Pith. sign in

REVIEW 4 major objections 5 minor 42 references

VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a 3D-Gaussian RGBD SLAM system can drop learned Gaussian positions entirely, anchoring each Gaussian to a depth pixel and learning only radius, opacity, and color, which frees memory and unlocks larger scenes.

desk verdict A clean view-tied Gaussian representation for RGBD SLAM, but the headline rendering gains largely reflect memorization of training views rather than scene quality. read the letter →

arxiv 2506.02741 v1 pith:XKQ4O3HO submitted 2025-06-03 cs.CV

Pengchong Hu , Zhizhong Han This is my paper
classification cs.CV
keywords RGBDSLAM3DGaussianSplattingview-tiedGaussianslarge-scalescenemappingcameratrackingdifferentiablerenderingrepresentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a 3D-Gaussian SLAM system does not need to learn or store where its Gaussians are. Instead, it ties each Gaussian to a pixel that has a valid depth reading, fixes the Gaussian's center at the back-projected depth point, and learns only the radius, opacity, and color of each Gaussian. Because each Gaussian drops from 14 parameters to 5, the system can keep far more Gaussians in limited GPU memory and can stop optimizing all past Gaussians, organizing frames into local sections instead of keyframes. The authors report that this design yields better rendering and competitive camera tracking on standard RGBD benchmarks while scaling to much larger scenes, including city-scale sequences.

What carries the argument

The load-bearing object is the view-tied 3D Gaussian: a spherical Gaussian whose center is the back-projection of a depth pixel through the camera pose, carrying only color, radius, and opacity as learnable attributes. Around it, the method's machinery is the section: a block of N consecutive frames whose head-frame Gaussian initialization and per-section optimization let the system keep the most relevant Gaussians learnable while freezing the rest, and whose visibility-based selection of an overlapping front section for head-frame tracking suppresses pose-error accumulation. These two mechanisms together replace the global joint optimization of all Gaussians over all keyframes that the authors identify as the scalability bottleneck of prior 3DGS SLAM.

What would settle it

Run the system on an RGBD sequence whose depth has a known artificial scale error (for example, multiply all depths by 1.02) and compare rendering PSNR and ATE against a version with untouched depth: if the biased-depth map shows a systematic geometric shift with no mechanism to correct it, the claim that frozen depth positions suffice would be put in question.

Watch

Extended reading notes

Core claim

The central claim is that Gaussian positions in RGBD SLAM can be determined entirely by the sensor depth and the camera pose, never learned. The paper's view-tied Gaussians are spheres anchored to depth pixels, so the representation stores color, radius, and opacity only, saving about 64.3% of per-Gaussian storage. Framing consecutive frames as sections, marking the first frame of each section as a head, tracking head frames against a selected overlapping previous section, and freezing all sections except the one containing the latest view removes the need for keyframe lists and for optimizing all Gaussians throughout training; the authors argue this is what allows the map to grow to much larger scenes while improving rendering quality and maintaining tracking accuracy.

Load-bearing premise

The load-bearing premise is that the sensor's depth values are an unbiased geometric prior: Gaussian centers are frozen at back-projected pixels, so any systematic depth bias or pose drift during tracking cannot be corrected by moving the map points.

Editorial extensions

If this is right

  • With 5 parameters per Gaussian instead of 14, the method reports using roughly 98 million total Gaussians on Replica while keeping at most 2.7 million in GPU memory at once, so scene coverage and local detail both rise.
  • Because only the section containing the latest view is optimized, the method does not need to re-render or re-optimize against a keyframe list, which the authors identify as the path to much larger scenes.
  • The reported rendering gains on real indoor datasets (TUM-RGBD, ScanNet, ScanNet++) indicate that tying Gaussians to depth does not overfit a synthetic benchmark.
  • On KITTI city-scale sequences in the supplementary material, the method reports markedly higher rendering PSNR and lower ATE than the same baselines, supporting the scalability claim.

