Pith. sign in

REVIEW 3 major objections 6 minor 80 references

Dropping color from Gaussian Splatting improves SLAM geometry by 25-30%

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 · glm-5.2

2026-07-09 10:22 UTC pith:CYY5WXC7

load-bearing objection Solid systems paper with a clean idea; main gap is missing failure-mode analysis for the multi-view photometric loss on non-Lambertian surfaces. the 3 major comments →

arxiv 2607.07452 v1 pith:CYY5WXC7 submitted 2026-07-08 cs.RO

GeoGS-SLAM: Geometry-Only Gaussian Splatting for Dense Monocular SLAM

Lipu Zhou , Yaoyun Kang , Junxiang Pang , Shengkai Sun , Tingting Bao , Kehan Wang This is my paper
classification cs.RO
keywords Gaussian Splattingmonocular SLAMgeometry reconstruction3D Gaussian Splattingdense mappingloop closurephotometric consistencySim(3) transformation
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.

This paper argues that 3D Gaussian Splatting can perform dense monocular SLAM using only geometric parameters—position, rotation, scale, and opacity—without any appearance modeling (no spherical harmonics, no color). The authors call this representation GeoGS. The motivation is straightforward: SLAM downstream tasks like navigation and obstacle avoidance need accurate spatial geometry, not photorealistic rendering. By stripping color, each Gaussian primitive retains only 17% of the parameters of a standard 3DGS primitive, the scene requires far fewer primitives (23k vs 161k-231k in comparisons), and geometric convergence is faster because optimization effort is not split between appearance and geometry. The key technical challenge is that removing color rendering eliminates the standard photometric loss used to train Gaussians. The authors solve this with a two-part supervision scheme: single-view geometric losses (depth smoothness, normal consistency, depth distortion, and optional depth/normal priors from a SLAM frontend) and multi-view photometric losses that use rendered depth to reproject pixels between views and enforce cross-view RGB consistency without ever rendering color from the Gaussians themselves. They also introduce a local-plane PCA-based initialization that aligns new Gaussians with estimated surface normals from the start, and a coherent map update strategy for loop closure that applies a single unified Sim(3) transformation to all affected Gaussians rather than propagating per-keyframe corrections, which avoids surface tearing. On Replica, GeoGS-SLAM improves mean accuracy by 25% over the best prior method; on ScanNet++, it reduces Chamfer Distance by 30%.

Core claim

The central finding is that appearance modeling is not merely unnecessary for geometry-focused SLAM—it is actively counterproductive. When Gaussians carry color parameters, the optimizer can reduce photometric error by placing floating primitives that are geometrically wrong but visually helpful, and the larger parameter count per primitive slows convergence under the tight per-keyframe budget of online mapping. By removing all appearance parameters and supervising geometry through depth-based multi-view photometric consistency instead of color rendering, the system achieves better geometric reconstruction with fewer primitives, less storage, and faster convergence. The multi-view loss is负载:

What carries the argument

GeoGS primitive (geometry-only Gaussian with position, rotation, scale, opacity; no spherical harmonics), multi-view photometric supervision via depth-based reprojection (L1, SSIM, NCC on cross-view RGB patches without color rendering), local-plane PCA initialization, and coherent Sim(3) map update for loop closure

Load-bearing premise

The multi-view photometric losses assume that cross-view RGB patch consistency, computed via depth-based reprojection, provides a reliable gradient signal for optimizing Gaussian geometry without any appearance model. This depends on the rendered depth being accurate enough to produce correct correspondences—wrong depth yields wrong correspondences, which makes the photometric signal noise rather than supervision. The paper does not analyze failure modes of this feedback loop

What would settle it

If a scene with strong view-dependent surfaces (mirrors, glass, glossy materials) were added to the benchmark and GeoGS-SLAM's reconstruction quality degraded significantly more than appearance-coupled methods, the core premise that geometry-only supervision suffices would be challenged.

