Pith. sign in

REVIEW 4 major objections 7 minor 40 references

LEG-SLAM: Real-Time Language-Enhanced Gaussian Splatting for SLAM

T0 review · 4 major / 7 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 The paper claims LEG-SLAM is the first real-time SLAM system that unifies language features, open-vocabulary segmentation, and Gaussian-splatting mapping, running at more than 10 fps on Replica and 18 fps on ScanNet.

desk verdict The speed claim is real and the integration is a genuine systems contribution, but the semantic evaluation is contaminated by PCA trained on the test categories, so the open-vocabulary claim is not yet backed by clean numbers. read the letter →

arxiv 2506.03073 v1 pith:5J47VCGT submitted 2025-06-03 cs.CV

Roman Titkov , Egor Zubkov , Dmitry Yudin , Jaafar Mahmoud , Malik Mohrat , Gennady Sidorov This is my paper
classification cs.CV
keywords 3DGaussianSplattingSimultaneousLocalizationandMappingopen-vocabularysegmentationDINOv2language-enhancedPCAfeaturecompressionRGB-DSLAMreal-timesemantic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to prove that a single online SLAM system can build a dense 3D Gaussian scene that is both photorealistic and queryable by arbitrary text, with no precomputed semantic map or offline preparation. The central claim is that LEG-SLAM is the first real-time system to do this, running at more than 10 frames per second on Replica and 18 frames per second on ScanNet while segmenting scenes by text queries such as “vase” or “chair.” It attributes this to an efficient combination of DINOv2 feature extraction, PCA-based compression of semantic features, and a Gaussian-splatting mapper that optimizes color, depth, and language features together. If the claim holds, interactive applications such as robots and augmented reality could ask a map where an object is while the map is being built, at sensor frame rates instead of after minutes of offline processing.

What carries the argument

The central object is the semantic 3D Gaussian: each Gaussian’s parameters are augmented with a K-dimensional language-feature vector, initialized to zero and updated during mapping. The shared differentiable rasterizer is the mechanism that carries the argument, rendering color, depth, and semantic features in one pass so that semantic consistency and geometry are optimized on the same cloud. The DINOv2 feature encoder supplies the semantic signal; a PCA projector, trained on CLIP-projected, Talk2DINO-aligned ImageNet classes with noise augmentation, compresses 768-dimensional maps to 64 dimensions with negligible cosine-similarity loss; and Talk2DINO, a learned CLIP-to-DINOv2 bridge, lets arbitrary text queries enter the same compressed space. A coarse-to-fine Gaussian pyramid accelerates convergence, and ORB-SLAM3 provides poses.

What would settle it

Train the PCA projector on a held-out set of categories (for example, only 10 of the 20 ScanNet classes, or ImageNet classes with ScanNet categories removed) and evaluate mIoU on all 20 classes; if held-out categories show a clear drop well below the 41.4 mIoU reported when the compressor saw the evaluation classes, the generalization claim behind the open-vocabulary numbers is falsified. A second check is to run the full pipeline with a 10 Hz RGB-D stream and verify end-to-end per-frame latency under 100 ms including a text query, since the real-time claim depends on that budget.

Watch

Extended reading notes

Core claim

LEG-SLAM promises a unified pipeline in which each 3D Gaussian carries not only position, scale, orientation, and color but also a compressed K-dimensional DINOv2 language-feature vector, with K set to 64 in the final configuration. These semantic vectors are optimized jointly with the geometric and photometric parameters, so a single differentiable rasterizer renders RGB, depth, and a semantic feature map at the same time. At query time, a CLIP text embedding is mapped into DINOv2 space by Talk2DINO, compressed with the same PCA projector, and matched against the rendered feature map by dot product. The authors report that this yields open-vocabulary segmentation on ScanNet at 41.4 mIoU, reconstruction at about 32 dB PSNR on Replica, and a roughly 30x speedup over the compared language-Gaussian baselines, with a full ScanNet scene processed in about 1.5 minutes.

Load-bearing premise

The load-bearing premise is that the PCA compressor trained to reconstruct the 20 ScanNet semantic categories generalizes to arbitrary text queries, because the reported 41.4 mIoU is measured with a compressor that has already seen those same 20 categories; if that transfer fails, the open-vocabulary accuracy is unmeasured.