Reading between the lines

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

  • Because positions are frozen, the map inherits any systematic depth-sensor bias; a natural extension would be to learn a small per-ray offset along the projection direction while keeping position storage cheap.
  • The section organization makes each local map an independent splatting unit, so merging scene parts via pose-graph registration is an obvious route to multi-session or multi-agent SLAM.
  • A monocular or estimated-depth variant could inherit the same storage savings, with the frozen-position assumption tested against depth-estimation error.
  • The reported tolerance to added depth noise (up to 30% of pixels) suggests the method's limit is bias rather than noise; injecting a fixed depth-scale error would pinpoint that limit.
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 VTGaussian-SLAM, an RGBD SLAM system that represents scenes with ``view-tied 3D Gaussians''. Each Gaussian is tied to a depth pixel, so its position is determined by the back-projected depth and camera pose and is never learned; only color, radius, and opacity are optimized. The method organizes frames into sections, keeps only the current section learnable, and uses overlapping sections for tracking head frames to reduce drift. The authors claim improved rendering quality, tracking accuracy, and scalability relative to prior 3D Gaussian splatting SLAM methods, and support this with experiments on Replica, TUM-RGBD, ScanNet, ScanNet++, and KITTI, plus ablations and memory/runtime comparisons.

Significance. The view-tied representation is a simple and memory-efficient idea: reducing per-Gaussian parameters from 14 to 5 is a concrete contribution, and the section-based scheduling that avoids optimizing all Gaussians globally is a plausible mechanism for scaling to longer sequences. The paper provides extensive benchmark tables, per-scene results in the supplement, and ablations of the main design choices. However, the headline claims are not supported by the evaluation protocol. Training-view rendering metrics are reported for most benchmarks, and those metrics are structurally inflated by the fact that Gaussians are initialized at the pixels of the evaluated frames. The only held-out novel view synthesis experiment shows parity, not superiority. Tracking claims are also contradicted by the paper's own tables when prior-based methods such as LoopSplat and CG-SLAM are included. The core idea has merit, but the evidence as presented overstates the method's advantages.

major comments (4)
  1. [Sec. 4, Tab. 2, Tab. 17] The rendering evaluation is not held out. The paper states that all rendering metrics are computed by rendering full-resolution images along the estimated trajectory every 5 frames, i.e., at training views. Because Sec. 3.2 initializes a Gaussian at every pixel with valid depth using that frame's color, and Sec. 3.4 adds complement Gaussians on each regular frame, the large PSNR/SSIM/LPIPS gaps in Tab. 2 largely measure per-frame overfitting rather than scene representation quality. The only held-out test, ScanNet++ novel view synthesis (Tab. 17), shows parity: Ours 21.46 dB vs LoopSplat 21.30 dB and Gaussian-SLAM 21.27 dB. The unqualified abstract claim of ``better performance'' in rendering is therefore not supported. The authors should either report novel view synthesis results on all four benchmarks or explicitly restrict the rendering claims to training-view reconstruction and discuss why the representation does not generalize better.
  2. [Sec. 3.2, Sec. 3.5, Supp. D.1] The method never optimizes Gaussian positions. Sec. 3.2 freezes positions at the back-projected depth, and the bundle adjustment in Sec. 3.5 only updates the head-frame camera pose. Systematic depth bias or accumulated pose drift therefore directly enters the final map with no corrective mechanism. The noise robustness study in Supp. D.1 (Tab. 22) only adds random noise to a fraction of pixels and measures rendering PSNR/SSIM/LPIPS, not geometric accuracy; the ``movable along ray'' variant is not a full position optimization. Reconstruction F1 on Replica (Tab. 3) is 90.0, slightly below LoopSplat's 90.4, which does not demonstrate a geometry advantage. The paper should analyze the effect of biased or drifting depth on the reconstructed mesh, not just on rendered images, and should acknowledge this limitation explicitly.
  3. [Abstract, Sec. 4.1, Tabs. 1, 4, 5] The abstract and conclusion claim better tracking accuracy over the latest methods, but the paper's own tables show the opposite when prior-based methods are included. On Replica, Ours ATE RMSE is 0.28 cm vs LoopSplat 0.26 cm and CG-SLAM 0.27 cm (Tab. 1); on TUM-RGBD, Ours is 2.6 cm vs LoopSplat 2.3 cm and CG-SLAM 2.0 cm (Tab. 4); on ScanNet, Ours is 11.3 cm vs LoopSplat 7.7 cm and CG-SLAM 8.1 cm (Tab. 5). The text in Sec. 4.1 argues that prior-based methods are ``not a fair experimental setting,'' but this caveat does not appear in the abstract or contribution list. The tracking claim should be restricted to prior-free methods, or the comparison with prior-based methods should be discussed without claiming overall superiority.
  4. [Supp. B, Tabs. 20-21] The KITTI experiments used to support the large-scale claim do not state how depth maps were obtained for the RGBD SLAM pipeline. KITTI provides LiDAR point clouds but not per-frame dense aligned depth maps. Without specifying the depth generation procedure (projection, interpolation, occlusion handling, alignment to the RGB image), the large-scale results are not reproducible and the comparison with RGBD SLAM baselines is not interpretable. Please add this detail, or remove the KITTI section if the depth source is not well-defined.
minor comments (5)
  1. [Tab. 17] The table title contains a typo: ``Synsthesis'' should be ``Synthesis.''
  2. [Supp. Tab. 18] The baseline SGS-SLAM appears in the Replica rendering table but is not listed in the main text's baseline list (Sec. 4) or described in the related work; please clarify its origin and configuration.
  3. [Sec. 3.3] The visibility mask W_i is used in Eq. (1), but its construction is only described later in Sec. 3.3 after the optimization objective is introduced; moving the definition before Eq. (1) would improve readability.
  4. [Sec. 2] The phrase ``multi-view variances'' in the discussion of 3D Gaussian attributes should be ``anisotropic scales'' or ``covariances'' to use standard terminology.
  5. [Sec. 4, Implementation Details] The section length N is dataset-dependent (N=40 on Replica, N=30/50 on ScanNet, N=100 on ScanNet++), but the sensitivity analysis in Tab. 8 is only on Replica. A brief justification for the real-dataset choices, or an ablation on one real dataset, would strengthen the claim that the method is not heavily tuned to N.