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

If this is right

  • Geometry-only Gaussian representations could become the default for robotics-focused SLAM where photorealistic rendering is not needed, separating the SLAM mapping problem from the novel-view-synthesis problem that motivated 3DGS.
  • The multi-view photometric supervision scheme—using rendered depth for reprojection rather than rendered color—could be adopted by other geometry-focused 3DGS methods that still carry appearance parameters, as a way to decouple geometric convergence from appearance fitting.
  • The coherent Sim(3) map update strategy addresses a general problem in explicit-map SLAM: post-optimization trajectory corrections can tear maps when applied per-keyframe. This unified-transformation approach could transfer to surfel-based or point-based SLAM systems beyond Gaussians.
  • The 80%+ reduction in per-primitive parameters and the 7-10x reduction in primitive count suggest that geometry-only Gaussians could enable dense SLAM on lower-power edge devices where memory and compute budgets are tight.

Where Pith is reading between the lines

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

  • The multi-view photometric loss creates a feedback loop: rendered depth drives cross-view correspondences, which provide gradients to improve depth. If depth is initially poor (e.g., in textureless regions or under aggressive motion), the correspondences may be wrong, and the photometric signal could reinforce errors rather than correct them. The paper's ablation shows this loss is the single most
  • Scenes with strong view-dependent appearance—specular surfaces, glass, transparent objects—could break the cross-view photometric consistency assumption that underpins the multi-view loss, since the same 3D point looks different from different viewpoints. The paper does not report results on such scenes, and the illumination-robustness claim is only qualitatively demonstrated, not quantitatively t
  • The coherent Sim(3) update assumes the trajectory correction can be well-approximated by a single global similarity transformation over the revisited region. For large loop closures with spatially varying corrections (e.g., bending along a long corridor), a single Sim(3) may underfit, and a piecewise or locally adaptive transformation might be needed.

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. This paper proposes GeoGS-SLAM, a dense monocular SLAM system that removes all appearance-related parameters (spherical harmonics) from 3D Gaussian Splatting, retaining only geometry-related parameters (position, rotation, scale, opacity). The authors introduce a training framework using single-view geometric losses (depth/normal supervision from frontend priors) and multi-view photometric losses (L1, SSIM, NCC on reprojected RGB patches), a local-plane driven initialization via PCA, and a coherent Sim(3)-based map update strategy for loop closure/global BA. The system is evaluated on Replica, ScanNet++, ScanNet, and DTU, showing improvements in reconstruction quality and tracking accuracy over recent baselines.

Significance. The paper addresses a well-motivated question: whether 3DGS can perform dense SLAM without appearance modeling. The GeoGS formulation is clean and the parameter reduction (>80% per primitive, 23k vs 161k-231k Gaussians on DTU) is a concrete, measurable efficiency gain. The local-plane initialization and coherent map update are reasonable engineering contributions. The DTU time-constrained evaluation (Table IV) provides a controlled comparison showing strong geometric convergence. The ablation isolating L_mv as the most critical loss term (Table X) is informative. Code release is promised.

major comments (3)
  1. §VI.D, Table II: The ScanNet++ reconstruction comparison is only against HI-SLAM2. This is the sole real-world dataset with geometric ground truth, and the 30% CD improvement claim rests on a single baseline. Other recent GS-based SLAM methods (e.g., Splat-SLAM) are compared on Replica and ScanNet tracking but omitted here. Adding at least one more baseline on ScanNet++ would substantially strengthen the central SOTA claim on real-world data.
  2. §I and Fig. 10: The paper claims 'enhanced robustness to illumination variations' as a key benefit of removing appearance modeling. However, the multi-view loss L_mv (Eq. 27-29) still operates on raw RGB values, so view-dependent radiance differences (specular surfaces, glass) create photometric residuals that the optimizer attributes entirely to geometry error, since there is no appearance model to absorb them. Fig. 10 provides only qualitative evidence under moderate lighting changes. The claim should either be tested quantitatively on scenes with strong view-dependent appearance, or the scope should be explicitly qualified to Lambertian-dominant scenes. As stated, the claim is stronger than what the evidence supports.
  3. §VI.G, Table X: The loss-term ablation is performed on DTU Scan24 'without optional geometric priors,' which is not representative of the SLAM pipeline where L_d and L_n (Eq. 20-21) are active. While the stress-test note correctly observes that single-view losses stabilize the system in the SLAM setting, the ablation does not isolate L_mv's contribution within the full pipeline. An ablation on Replica or ScanNet++ (where geometric priors are active) would better characterize the relative importance of each loss term in the actual SLAM configuration.
minor comments (6)
  1. Table VI: The storage entry for GeoGS is listed as '1.2' without a unit. Based on context (other entries in MB), this should be '1.2 MB'.
  2. Fig. 2 caption: 'corss-view' should be 'cross-view'.
  3. §IV.A, Eq. (7): s_i is stated as ∈ R^2, consistent with the 2DGS disk parameterization, but this should be noted explicitly as a departure from standard 3DGS (where s ∈ R^3) to avoid confusion for readers familiar with the original formulation.
  4. §V.D, Eq. (35): The notation T'^{-1}_i is used but not explicitly defined as the inverse of the corrected pose. A brief clarification would help readability.
  5. §VI.G, Table VIII: The coherent map update ablation is performed on Replica without loop closure (only final global BA). Since the coherent update is primarily motivated by loop closure scenarios, an ablation with loop closure active would better demonstrate its intended benefit.
  6. The paper would benefit from reporting runtime/latency numbers for the mapping and loop closure update steps, given that the motivation emphasizes online efficiency.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. The referee correctly identifies the core contributions of GeoGS-SLAM and raises three substantive points: (1) the ScanNet++ reconstruction comparison is limited to a single baseline, (2) the robustness-to-illumination claim is stronger than the evidence supports, and (3) the loss-term ablation is conducted in a configuration that does not represent the full SLAM pipeline. We address each point below and commit to revisions where the referee's critique is valid.