Editorial extensions

If this is right

  • A robot or AR device could query a scene while mapping it, without a pre-built semantic map or precomputed poses, which the paper argues is the first such capability.
  • Scene-level open-vocabulary analysis that previously took tens of minutes to hours (LERF, LangSplat, Feature3DGS) would drop to about 1.5 minutes per ScanNet scene, making interactive querying practical.
  • Because semantic features live inside the Gaussian cloud, the same map can be re-queried for any number of text concepts without rerunning segmentation for each query.
  • The 64-dimensional compressed features keep segmentation near 41.4 mIoU while preserving real-time rendering, so language-enhanced maps can be built at sensor frame rate with modest GPU memory.

Reading between the lines

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

  • If the PCA compressor truly learns a general projection rather than the 20 benchmark classes, the same 64-dimensional features should serve queries outside the training labels; a direct test would train PCA on a disjoint label set and measure held-out mIoU.
  • The timing profile suggests feature extraction (about 32 ms at 640x480) dominates the 57 ms frame budget, so faster visual backbones or lower-resolution extraction would translate almost directly into higher frame rates or lower-power operation.
  • The paper’s Figure 2 names ORB-SLAM3 as the tracker, but the bibliography contains no ORB-SLAM3 citation; locating and citing that component matters for reproducing the SLAM accuracy and speed numbers, and the absence is a support gap rather than a refutation of the speed claim.
  • Because the semantic vectors are optimized online, a natural extension is to refine the map from user text feedback during exploration, for example by using a query heatmap as a pseudo-label, which the paper does not explore.
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 / 7 minor

Summary. LEG-SLAM proposes a real-time dense RGB-D SLAM system that jointly renders color, depth, and language-semantic features by integrating ORB-SLAM3 tracking, DINOv2 visual features, a PCA-based feature compressor, Talk2DINO text-to-vision embedding alignment, and a 3D Gaussian Splatting mapper. The paper reports per-frame processing times (57.02 ms at 640x480, 95.21 ms at 1200x680), reconstruction quality on Replica, and open-vocabulary segmentation results on ScanNet, claiming more than 10 fps on Replica and 18 fps on ScanNet and up to a 30x speedup over prior language-augmented Gaussian Splatting methods.

Significance. If the quantitative claims hold, the speed component is a useful engineering contribution: the per-stage timing breakdown in Table 1 is internally consistent, and the ablations (Tables 4-5) provide a clear view of the accuracy/speed trade-off from PCA dimensionality. The paper also delivers a concrete system description and comparisons across multiple baselines. However, the central novelty is open-vocabulary semantic SLAM, and the evidence for that claim is weakened by a circularity in the ScanNet evaluation and by an apparent copy-paste error in the comparison table. The contributions are significant if the semantic evaluation is re-done cleanly, but they are not presently established beyond the real-time mapping component.