Circularity Check

1 steps flagged · score 6.0 of 10

Training-view rendering advantage is a fit of per-pixel view-tied Gaussians; the only held-out novel view test shows parity.

  1. fitted input called prediction [Sec. 4 (Datasets and Metrics), with Secs. 3.2, 3.4 and Eq. 2; held-out result in Sec. C / Tab. 17]
    "Sec. 4: "all the rendering metrics are computed by rendering the full resolution images along the estimated trajectory every 5 frames." Sec. 3.2: "We remove the need of learning and storing locations by tying a Gaussian g at each pixel with a valid depth value on the depth map. We center g at the 3D location back-projected from the depth value." Sec. 3.4: "we first initialize Gaussians at all pixels on the depth map at the head frame." Eq. 2 minimizes ρ||Vi − V′i||1 + τLS(Vi, V′i) + σUi||Di − D′i||1 against those observations."

    The headline rendering claim (Abstract: "improving rendering quality"; Sec. 4.1: "our method produces the best rendered images, as shown in Tab. 2") is supported by metrics computed at poses on the estimated trajectory, i.e., at the same frames used during mapping. The representation is defined by tying each Gaussian to a pixel of such a frame and back-projecting it from that frame's depth and pose; the mapping step then minimizes the RGB-D residual at exactly those frames. Thus Tab. 2's PSNR/SSIM/LPIPS values are a measure of how well the model fits its own per-pixel store, not of predicted novel views. This is confirmed by the only held-out evaluation: ScanNet++ novel view synthesis (Tab. 17) gives 21.46 dB vs LoopSplat 21.30 and Gaussian-SLAM 21.27, i.e., parity.

full rationale

This paper contains no derivation chain in the usual sense; its contribution is a directly defined representation plus tracking/mapping strategies, and most claims are checked against external baselines and ablations. There is no load-bearing self-citation chain: DF-Prior (Hu & Han 2023) appears only in related work and as a baseline, and no uniqueness theorem is imported from the authors' prior work. Tracking accuracy, reconstruction F1, memory, and runtime results are measured against independent systems and are not circular. The one load-bearing circularity is the rendering-quality claim: the flagship Tab. 2 metric is computed on training frames, while Secs. 3.2 and 3.4 initialize Gaussians at exactly those frames' valid-depth pixels and Eq. 2 optimizes the same frames' RGB-D errors. Consequently the reported rendering advantage is a fitted input called a prediction, and it collapses to parity on the paper's only held-out novel-view benchmark (ScanNet++, Tab. 17). This warrants a partial-circularity score of 6 rather than a higher score because tracking, scalability, memory usage, and reconstruction remain independently evaluated. The frozen-depth bias concern noted by the reader is a robustness limitation, not a circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central result rests on depth-as-geometry and local-consistency assumptions, not on new physical entities. View-tied Gaussians are a representation, not a hypothesized entity, so no invented entities are listed. The main cost is in tuned hyperparameters and strong assumptions about depth quality.