read point-by-point responses

  1. Referee: §VI.D, Table II: The ScanNet++ reconstruction comparison is only against HI-SLAM2. This is the sole real-world dataset with geometric ground truth, and the 30% CD improvement claim rests on a single baseline. Other recent GS-based SLAM methods (e.g., Splat-SLAM) are compared on Replica and ScanNet tracking but omitted here. Adding at least one more baseline on ScanNet++ would substantially strengthen the central SOTA claim on real-world data.

    Authors: The referee is correct that the ScanNet++ reconstruction comparison in Table II includes only HI-SLAM2 as a baseline, and that adding at least one more GS-based SLAM method would strengthen the claim. The reason for the current omission is practical: Splat-SLAM does not provide ScanNet++ reconstruction results in its published evaluation, and reproducing Splat-SLAM's full pipeline on ScanNet++ with geometric ground-truth evaluation requires non-trivial setup (mesh extraction and metric computation against the ScanNet++ ground-truth meshes using the Splat-SLAM evaluation toolbox). However, we agree this is not a sufficient justification for omission given the centrality of the claim. We will add Splat-SLAM reconstruction results on the four ScanNet++ sequences in the revised manuscript. If reproducibility issues arise with Splat-SLAM's mesh extraction pipeline on ScanNet++, we will document them transparently and report whatever results we can obtain. revision: yes

  2. Referee: §I and Fig. 10: The paper claims 'enhanced robustness to illumination variations' as a key benefit of removing appearance modeling. However, the multi-view loss L_mv (Eq. 27-29) still operates on raw RGB values, so view-dependent radiance differences (specular surfaces, glass) create photometric residuals that the optimizer attributes entirely to geometry error, since there is no appearance model to absorb them. Fig. 10 provides only qualitative evidence under moderate lighting changes. The claim should either be tested quantitatively on scenes with strong view-dependent appearance, or the scope should be explicitly qualified to Lambertian-dominant scenes.

    Authors: The referee raises a valid and important point. We agree that the claim as currently stated is stronger than what the evidence supports. The referee's technical observation is correct: because L_mv operates on raw RGB values, view-dependent radiance differences (specular highlights, reflective surfaces) create photometric residuals that, in the absence of an appearance model, the optimizer may attribute to geometry error. This means GeoGS is not inherently robust to all illumination variations—specifically, it is vulnerable to strong view-dependent appearance effects. What our design does provide is robustness to global illumination changes across views (e.g., exposure differences, ambient lighting shifts) in the sense that the optimizer does not spend capacity fitting view-dependent color, and the single-view geometric losses (L_d, L_n) provide supervision that is independent of appearance. However, this is a narrower claim than 'enhanced robustness to illumination variations.' We will revise the manuscript to explicitly qualify the scope: (1) in the abstract and introduction, we will clarify that the robustness benefit applies primarily to Lambertian-dominant scenes and global illumination changes, not to strong view-dependent effects; (2) we will add a discussion of the limitation that L_mv's use of raw RGB means specular and reflective surfaces can introduce geometry errors; (3) we will reframe Fig. 10's caption to accurately describe the moderate nature of the lighting changes shown. We will not claim quantitative robustness to strong view-dependent appearance, as we do not have experiments supporting such a claim. revision: yes