major comments (4)
  1. [Supplementary Material, ScanNet dataset; §4.4; §3.3] The ScanNet semantic evaluation uses a PCA compressor trained on exactly the same 20 categories used to compute mIoU and mAcc, as stated in the supplementary material. This contradicts the main-text claim (§3.3) that PCA is trained on 1,000 ImageNet text classes to learn a general projection. Since the same PCA projection is applied to both the visual embeddings and the text-query embedding (§3.5), training on the evaluation classes can only make dot-product matching easier for those classes. The reported 41.4 mIoU and 74.3 mAcc therefore do not measure transfer to unseen categories and cannot support the open-vocabulary claim. The authors must re-run the quantitative evaluation with a PCA trained on the 1,000 ImageNet classes (or another held-out class set) and report the resulting metrics, or clearly label the current numbers as closed-vocabulary results.
  2. [Table 2] The row for OVO-Gaussian-SLAM is numerically identical to the row for Gaussian-SLAM: ATE RMSE 0.31, Depth L1 0.68, PSNR 42.08, SSIM 0.996, LPIPS 0.018. Because Table 3 lists different semantic metrics for OVO-Gaussian-SLAM and the two methods are distinct, this appears to be a copy-paste error. The comparison is not trustworthy as presented. The authors must provide the correct OVO-Gaussian-SLAM SLAM metrics and clarify the source of these numbers.
  3. [§4.5; Table 2] The claim that LEG-SLAM is 'the only method in this comparison capable of integrating real-time SLAM with open-vocabulary scene understanding' is not supported by the table as printed because OVO-Gaussian-SLAM is also marked 'Open-Vocabulary'. Depending on the corrected OVO-Gaussian-SLAM row, the 'first real-time' claim may need to be re-scoped. A direct runtime and capability comparison with OVO-Gaussian-SLAM (and OVO-SLAM) under identical hardware and input conditions is needed to substantiate the speed/novelty claim.
  4. [§4.4; Table 3] The 'Training Time' comparison in Table 3 mixes settings: LangSplat, Feature3DGS, and Semantic Gaussians are per-scene offline optimization methods without SLAM tracking, while LEG-SLAM is an online SLAM system. Comparing 1.5 minutes per scene to those methods conflates unrelated workloads. The speed advantage should be reported against methods that operate in the same online SLAM setting, and the exact timing protocol (per-frame from Table 1 vs. total scene processing) should be stated.
minor comments (7)
  1. [Abstract; §3.3] The abstract describes PCA as a 'learnable feature compressor,' but PCA is normally a fixed linear projection fit once, not learned online. Please adjust the wording to 'learned' or 'fit' and use consistent terminology with §3.3.
  2. [§3.1] The text says the DINOv2 embeddings are 'compressed using an autoencoder,' but the actual method uses PCA. Please align the terminology to avoid confusion.
  3. [Figure 2 caption] The caption mentions a 'learning-based approach to a language-enhanced rasterizer for Gaussian Splatting,' but no learning of the rasterizer is described in the paper. Please clarify or rephrase.
  4. [Table 2 header] The 'Closed-V ocabulary' and 'Open-V ocabulary' entries contain an extra space due to a line break in the source. Please fix the formatting.
  5. [§4.6.1] The timing '33 ms' for DINOv2 + PCA is consistent with Table 1 (31.91 ms feature extraction plus 1.57 ms compression), but the sentence 'including embedding compression' could be clearer; please state the decomposition.
  6. [Supplementary Material, Table 2] The entries 'Counter not presented' and 'Person not presented' are ambiguous: do these classes not appear in the tested scenes, or are they excluded from evaluation due to annotation errors? Please clarify.
  7. [§3.4, step 1] The description 'the semantic vector H×W×K is assigned zero values' conflates an image-sized feature map with per-Gaussian vectors. Please rephrase to indicate that each Gaussian is initialized with a K-dimensional semantic vector of zeros.

Circularity Check

1 steps flagged · score 4.0 of 10

ScanNet mIoU is partially circular: the PCA compressor is trained on the same 20 categories used for the reported open-vocabulary mIoU.

  1. fitted input called prediction [Supplementary Material, Section 1 ('ScanNet dataset'); main text §4.4 Table 3; §3.3 and §3.5 describe the PCA use.]
    "To ensure effective feature compression, the same 20 categories were used to train the PCA encoder for ScanNet dataset evaluation. This alignment allows PCA to learn the optimal projection for retaining key semantic information while reducing computational complexity. However, since PCA learns a generalized compression strategy rather than memorizing specific class distributions, our method remains effective even on previously unseen categories."

    The paper's quantitative open-vocabulary evidence, 41.4 mIoU and 74.3 mAcc on ScanNet (main text Table 3), is produced by a compression projection that was itself fitted on the same 20 ScanNet categories being scored. Both the visual embeddings and the text query embeddings are passed through this PCA (§3.5 steps 3-4), so the dot-product matching is optimized for exactly the benchmark classes. The supplement explicitly states: 'the same 20 categories were used to train the PCA encoder for ScanNet dataset evaluation.' The claimed transfer to unseen categories is only asserted, not measured; the only unseen-category illustration, Figure 1 ('backpack'), is qualitative.