free parameters (6)
  • Section length N = 40 (Replica); 30 or 50 (ScanNet); 30 (TUM-RGBD); 100 (ScanNet++)
    Chosen per dataset based on image resolution and mapping iterations; Tab. 8 shows performance degrades at 20, 60, 80, and 100 on Replica, so the choice affects results.
  • Tracking weights alpha and beta = alpha=0.5; beta=0.025 (Replica), 1.0 (TUM-RGBD and ScanNet++), 0.9 (ScanNet)
    Set to balance RGB and depth tracking losses in Eq. 1; beta varies with depth quality.
  • Mapping weights rho, tau, sigma = rho=0.8, tau=0.2, sigma=1.0
    Balance L1 RGB, SSIM, and L1 depth in Eq. 2; fixed across datasets.
  • Overlap threshold gamma and coverage mask threshold = gamma=0.26 (TUM), 0.24 (ScanNet and ScanNet++); mask threshold 0.5
    Hand-chosen thresholds for selecting candidate views and deciding where to add complementary Gaussians.
  • Learning rates = color 0.0025, radius 0.005, opacity 0.05; pose lrrot and lrtrans differ per dataset
    Optimizer hyperparameters; pose learning rates are set separately for Replica, TUM-RGBD, ScanNet, and ScanNet++.
  • Candidate view interval N1 = 5
    Every 5th frame is a candidate view for overlapping-section selection during tracking.
assumptions (5)
  • domain assumption Differentiable Gaussian splatting provides an accurate image formation model for optimization.
    The tracking and mapping losses in Eq. 1 and Eq. 2 assume rendered RGB and depth match observations when pose and Gaussian attributes are correct.
  • domain assumption Depth maps are an unbiased geometric prior; Gaussian centers need no position optimization.
    Sec. 3.2 fixes Gaussian centers at back-projected depth pixels; the paper tests noise robustness in Tab. 22 but assumes no systematic bias.
  • domain assumption Local section-level optimization maintains global consistency; frozen sections remain valid.
    Sec. 3.1 and Sec. 3.4 optimize only the current section plus overlapping sections; there is no global consistency step or loop closure.
  • domain assumption Pose initialization by constant-speed assumption, or odometry on ScanNet++, is accurate enough for tracking convergence.
    Sec. 3.3 and Supplementary A rely on this for all datasets; on ScanNet++ an external RGBD odometry method is used when constant-speed initialization fails.
  • domain assumption Visibility masks computed by depth interpolation and silhouette projection capture occlusions.
    Eq. 1 uses mask W_i built from three-frame visibility and projection; incorrect masking would bias pose optimization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians." pith.science (2026). https://pith.science/paper/XKQ4O3HO

@misc{pith2026250602741,
  author       = {Pith},
  title        = {Pith review of: VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XKQ4O3HO}},
  note         = {Machine review of arXiv:2506.02741}
}
read the original abstract

Jointly estimating camera poses and mapping scenes from RGBD images is a fundamental task in simultaneous localization and mapping (SLAM). State-of-the-art methods employ 3D Gaussians to represent a scene, and render these Gaussians through splatting for higher efficiency and better rendering. However, these methods cannot scale up to extremely large scenes, due to the inefficient tracking and mapping strategies that need to optimize all 3D Gaussians in the limited GPU memories throughout the training to maintain the geometry and color consistency to previous RGBD observations. To resolve this issue, we propose novel tracking and mapping strategies to work with a novel 3D representation, dubbed view-tied 3D Gaussians, for RGBD SLAM systems. View-tied 3D Gaussians is a kind of simplified Gaussians, which is tied to depth pixels, without needing to learn locations, rotations, and multi-dimensional variances. Tying Gaussians to views not only significantly saves storage but also allows us to employ many more Gaussians to represent local details in the limited GPU memory. Moreover, our strategies remove the need of maintaining all Gaussians learnable throughout the training, while improving rendering quality, and tracking accuracy. We justify the effectiveness of these designs, and report better performance over the latest methods on the widely used benchmarks in terms of rendering and tracking accuracy and scalability. Please see our project page for code and videos at https://machineperceptionlab.github.io/VTGaussian-SLAM-Project .

Figures

Figures reproduced from arXiv: 2506.02741 by the authors.