  3. Referee: §VI.G, Table X: The loss-term ablation is performed on DTU Scan24 'without optional geometric priors,' which is not representative of the SLAM pipeline where L_d and L_n (Eq. 20-21) are active. While the stress-test note correctly observes that single-view losses stabilize the system in the SLAM setting, the ablation does not isolate L_mv's contribution within the full pipeline. An ablation on Replica or ScanNet++ (where geometric priors are active) would better characterize the relative importance of each loss term in the actual SLAM configuration.

    Authors: The referee is correct that the ablation in Table X is conducted in a stripped-down configuration (DTU Scan24 without geometric priors) that does not represent the full SLAM pipeline where L_d and L_n are active. The original motivation for this design was to stress-test the loss terms in a setting where geometric priors do not dominate, thereby exposing the relative contributions of each term more clearly. However, we agree that this choice limits the interpretability of the ablation for the SLAM setting. The stress-test observation—that removing single-view losses destabilizes the system in the full SLAM configuration—already hints at the interaction, but it does not provide a clean isolation of L_mv's contribution when L_d and L_n are active. We will add an ablation on the Replica dataset (where geometric priors from the frontend are active) that removes each loss term from the full pipeline and reports reconstruction metrics. This will complement the existing DTU stress-test ablation and provide the characterization the referee requests. We will retain the DTU ablation as well, with a clearer explanation of its purpose as a stress-test configuration, while the new Replica ablation will serve as the representative SLAM-pipeline ablation. revision: yes

Circularity Check

0 steps flagged

No circularity found: GeoGS-SLAM's derivation is self-contained, with externally falsifiable benchmarks and no self-citation chains.

full rationale

The paper's central claims are not circular. (1) GeoGS (Eq. 7) is a straightforward restriction of standard 3DGS (Eq. 1-2) by removing SH parameters — a design choice validated against external baselines (2DGS, PGSR, QGS) on DTU (Table IV), not a self-citation. (2) The loss functions (Eqs. 16-31) are adapted from PGSR [9] and 2DGS [10] with explicit attribution; the multi-view photometric loss (Eqs. 23-30) uses rendered depth for reprojection, which is a standard MVS-style formulation, not a definition that reduces to its input. (3) The coherent map update (Eqs. 35-37) is a least-squares Sim(3) fit — a standard Procrustes-type alignment, not a self-definitional loop. (4) The frontend (DROID-SLAM [29]) and depth/normal priors (Omnidata [73]) are external, independently developed components. (5) No 'prediction' reduces to a fitted parameter: the 80% parameter reduction is arithmetic (14 params vs 59+ for SH-3DGS), the primitive count reduction (23k vs 161k-231k) is measured empirically on DTU Scan55 (Table VI), and the reconstruction improvements on Replica (Table I) and ScanNet++ (Table II) are externally benchmarked against HI-SLAM2, Splat-SLAM, and others. (6) The ablation (Table X) confirms L_mv is load-bearing but does not make the result circular — it shows the loss contributes, which is the expected behavior of a genuine optimization signal. The paper is self-contained against external benchmarks with no self-citation chain supporting its central premise.

Axiom & Free-Parameter Ledger

13 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or dimensions. The GeoGS primitive is a parameter-reduced variant of the standard 3DGS primitive, not a new entity. The free parameters are all standard loss weights and thresholds that would be tuned during development; their values are not reported, which limits reproducibility. The most consequential axioms are the multi-view photometric consistency assumption (which the ablation confirms is load-bearing) and the single-Sim(3) approximation for loop closure regions.