full rationale

The reconstruction-quality and frame-rate claims of LEG-SLAM are independent and plausible: Table 1 reports measured per-frame timings (57 ms at 640x480), and Table 2 compares rendering/ATE numbers against external baselines. No self-citation chain or uniqueness theorem is used to force the architecture. The circularity is concentrated in the semantic evaluation. The main text says PCA is trained on 1,000 ImageNet text classes 'beyond ImageNet classes by learning a general feature projection,' but the supplementary material admits that for the ScanNet benchmark the same 20 evaluation categories were also used to train the PCA encoder. Since both the DINOv2 visual embeddings and the CLIP text queries are compressed by this same fitted projection before dot-product comparison, the reported mIoU is partially circular: the compressor has been optimized on the label categories it is then used to recognize. The assertion that PCA 'learns a generalized compression strategy' is not quantitatively validated on held-out categories. This contaminates the open-vocabulary semantic claim but does not invalidate the independent real-time mapping contributions. Score 4 reflects partial circularity with substantial independent content remaining.

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

The central claim rests on pretrained modules (DINOv2, CLIP, Talk2DINO), a PCA compressor that is fitted to data, and an external ORB-SLAM3 front-end. The most fragile entry is the PCA projection: the supplement states that a version is trained on the very 20 categories used for the ScanNet mIoU evaluation, which weakens the open-vocabulary claim. No new entities are invented.

free parameters (3)
  • PCA embedding dimension K = 64 dims
    Selected by an ablation sweep (Table 5) as the trade-off point between mIoU, PSNR, VRAM, and train FPS. The choice is data-driven and affects both semantic quality and real-time performance.
  • PCA projection matrix for visual embeddings = Trained on 1,000 ImageNet text classes (main text) or on the 20 ScanNet categories (supplement)
    The projection is fitted to a distribution of text-based DINOv2 embeddings and is the key compression mechanism. The supplement confirms a variant trained on the same classes used for the ScanNet mIoU evaluation.
  • Loss weight lambda (color vs SSIM blend) = Not reported
    Equation (1) defines lambda as a blend of L1 and SSIM for the color term, but its value is not given in the paper. The semantic and depth terms have implicit weights of 1.
assumptions (4)
  • domain assumption DINOv2 features are semantic enough to support open-vocabulary segmentation after PCA compression to 64 dimensions.
    Invoked in Sec. 3.2-3.4; the system distills DINOv2 features into Gaussians and queries them with Talk2DINO text embeddings. The paper provides only indirect evidence via mIoU, and that evidence is weakened by the PCA training protocol.
  • domain assumption Talk2DINO correctly projects CLIP text embeddings into DINOv2 latent space and this alignment survives PCA compression.
    Invoked in Sec. 3.5; the text query is projected to DINOv2 space and compressed with the same PCA used for visual features. No evaluation of the alignment quality after compression is provided beyond the final mIoU.
  • domain assumption ORB-SLAM3 provides poses that are accurate enough for joint Gaussian optimization of color, depth, and semantic features.
    Figure 2 shows ORB-SLAM3 as the pose source, but the text does not describe how tracking errors propagate into the semantic map. The reported ATE RMSE 0.94 on Replica is higher than most baselines in Table 2.
  • standard math The 3D Gaussian optimization converges to a consistent semantic scene under the combined loss in Eq. (1).
    The loss combines L1 color, SSIM, cosine similarity on semantic features, and L1 depth. Convergence is assumed; no convergence analysis is provided, which is standard practice in 3DGS SLAM systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LEG-SLAM: Real-Time Language-Enhanced Gaussian Splatting for SLAM." pith.science (2026). https://pith.science/paper/5J47VCGT

@misc{pith2026250603073,
  author       = {Pith},
  title        = {Pith review of: LEG-SLAM: Real-Time Language-Enhanced Gaussian Splatting for SLAM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5J47VCGT}},
  note         = {Machine review of arXiv:2506.03073}
}
read the original abstract