Figure 1
Figure 1. Overview. (a) and (c) are tracking strategies, while (b) and (d) are mapping strategies. Please refer to Sec. 3.1 for more details. geometry of the scene, dedicated to reducing storage but pursuing better rendering quality. Different from the origi￾nal 3D Gaussians (Kerbl et al., 2023), we tie a 3D Gaussian to each pixel on the depth, which makes their positions only determined by depth and camera poses, without a n… view at source ↗
Figure 2
Figure 2. Illustration of selecting overlapping section. We show Gaussian centers and colors in each section [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Issue of pose error cumulation. This design aims to find a balance between the rendering quality and the spatial consistency of the current sec￾tion to previous frames in a long image sequence. Obviously, ren￾dering Gaussians in the same section will produce better renderings since neighboring frames usually have larger overlaps with the latest frame. Although better renderings are helpful for more accurate camera p… view at source ↗
Figures from the paper (14 more)
Figure 5
Figure 5. Figure 5: Illustration of optimizing view-tied Gaussians initialized on a head frame. Error maps are shown at different iterations. Implementation Details. For neighboring views in a sec￾tion Sk, we choose N = 40 on Replica (Straub et al., 2019), N = 30 on TUM-RGBD (Sturm et al.…
Figure 6
Figure 6. Figure 6: Visual comparisons in reconstruction on Replica [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Error map comparisons in rendering on Replica. ing quality, we measure PSNR, SSIM (Wang et al., 2004), and LPIPS (Zhang et al., 2018). Similar to (Sandstrom et al. ¨ , 2023a; Liso et al., 2024; Zhu et al., 2024; Yugay et al., 2023), all the rendering metrics are comput…
Figure 8
Figure 8. Figure 8: Visual comparisons in rendering on TUM-RGBD. Gaussians to represent local details in a more efficient man￾ner, leading to better renderings to compare with the obser￾vations during tracking. However, relying on data-driven priors, LoopSplat (Zhu et al., 2024) reported …
Figure 9
Figure 9. Figure 9: The rendering improvement also justifies our advan￾tages of using view-tied Gaussians on real-captured scenes. Besides good ability of recovering appearance details, with motion blur and low image quality in real images, our view￾tied Gaussians can also limit these neg…
Figure 10
Figure 10. Figure 10: Visualization of visibility and Gaussian projections on a head frame during tracking in a section [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Visualization of Gaussian centers with colors in the selected overlapping section for tracking (a) head frames and (b) regular frames. The Gaussian centers nearby are shown without color [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Visual comparisons in camera tracking on Replica. We also show error maps on reconstructions. Please refer to our video for a more complete comparison during scanning [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Visual comparisons in training view rendering on ScanNet++ (Yeshwanth et al., 2023). b20a261fdf 281bc17764 Zoom In SplaTAM Gaussian-SLAM LoopSplat Ours GT Zoom In [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Visual comparisons in novel view rendering on ScanNet++ (Yeshwanth et al., 2023). 20 [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Visual comparisons on fixed Gaussians or movable Gaussians (along the ray) in rendering on Replica. We provide rendering results, including RGB error maps and depth error maps rendered by using either our view-tied Gaussians or movable Gaussians along the ray. Additio…
Figure 18
Figure 18. Figure 18: Visualization of optimized 3D Gaussians [PITH_FULL_IMAGE:figures/full_fig_p024_18.png]
Figure 16
Figure 16. Figure 16: Comparison on w/ or w/o overlap selection when track￾ing [PITH_FULL_IMAGE:figures/full_fig_p024_16.png]
Figure 19
Figure 19. Figure 19: Visual comparisons of rendered images and depths. We also show error maps (large rendering errors are shown in red). Please refer to our video for more visual comparisons of rendered images. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 25 canonical work pages

  1. [3]

    Geiger, A., Lenz, P., and Urtasun, R

    URL https: //arxiv.org/abs/2402.04796. Geiger, A., Lenz, P., and Urtasun, R. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition,

    arXiv

  2. [5]

    2d gaussian splatting for geometrically accurate radi- ance fields

    Huang, B., Yu, Z., Chen, A., Geiger, A., and Gao, S. 2d gaussian splatting for geometrically accurate radi- ance fields. In SIGGRAPH 2024 Conference Papers . Association for Computing Machinery, 2024a. doi: 10.1145/3641519.3657428. Huang, H., Li, L., Hui, C., and Yeung, S.-K. Photo-slam: Real-time simultaneous localization and photorealistic mapping for m...

    arXiv 2024

  3. [6]

    Kong, X., Liu, S., Taher, M., and Davison, A

    URL https://repo-sam.inria.fr/ fungraph/3d-gaussian-splatting/. Kong, X., Liu, S., Taher, M., and Davison, A. J. vmap: Vectorised object mapping for neural field slam. arXiv preprint arXiv:2302.01838,

    arXiv

  4. [7]

    org/abs/2402.03246

    URL https://arxiv. org/abs/2402.03246. Li, Z., Lyu, X., Ding, Y ., Wang, M., Liao, Y ., and Liu, Y . Rico: Regularizing the unobservable for indoor composi- tional reconstruction,

    arXiv

  5. [8]

    org/abs/2405.06945

    URL https://arxiv. org/abs/2405.06945. Liso, L., Sandstr ¨om, E., Yugay, V ., Van Gool, L., and Os- wald, 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,

    arXiv

  6. [9]

