scene-adaptive encoder is a trained module with no independent evidence outside the paper.

free parameters (6)
  • language feature dimension d = 16
    Selected after ablation on the three evaluation datasets (Tab. 7); d=8 gives higher PSNR, d>=32 gives lower accuracy; 16 is the post-hoc best trade-off.
  • pruning distance threshold tau_dist = not reported
    Eq. (4); value never given, needed to reproduce pruning.
  • pruning similarity threshold tau_sim = not reported (0.7 used for loop detection)
    Eq. (4) and Sec. 3.4; the pruning threshold is hand-set and unreported.
  • loss weights w_depth, w_feat = not reported
    Eq. (1); no values given for the joint loss.
  • loop codebook size k = 64
    Sec. 4.1; k-means over "millions of feature vectors"; source data unspecified.
  • K (nearest neighbors for pruning) = not reported
    Sec. 3.3; neighborhood size for the redundancy check is not given.
axioms (5)
  • domain assumption LSeg per-pixel embeddings are a valid ground-truth language representation for both distillation and evaluation.
    The entire semantic objective (Eq. 2) and the open-vocabulary metric use LSeg features; the 'open-vocabulary' space is bounded by LSeg/CLIP alignment.
  • domain assumption G-ICP can robustly register depth maps against the Gaussian map using per-Gaussian covariances as a point cloud.
    Tracking (Sec. 3.1) relies on this; no failure analysis for texture-less or specular regions.
  • ad hoc to paper The 16-dimensional compact space is expressive enough for language queries after distillation.
    This is the central compression hypothesis, validated only by the internal ablation (Tab. 7) on the same benchmark scenes.
  • domain assumption The offline k-means codebook generalizes from its (unspecified) training features to the evaluation scenes.
    Sec. 3.4; the source of the 'millions of feature vectors' is not stated, so the histogram-based loop detection might be fitted to the test distribution.
  • domain assumption No post-run optimization is applied; reported maps come from the online SLAM process.
    Asserted in Sec. 4.1; this is a self-imposed protocol statement that the authors verify but external groups cannot.

pith-pipeline@v1.3.0-alltime-deepseek · 12690 in / 14117 out tokens · 126684 ms · 2026-08-03T21:11:10.547658+00:00 · methodology

0 comments
read the original abstract

Recent advances in 3D Gaussian Splatting (3DGS) have enabled Simultaneous Localization and Mapping (SLAM) systems to build photorealistic maps. However, these maps lack the open-vocabulary semantic understanding required for robotic interaction. Integrating language features into SLAM remains a significant challenge, as storing high-dimensional features incurs excessive memory and rendering overhead, while existing methods with static models lack adaptability for novel environments. We propose LEGO-SLAM (Language-Embedded Gaussian Optimization SLAM), a framework that achieves real-time, open-vocabulary mapping within a 3DGS-based SLAM system. At the core of our method is a scene-adaptive autoencoder that distills high-dimensional language embeddings into a compact 16-dimensional feature space, reducing the memory per Gaussian and accelerating rendering. Unlike static approaches, our encoder adapts online to unseen scenes. These compact features also enable a language-guided pruning strategy that identifies semantic redundancy, reducing the map's Gaussian count by up to 58% while maintaining rendering quality. Furthermore, we introduce a language-based loop detection approach that reuses the language features already extracted for mapping, eliminating the need for a separate detection model. Experiments demonstrate that LEGO-SLAM achieves competitive mapping quality and tracking accuracy, all while providing open-vocabulary capabilities at 15 FPS. Our project page is available at https://lab-of-ai-and-robotics.github.io/LEGO-SLAM/

Figures

Figures reproduced from arXiv: 2511.16144 by Hyeonwoo Yu, Joonyeol Choi, Kyeongsu Kang, Seongbo Ha, Seungjun Tak, Sibaek Lee.