Modern Gaussian Splatting methods have proven highly effective for real-time photorealistic rendering of 3D scenes. However, integrating semantic information into this representation remains a significant challenge, especially in maintaining real-time performance for SLAM (Simultaneous Localization and Mapping) applications. In this work, we introduce LEG-SLAM -- a novel approach that fuses an optimized Gaussian Splatting implementation with visual-language feature extraction using DINOv2 followed by a learnable feature compressor based on Principal Component Analysis, while enabling an online dense SLAM. Our method simultaneously generates high-quality photorealistic images and semantically labeled scene maps, achieving real-time scene reconstruction with more than 10 fps on the Replica dataset and 18 fps on ScanNet. Experimental results show that our approach significantly outperforms state-of-the-art methods in reconstruction speed while achieving competitive rendering quality. The proposed system eliminates the need for prior data preparation such as camera's ego motion or pre-computed static semantic maps. With its potential applications in autonomous robotics, augmented reality, and other interactive domains, LEG-SLAM represents a significant step forward in real-time semantic 3D Gaussian-based SLAM. Project page: https://titrom025.github.io/LEG-SLAM/

Figures

Figures reproduced from arXiv: 2506.03073 by the authors.

Figure 1
Figure 1. Simplified diagram of the proposed LEG-SLAM ap [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A detailed scheme of the developed LEG-SLAM method. Its distinctive features include a fast approach to obtaining visual [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of RGB and semantic heatmap outputs for different text queries. Each image contains both an RGB rendering (top) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of semantic segmentation performance across different pipelines on the Replica dataset, including ground truth for [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: VRAM consumption across different feature compressor [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 1
Figure 1. Figure 1: Comparison of PCA embedding compression on the [PITH_FULL_IMAGE:figures/full_fig_p012_1.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages

  1. [1]

    Talking to DINO: Bridging Self- Supervised Vision Backbones with Language for Open- V ocabulary Segmentation, 2024

    Luca Barsellotti, Lorenzo Bianchi, Nicola Messina, Fabio Carrara, Marcella Cornia, Lorenzo Baraldi, Fabrizio Falchi, and Rita Cucchiara. Talking to DINO: Bridging Self- Supervised Vision Backbones with Language for Open- V ocabulary Segmentation, 2024. arXiv:2411.19331 [cs]. 3, 4

    arXiv 2024

  2. [2]

    RGBDS-SLAM: A RGB-D Semantic Dense SLAM Based on 3D Multi Level Pyramid Gaussian Splatting

    Zhenzhong Cao, Chenyang Zhao, Qianyi Zhang, Jinzheng Guang, and Yinuo Song Jingtai Liu. RGBDS-SLAM: A RGB-D Semantic Dense SLAM Based on 3D Multi Level Pyramid Gaussian Splatting, 2024. arXiv:2412.01217 [cs]. 6

    work page Pith review arXiv 2024

  3. [3]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Niebner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Niebner. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes

    work page

  4. [4]

    Scal- ing Open-V ocabulary Image Segmentation with Image-Level Labels

    Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scal- ing Open-V ocabulary Image Segmentation with Image-Level Labels. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceed- ings, Part XXXVI, pages 540–557, Berlin, Heidelberg, 2022. Springer-Verlag. 2

    work page 2022

  5. [5]

    Semantic Gaussians: Open-V ocabulary Scene Understand- ing with 3D Gaussian Splatting, 2024

    Jun Guo, Xiaojian Ma, Yue Fan, Huaping Liu, and Qing Li. Semantic Gaussians: Open-V ocabulary Scene Understand- ing with 3D Gaussian Splatting, 2024. arXiv:2403.15624 [cs]. 2

    arXiv 2024

  6. [6]

    RGBD GS- ICP SLAM, 2024

    Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu. RGBD GS- ICP SLAM, 2024. eprint: 2403.12550. 2

    arXiv 2024

  7. [7]

    Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB- D Cameras

    Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Ye- ung. Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB- D Cameras. pages 21584–21593, 2024. 2, 6

    work page 2024

  8. [8]

    NEDS-SLAM: A Neural Explicit Dense Semantic SLAM Framework using 3D Gaussian Splatting.IEEE Robotics and Automation Letters, 9(10):8778–8785, 2024

    Yiming Ji, Yang Liu, Guanghu Xie, Boyu Ma, and Zongwu Xie. NEDS-SLAM: A Neural Explicit Dense Semantic SLAM Framework using 3D Gaussian Splatting.IEEE Robotics and Automation Letters, 9(10):8778–8785, 2024. arXiv:2403.11679 [cs]. 2, 6

    arXiv 2024

Show all 40 references
  1. [9]