    Instant neural graphics primitives with a multiresolution hash encoding

    M¨uller, T., Evans, A., Schied, C., and Keller, A. Instant neural graphics primitives with a multiresolution hash encoding. arXiv:2201.05989,

    arXiv

  7. [10]

    ∗ indicates methods relying on pre-trained data-driven priors

    Tracking performance comparisons in ATE RMSE ↓ [cm] on Replica (Straub et al., 2019). ∗ indicates methods relying on pre-trained data-driven priors. Method Rm0 Rm1 Rm2 Off0 Off1 Off2 Off3 Off4 Avg. Neural Implicit Fields NICE-SLAM (Zhu et al.,

    work page 2019

  8. [11]

    ∗ indicates methods relying on pre-trained data-driven priors

    Tracking performance comparisons in ATE RMSE ↓ [cm] on TUM-RGBD (Sturm et al., 2012). ∗ indicates methods relying on pre-trained data-driven priors. Method fr1/desk fr2/xyz fr3/office Avg. Neural Implicit Fields NICE-SLAM (Zhu et al.,

    work page 2012

Show all 42 references
  1. [12]

    org/abs/2406.04861

    URL https://arxiv. org/abs/2406.04861. R¨uckert, D., Franke, L., and Stamminger, M. Adop: Approximate differentiable one-pixel point rendering. arXiv:2110.06635,

    arXiv

  2. [13]

    R., and Tombari, F

    Sandstr¨om, E., Tateno, K., Oechsle, M., Niemeyer, M., Van Gool, L., Oswald, M. R., and Tombari, F. Splat- slam: Globally optimized rgb-only slam with 3d gaus- sians. arXiv preprint arXiv:2405.16544,

    arXiv

  3. [14]

    Straub, J., Whelan, T., Ma, L., Chen, Y ., Wijmans, E., Green, S., Engel, J

    URL https://arxiv.org/abs/2409.02104. Straub, J., Whelan, T., Ma, L., Chen, Y ., Wijmans, E., Green, S., Engel, J. J., Mur-Artal, R., Ren, C., Verma, S., Clark- son, A., Yan, M., Budge, B., Yan, Y ., Pan, X., Yon, J., Zou, Y ., Leon, K., Carter, N., Briales, J., Gillingham, T....

    1906 arXiv

  4. [15]

    ∗ indicates methods relying on pre-trained data-driven priors

    Tracking performance comparisons in ATE RMSE ↓ [cm] on ScanNet++ (Yeshwanth et al., 2023). ∗ indicates methods relying on pre-trained data-driven priors. Method a b c d e Avg. Neural Implicit Fields Point-SLAM (Sandstr¨om et al., 2023a) 246.16 632.99 830.79 271.42 574.86 511.2...

    2023

  5. [16]

    We additionally report our performance on extremely large scenes, such as city-level scenes in KITTI (Geiger et al., 2012)

    Large-scale scenes Results. We additionally report our performance on extremely large scenes, such as city-level scenes in KITTI (Geiger et al., 2012). Since many moving objects exist in KITTI sequences, we only select part of the sequences to evaluate our tracking and renderi...

    2012

  6. [17]

    ∗ indicates methods relying on pre-trained data-driven priors

    Novel View Synsthesis performance comparison in PSNR ↑ on ScanNet++ (Yeshwanth et al., 2023). ∗ indicates methods relying on pre-trained data-driven priors. Method a b c d e Avg. 3D Gaussian Splatting SplaTAM (Keetha et al.,

    2023

  7. [18]

    Wolf, Y ., Bracha, A., and Kimmel, R

    URL https://arxiv.org/abs/2408.12677. Wolf, Y ., Bracha, A., and Kimmel, R. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views. arXiv preprint arXiv:2404.01810,

    arXiv

  8. [19]

    V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation

    Yang, X., Li, H., Zhai, H., Ming, Y ., Liu, Y ., and Zhang, G. V ox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In 2022 IEEE International Symposium on Mixed and Aug- mented Reality (ISMAR) , Dec

    2022

  9. [20]

    URL http://dx.doi

    doi: 10.1109/ ISMAR55827.2022.00066. URL http://dx.doi. org/10.1109/ismar55827.2022.00066. Yeshwanth, C., Liu, Y .-C., Nießner, M., and Dai, A. Scan- net++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV),

    2022

  10. [21]

    MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction

    Yu, Z., Peng, S., Niemeyer, M., Sattler, T., and Geiger, A. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. ArXiv, abs/2022.00665,

    2022

  11. [22]

    Impact of depth noise and movability of Gaussians on the rendering performance in PSNR ↑, SSIM ↑, and LPIPS ↓ on Replica (Straub et al., 2019). Metric 10% pixels w/ noises 20% pixels w/ noises 30% pixels w/ noises Gaussians movable along ray Ours(w/o additional noises & fix) P...