Figure 1
Figure 1. Figure 1: LEGO-SLAM: real-time, open-vocabulary 3DGS-SLAM. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System Overview. LEGO-SLAM architecture, where the Tracking module estimates pose and the Mapping module optimizes the 3D Gaussian Map via language distillation. This map is refined by Language Pruning and Loop Detection, enabling 3D Object Localization. is that we directly track against our unified 3D Gaussian map, eliminating the need for a separate tracking map as seen in decoupled systems [24]. Pose Es… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Mapping Comparison. We compare the rendered maps of LEGO-SLAM against baselines on the TUM-RGBD, and ScanNet datasets. All maps shown are captured directly from the online SLAM process without any post-run op￾timization [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scene-Adaptive Encoder Adaptation. Our Adaptive, online-tuned encoder generates accurate relevancy maps for 3D object queries, while the Frozen baseline fails. Impact of Feature Dimension. We investigate the impact of our compact feature dimension (d) on mapping quality, semantic accuracy, and memory us￾age in Tab. 7. Selecting the feature dimension involves a critical trade-off. A larger d increases memor… view at source ↗
Figure 5
Figure 5. Figure 5: Pruning Performance Comparison. As the pruning ratio increases, our language-guided method shows significantly less degradation in rendering quality com￾pared to the geometric approach on the Replica Room0 scene [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 7 linked inside Pith

  1. [1]

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

    Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn ar- chitecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297–5307 (2016)

    2016

  2. [2]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  3. [3]

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

    Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1316–1326 (2023)

    2023

  4. [4]

    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

  5. [5]

    Georgia Institute of Technology, Tech

    Dellaert, F.: Factor graphs and gtsam: A hands-on introduction. Georgia Institute of Technology, Tech. Rep2(4) (2012)

    2012

  6. [6]

    In: European conference on computer vision

    Ghiasi,G.,Gu,X.,Cui,Y.,Lin,T.Y.:Scalingopen-vocabularyimagesegmentation with image-level labels. In: European conference on computer vision. pp. 540–557. Springer (2022)

    2022

  7. [7]

    In: European Conference on Computer Vision

    Ha, S., Yeon, J., Yu, H.: Rgbd gs-icp slam. In: European Conference on Computer Vision. pp. 180–197. Springer (2024)

    2024

  8. [8]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Hanson, A., Tu, A., Singla, V., Jayawardhana, M., Zwicker, M., Goldstein, T.: Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5949–5958 (2025)

    2025

  9. [9]

    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

  10. [10]

    In: European Conference on Computer Vision

    Kant,Y.,Ramachandran,A.,Yenamandra,S.,Gilitschenski,I.,Batra,D.,Szot,A., Agrawal, H.: Housekeep: Tidying virtual households using commonsense reasoning. In: European Conference on Computer Vision. pp. 355–373. Springer (2022) 16 S. Lee et al

    2022

  11. [11]

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

    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. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21357–21366 (2024)

    2024

  12. [12]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  13. [13]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language em- bedded radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 19729–19739 (2023)

    2023

  14. [14]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023)

    2023

  15. [15]

    IEEE Robotics and Automation Letters (2025)

    Lee, S., Yu, H., Kim, G., Choi, S.: Lamp: Implicit language map for robot naviga- tion. IEEE Robotics and Automation Letters (2025)

    2025

  16. [16]

    arXiv preprint arXiv:2201.03546 (2022)

    Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546 (2022)

    Pith/arXiv arXiv 2022

  17. [17]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Li, K., Niemeyer, M., Navab, N., Tombari, F.: Dns-slam: Dense neural semantic- informed slam. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7839–7846. IEEE (2024)

    2024

  18. [18]

    In: European Conference on Computer Vision

    Li, M., Liu, S., Zhou, H., Zhu, G., Cheng, N., Deng, T., Wang, H.: Sgs-slam: Semantic gaussian splatting for neural dense slam. In: European Conference on Computer Vision. pp. 163–179. Springer (2024)

    2024

  19. [19]

    IEEE Transactions on Circuits and Systems for Video Technology (2024)

    Liao, G., Zhou, K., Bao, Z., Liu, K., Li, Q.: Ov-nerf: Open-vocabulary neural radiance fields with vision and language foundation models for 3d semantic un- derstanding. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  20. [20]

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

    Liso, L., Sandström, E., Yugay, V., Van Gool, L., Oswald, M.R.: Loopy-slam: Dense neural slam with loop closures. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20363–20373 (2024)

    2024

  21. [21]

    In: European conference on computer vision

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In: European conference on computer vision. pp. 38–55. Springer (2024)

    2024

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 18039–18048 (2024)

    2024

  23. [23]

    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

  24. [24]

    IEEE transactions on robotics33(5), 1255–1262 (2017)

    Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monoc- ular, stereo, and rgb-d cameras. IEEE transactions on robotics33(5), 1255–1262 (2017)

    2017

  25. [25]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

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

    Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: Langsplat: 3d language gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20051–20060 (2024)

    2024

  27. [27]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from Abbreviated paper title 17 natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  28. [28]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021)

    2021

  29. [29]

    In: 2020 IEEE International Conference on Robotics and Automation (ICRA)

    Rosinol, A., Abate, M., Chang, Y., Carlone, L.: Kimera: an open-source library for real-time metric-semantic localization and mapping. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 1689–1696. IEEE (2020)

    2020

  30. [30]

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

    Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433–18444 (2023)

    2023

  31. [31]

    In: Robotics: science and sys- tems

    Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and sys- tems. vol. 2, p. 435. Seattle, WA (2009)

    2009

  32. [32]

    arXiv preprint arXiv:2210.05663 (2022)

    Shafiullah, N.M.M., Paxton, C., Pinto, L., Chintala, S., Szlam, A.: Clip- fields: Weakly supervised semantic fields for robotic memory. arXiv preprint arXiv:2210.05663 (2022)

    Pith/arXiv arXiv 2022

  33. [33]

    In: Conference on robot learning

    Shah, D., Osiński, B., Levine, S., et al.: Lm-nav: Robotic navigation with large pre- trained models of language, vision, and action. In: Conference on robot learning. pp. 492–504. PMLR (2023)

    2023

  34. [34]

    In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS)

    Shan, T., Englot, B., Meyers, D., Wang, W., Ratti, C., Rus, D.: Lio-sam: Tightly- coupled lidar inertial odometry via smoothing and mapping. In: 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 5135–5142. IEEE (2020)

    2020

  35. [35]

    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)

    Pith/arXiv arXiv 1906

  36. [36]

    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

  37. [37]

    arXiv preprint arXiv:2402.132554, 1 (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:2402.132554, 1 (2024)

    Pith/arXiv arXiv 2024

  38. [38]

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

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

    2024

  39. [39]

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

    Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10371–10381 (2024)

    2024

  40. [40]

    Advances in Neural Information Processing Systems37, 5850–5869 (2024)

    Ye, Z., Wan, C., Li, C., Hong, J., Li, S., Li, L., Zhang, Y., Lin, Y.C.: 3d gaus- sian rendering can be sparser: Efficient rendering via learned fragment pruning. Advances in Neural Information Processing Systems37, 5850–5869 (2024)

    2024

  41. [41]

    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)

    Pith/arXiv arXiv 2023

  42. [42]

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

    Zhang,Y.,Tosi,F.,Mattoccia,S.,Poggi,M.:Go-slam:Globaloptimizationforcon- sistent 3d instant reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3727–3737 (2023)

    2023

  43. [43]

    Advances in Neural Information Processing Systems37, 122434–122457 (2024) 18 S

    Zhang, Z., Song, T., Lee, Y., Yang, L., Peng, C., Chellappa, R., Fan, D.: Lp- 3dgs: Learning to prune 3d gaussian splatting. Advances in Neural Information Processing Systems37, 122434–122457 (2024) 18 S. Lee et al

    2024

  44. [44]

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

    Zhou,S.,Chang,H.,Jiang,S.,Fan,Z.,Zhu,Z.,Xu,D.,Chari,P.,You,S.,Wang,Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21676–21685 (2024)

    2024

  45. [45]

    In: 2025 International Conference on 3D Vision (3DV)

    Zhu, L., Li, Y., Sandström, E., Huang, S., Schindler, K., Armeni, I.: Loopsplat: Loop closure by registering 3d gaussian splats. In: 2025 International Conference on 3D Vision (3DV). pp. 156–167. IEEE (2025)

    2025

  46. [46]

    arXiv preprint arXiv:2403.07494 (2024)

    Zhu, S., Qin, R., Wang, G., Liu, J., Wang, H.: Semgauss-slam: Dense semantic gaussian splatting slam. arXiv preprint arXiv:2403.07494 (2024)

    Pith/arXiv arXiv 2024

  47. [47]

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

    Zhu, S., Wang, G., Blum, H., Liu, J., Song, L., Pollefeys, M., Wang, H.: Sni-slam: Semantic neural implicit slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21167–21177 (2024)

    2024

  48. [48]

    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