    BrightDreamer: Generic 3D Gaussian Genera- tive Framework for Fast Text-to-3D Synthesis, 2024

    Lutao Jiang, Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. BrightDreamer: Generic 3D Gaussian Genera- tive Framework for Fast Text-to-3D Synthesis, 2024. eprint: 2403.11273. 2

    2024 arXiv

  2. [10]

    ESLAM: Efficient Dense SLAM System Based on Hybrid Representation of Signed Distance Fields

    Mohammad Mahdi Johari, Camilla Carta, and Francois Fleuret. ESLAM: Efficient Dense SLAM System Based on Hybrid Representation of Signed Distance Fields. In 2023 Ieee/Cvf Conference On Computer Vision And Pat- tern Recognition (Cvpr), pages 17408–17419. Los Alamitos,

    2023

  3. [11]

    SplaTAM: Splat, Track & Map 3D Gaus- sians for Dense RGB-D SLAM spla-tam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. SplaTAM: Splat, Track & Map 3D Gaus- sians for Dense RGB-D SLAM spla-tam . github . io. 2, 6

  4. [12]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics, 42(4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics, 42(4):1–14, 2023. 2

    2023

  5. [13]

    Weinberger, Serge Belongie, Vladlen Koltun, and Ren ´e Ranftl

    Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and Ren ´e Ranftl. Language-driven Semantic Seg- mentation, 2022. arXiv:2201.03546 [cs]. 2

    2022 arXiv

  6. [14]

    DNS SLAM: Dense Neural Semantic-Informed SLAM, 2023

    Kunyi Li, Michael Niemeyer, Nassir Navab, and Federico Tombari. DNS SLAM: Dense Neural Semantic-Informed SLAM, 2023. eprint: 2312.00204. 2

    2023 arXiv

  7. [15]

    GS 3 LAM: Gaussian Semantic Splatting SLAM

    Linfei Li, Lin Zhang, Zhong Wang, and Ying Shen. GS 3 LAM: Gaussian Semantic Splatting SLAM. InProceedings of the 32nd ACM International Conference on Multimedia, pages 3019–3027, Melbourne VIC Australia, 2024. ACM. 6

    2024

  8. [16]

    SGS-SLAM: Semantic Gaussian Splatting For Neural Dense SLAM

    Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. SGS-SLAM: Semantic Gaussian Splatting For Neural Dense SLAM. pages 163–179. 2025. arXiv:2402.03246 [cs]. 1, 2, 6

    2025 arXiv

  9. [17]

    Editing Conditional Radiance Fields

    Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing Conditional Radiance Fields. In2021 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 5753–5763, Mon- treal, QC, Canada, 2021. IEEE. 2

    2021

  10. [18]

    Dynamic 3D Gaussians: Tracking by Per- sistent Dynamic View Synthesis, 2023

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3D Gaussians: Tracking by Per- sistent Dynamic View Synthesis, 2023. arXiv:2308.09713 [cs]. 2

    2023 arXiv

  11. [19]

    Oswald, and Javier Civera

    Tomas Berriel Martins, Martin R. Oswald, and Javier Civera. OVO-SLAM: Open-V ocabulary Online Simultaneous Local- ization and Mapping, 2024. arXiv:2411.15043 [cs] version:

    2024

  12. [20]

    Hidenobu Matsuki, Riku Murai, Paul H. J. Kelly, and Andrew J. Davison. Gaussian Splatting SLAM, 2024. arXiv:2312.06741 [cs]. 6

    2024 arXiv

  13. [21]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, 2020. arXiv:2003.08934 [cs]. 2

    2020 arXiv

  14. [22]

    DINOv2: Learning Robust Visual Features without Supervision.Transactions on Machine Learning Research Journal, pages 1–31, 2024

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, and others. DINOv2: Learning Robust Visual Features without Supervision.Transactions on Machine Learning Research...

    2024

  15. [23]

    Langsplat: 3d language gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 1

    2024

  16. [24]

    Learning Transferable Vi- sual Models From Natural Language Supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Vi- sual Models From Natural Language Supervision, 2021. arXiv:2103.0...