    2019

  12. [23]

    Mononeuralfusion: Online monocular neural 3d recon- struction with geometric priors

    Zou, Z., Huang, S., Cao, Y ., Mu, T., Shan, Y ., and Fu, H. Mononeuralfusion: Online monocular neural 3d recon- struction with geometric priors. CoRR, abs/2209.15153,

    arXiv

  13. [24]

    Addi- tionally, we also show more results regarding Novel View Synthesis on ScanNet++ (Yeshwanth et al., 2023)

    13 Supplementary Material In this supplementary material, we will cover more details about the implementation and results on each scene. Addi- tionally, we also show more results regarding Novel View Synthesis on ScanNet++ (Yeshwanth et al., 2023). A. Implementation Details We...

    2023

  14. [25]

    to help the pose initialization if the rendering error with the pose initialized by constant speed assumption is 50 times larger than the average of the rendering loss for previous frames after the tracking optimization. We set the learning rate of pose to lrrot = 0.0004 and l...

    2019

  15. [26]

    Please refer to our supplementary video for more details of this comparison. We also report comparisons in rendering in each scene from the training views in Replica (Straub et al., 2019), TUM- RGBD (Sturm et al., 2012), ScanNet (Dai et al., 2017a), and ScanNet++ (Yeshwanth et al.,

    2019

  16. [28]

    We use a little bit more memory, but we manage to use more Gaussians to produce much better rendering

    Each method uses the most Gaussians until no improvement can be made. We use a little bit more memory, but we manage to use more Gaussians to produce much better rendering. C. Novel View Synthesis We evaluate the performance in novel view synthesis on ScanNet++ (Yeshwanth et a...

    2023

  17. [31]

    ∗ indicates methods relying on pre-trained data-driven priors

    Rendering performance comparison in PSNR ↑, SSIM ↑, and LPIPS ↓ on TUM-RGBD (Sturm et al., 2012). ∗ indicates methods relying on pre-trained data-driven priors. Method Metric fr1/desk fr2/xyz fr3/office Avg. Neural Implicit Fields NICE-SLAM (Zhu et al.,

    2012

  18. [33]

    ∗ indicates methods relying on pre-trained data-driven priors

    Rendering performance comparison in PSNR ↑ on ScanNet++ (Yeshwanth et al., 2023). ∗ indicates methods relying on pre-trained data-driven priors. Method a b c d e Avg. 3D Gaussian Splatting SplaTAM (Keetha et al.,

    2023

  19. [35]

    b20a261fdf281bc17764 Zoom In LoopSplat Ours GT Gaussian-SLAM SplaTAM Zoom In Figure

    Visual comparisons in training view rendering on ScanNet++ (Yeshwanth et al., 2023). b20a261fdf281bc17764 Zoom In LoopSplat Ours GT Gaussian-SLAM SplaTAM Zoom In Figure

    2023

  20. [36]

    20 Table

    Visual comparisons in novel view rendering on ScanNet++ (Yeshwanth et al., 2023). 20 Table

    2023

  21. [37]

    ∗ indicates methods relying on pre-trained data-driven priors

    Rendering performance comparisons in PSNR ↑, SSIM ↑, and LPIPS ↓ on Replica (Straub et al., 2019). ∗ indicates methods relying on pre-trained data-driven priors. Method Metric Rm0 Rm1 Rm2 Off0 Off1 Off2 Off3 Off4 Avg. Neural Implicit Fields NICE-SLAM (Zhu et al.,

    2019

  22. [38]

    ∗ indicates methods relying on pre-trained data-driven priors

    Reconstruction performance comparison in Depth L1 [cm]↓ and F1 [%] ↑ on Replica (Straub et al., 2019). ∗ indicates methods relying on pre-trained data-driven priors. Method Metric Rm0 Rm1 Rm2 Off0 Off1 Off2 Off3 Off4 Avg. Neural Implicit Fields NICE-SLAM (Zhu et al.,