    2021 arXiv

  17. [25]

    Panoptic Lifting for 3D Scene Understanding with Neural Fields, 2022

    Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bul ´o, Nor- man M ¨uller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic Lifting for 3D Scene Understanding with Neural Fields, 2022. arXiv:2212.09802 [cs]. 2

    2022 arXiv

  18. [26]

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gi...

    2019 arXiv

  19. [27]

    Co- SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM, 2023

    Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Co- SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM, 2023. arXiv:2304.14377 [cs]. 2, 6

    2023 arXiv

  20. [28]

    GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting, 2024

    Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting, 2024. arXiv:2311.11700 [cs]. 2

    2024 arXiv

  21. [29]

    V ox-Fusion: Dense Tracking and Mapping with V oxel-based Neural Implicit Represen- tation

    Xingrui Yang, Hai Li, Hongjia Zhai, Yuhang Ming, Yuqian Liu, and Guofeng Zhang. V ox-Fusion: Dense Tracking and Mapping with V oxel-based Neural Implicit Represen- tation. In2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 499–507, 2022. arXiv:22...

    2022 arXiv

  22. [30]

    GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models, 2024

    Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models, 2024. eprint: 2310.08529. 2

    2024 arXiv

  23. [31]

    Os- wald

    Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Os- wald. Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting, 2024. arXiv:2312.10070 [cs]. 6

    2024 arXiv

  24. [32]

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-Place Scene Labelling and Un- derstanding with Implicit Scene Representation, 2021. arXiv:2103.15875 [cs]. 2

    2021 arXiv

  25. [33]

    SemGauss-SLAM: Dense Semantic Gaus- sian Splatting SLAM, 2024

    Siting Zhu, Renjie Qin, Guangming Wang, Jiuming Liu, and Hesheng Wang. SemGauss-SLAM: Dense Semantic Gaus- sian Splatting SLAM, 2024. eprint: 2403.07494. 2

    2024 arXiv

  26. [34]

    SNI-SLAM: Semantic Neural Implicit SLAM, 2024

    Siting Zhu, Guangming Wang, Hermann Blum, Jium- ing Liu, Liang Song, Marc Pollefeys, and Hesheng Wang. SNI-SLAM: Semantic Neural Implicit SLAM, 2024. arXiv:2311.11016 [cs]. 2, 6

    2024 arXiv

  27. [35]

    Oswald, and Marc Polle- feys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Polle- feys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM, 2022. arXiv:2112.12130 [cs]. 6 LEG-SLAM: Language-Enhanced Gaussian Splatting for Real-Time SLAM Supple...

    2022 arXiv

  28. [38]

    Table 1 presents per- scene results, where PSNR, SSIM, and LPIPS measure re- construction quality, while mIoU and mAcc assess semantic segmentation accuracy

    ScanNet dataset To further analyze the performance of LEG-SLAM, we evaluate the reconstruction and semantic segmentation met- rics on individual ScanNet scenes. Table 1 presents per- scene results, where PSNR, SSIM, and LPIPS measure re- construction quality, while mIoU and mA...

    1914

  29. [39]

    Table 3 summarizes the results, providing a de- tailed breakdown of reconstruction quality across multiple environments

    Replica dataset To contextualize LEG-SLAM’s performance, we compare it against existing SLAM methods on different Replica scenes. Table 3 summarizes the results, providing a de- tailed breakdown of reconstruction quality across multiple environments. The table reports PSNR, SS...

  30. [40]

    The vi- sualizations illustrate how varying the dimensionality from 3 to 64 affects the ability to preserve fine-grained semantic details while maintaining computational efficiency

    Impact of Embedding Dimensionality on Se- mantic Quality To analyze the effect of embedding compression, we com- pare the semantic reconstruction quality across different PCA embedding dimensions, as shown in Figure 1. The vi- sualizations illustrate how varying the dimensiona...

  31. [2017]

    Accepted: 2021-10-08T19:50:29Z. 5

    2021

  32. [2023]

    ISSN: 1063-6919. 2, 6

Pith tools

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