    2019

  23. [39]

    Depth L1 [cm]↓ 1.31 0.82 1.26 0.81 0.96 1.41 1.53 1.08 1.16 F1 [%]↑ 62.9 79.9 66.8 80.0 81.6 66.0 59.2 65.0 70.2 Gaussian-SLAM (Yugay et al., 2023)Depth L1 [cm]↓ 0.61 0.25 0.54 0.50 0.52 0.98 1.63 0.42 0.68 F1 [%]↑ 88.8 91.4 90.5 91.7 90.1 87.3 84.2 87.4 88.9 LoopSplat∗ (Zhu et al.,

    2023

  24. [40]

    Sequence Gaussian-SLAM SplaTAM LoopSplat Ours 00 3.02 58.83 2.22 2.06 01 77.51 84.45 74.47 29.01 05 128.88 80.39 117.43 7.74 10 10.60 43.82 11.39 4.54 Table

    Tracking performance comparisons in ATE RMSE ↓ [m] on KITTI (Geiger et al., 2012). Sequence Gaussian-SLAM SplaTAM LoopSplat Ours 00 3.02 58.83 2.22 2.06 01 77.51 84.45 74.47 29.01 05 128.88 80.39 117.43 7.74 10 10.60 43.82 11.39 4.54 Table

    2012

  25. [41]

    Rendering performance comparisons in PSNR ↑ on KITTI (Geiger et al., 2012). Sequence Gaussian-SLAM SplaTAM LoopSplat Ours 00 15.51 9.82 15.82 28.54 01 15.95 12.89 14.69 30.33 05 16.22 26.48 15.98 28.19 10 15.58 25.58 14.58 27.59 Peak GPU Use (GiB) 2.74 22.37 3.56 4.79 Table

    2012

  26. [2004]

    doi: 10.1109/TIP.2003.819861. Wei, J. and Leutenegger, S. Gsfusion: Online rgb-d mapping where gaussian splatting meets tsdf fusion,

    2003

  27. [2012]

    Sucar, E., Liu, S., Ortiz, J., and Davison, A

    doi: 10.1109/IROS.2012.6385773. Sucar, E., Liu, S., Ortiz, J., and Davison, A. J. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6229–6238,

    2012

  28. [2017]

    Park, K., Sinha, U., Barron, J

    doi: 10.1109/ICCV .2017.25. Park, K., Sinha, U., Barron, J. T., Bouaziz, S., Goldman, D. B., Seitz, S. M., and Martin-Brualla, R. Nerfies: De- formable neural radiance fields. IEEE International Con- ference on Computer Vision,

    2017 doi

  29. [2018]

    Zhang, W., Liu, Y .-S., and Han, Z

    URLhttps://arxiv.org/ abs/1801.03924. Zhang, W., Liu, Y .-S., and Han, Z. Neural signed distance function inference through splatting 3d gaussians pulled on zero-level set. In Advances in Neural Information Processing Systems, 2024b. Zhang, W., Shi, K., Liu, Y .-S., and Han, Z...

    arXiv

  30. [2019]

    A benchmark for the evaluation of rgb-d slam 11 systems

    Sturm, J., Engelhard, N., Endres, F., Burgard, W., and Cre- mers, D. A benchmark for the evaluation of rgb-d slam 11 systems. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems , pp. 573–580,

    2012

  31. [2021]

    Neural graph mapping for dense slam with efficient loop closure

    Bruns, L., Zhang, J., and Jensfelt, P. Neural graph mapping for dense slam with efficient loop closure. arXiv preprint arXiv:2405.03633,

    arXiv

  32. [2022]

    Colored point cloud registration revisited

    Park, J., Zhou, Q.-Y ., and Koltun, V . Colored point cloud registration revisited. In 2017 IEEE International Con- ference on Computer Vision (ICCV), pp. 143–152,

    2017

  33. [2023]

    Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field

    Hu, J., Chen, X., Feng, B., Li, G., Yang, L., Bao, H., Zhang, G., and Cui, Z. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. arXiv preprint arXiv:2403.16095,

    arXiv

  34. [2024]

    Dai, A., Chang, A

    URL https://arxiv.org/ abs/2406.06521. Dai, A., Chang, A. X., Savva, M., Halber, M., Funkhouser, T. A., and Nießner, M. Scannet: Richly-annotated 3d reconstructions of indoor scenes. CoRR, abs/1702.04405, 2017a. Dai, A., Nießner, M., Zoll¨ofer, M., Izadi, S., and Theobalt, C. ...

    arXiv

Pith tools

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