free parameters (13)
  • lambda_nc = not stated
    Weight for normal consistency loss (Eq. 22), tuned on training data
  • lambda_ds = not stated
    Weight for depth smoothness loss (Eq. 22)
  • lambda_dist = not stated
    Weight for distortion loss (Eq. 22)
  • lambda_d = not stated
    Weight for depth L1 loss (Eq. 22)
  • lambda_n = not stated
    Weight for normal L1 loss (Eq. 22)
  • lambda_1 = not stated
    Weight for multi-view L1 loss (Eq. 30)
  • lambda_ssim = not stated
    Weight for multi-view SSIM loss (Eq. 30)
  • lambda_ncc = not stated
    Weight for multi-view NCC loss (Eq. 30)
  • eta = not stated
    Weighting coefficient in edge-aligned depth smoothness (Eq. 18)
  • tau = not stated
    Occlusion threshold for multi-view validation mask (Eq. 25)
  • planarity_threshold = not stated
    Threshold below which Gaussian scale is reduced by 10x (Sec. IV-D)
  • scale_reduction_factor = 10
    Ad hoc factor for reducing scale in low-planarity regions (Sec. IV-D)
  • pose_lr = ~6e-4
    Learning rate for pose-only refinement; ablated in Table IX, optimal around 4e-4 to 6e-4
axioms (5)
  • domain assumption Scene geometry is more critical for SLAM than novel view synthesis
    Stated in Abstract and Sec. I; motivates the entire approach. Reasonable for robotic navigation but not universally true for all SLAM applications.
  • ad hoc to paper Multi-view photometric consistency (RGB patch agreement across views) provides sufficient signal for geometry optimization without appearance modeling
    Implicit in the design of L_mv (Eq. 27-30). The ablation (Table X) shows this is the most critical loss term, but no analysis of when it fails (e.g., specular surfaces, textureless regions) is provided.
  • domain assumption DROID-SLAM frontend provides sufficiently accurate poses and dense depth for Gaussian backend initialization
    Sec. V-A states the frontend is built on DROID-SLAM. The quality of GeoGS reconstruction depends on frontend depth quality, but this dependency is not analyzed.
  • ad hoc to paper A single Sim(3) transformation adequately approximates the correction needed for all revisited Gaussians in a loop closure region
    Sec. V-D, Eq. 36. This assumes the drift within a revisited region is well-approximated by a rigid similarity transform, which may not hold for large regions with non-uniform drift accumulation.
  • standard math Differentiable rasterization of 2D Gaussian disks (following 2DGS) produces accurate depth and normal renders
    Sec. IV-B follows the 2DGS formulation [10] for ray-splat intersection. This is a standard result from prior work.

pith-pipeline@v1.1.0-glm · 25886 in / 3649 out tokens · 405920 ms · 2026-07-09T10:22:52.050402+00:00 · methodology

0 comments
read the original abstract

Dense visual SLAM is a fundamental problem in robotics. Recent advances in 3DGS have demonstrated its potential for dense SLAM. Existing 3DGS frameworks focus on both appearance and geometry modeling. However, scene geometry is typically more critical for SLAM than novel view synthesis because downstream robotic tasks, such as navigation and obstacle avoidance, rely primarily on accurate spatial geometry rather than photorealistic rendering. This observation raises a natural question: Is it feasible for 3DGS to perform 3D reconstruction without scene appearance modeling? Motivated by this, we propose Geometry-only Gaussian Splatting (GeoGS), which directly reconstructs scene geometry, and further present GeoGS-SLAM, a dense visual SLAM system built upon this representation. Specifically, GeoGS retains only spatial parameters to reduce the number of per-primitive parameters by over 80%. In contrast to existing 3DGS methods, GeoGS focuses solely on geometric reconstruction, which significantly reduces the number of Gaussian primitives, accelerates geometric convergence, and enhances robustness to illumination variations. In addition, we present an effective training framework that optimizes the Gaussian primitives via single-view and multi-view geometric and photometric supervision, and speeds up geometry convergence with a local-plane driven initialization that better aligns primitives with local structures. Furthermore, we introduce a map update strategy for loop closure that globally transforms the Gaussian map to align it with the corrected pose estimates, thereby preventing map tearing caused by inconsistent per-viewpoint pose corrections in existing methods. Extensive experiments on synthetic and real-world benchmarks demonstrate that our method outperforms SOTA methods in terms of online mapping efficiency and geometric reconstruction quality.

Figures

Figures reproduced from arXiv: 2607.07452 by Junxiang Pang, Kehan Wang, Lipu Zhou, Shengkai Sun, Tingting Bao, Yaoyun Kang.

Figure 1
Figure 1. Figure 1: Illustration of the characteristics of GeoGS and the performance of GeoGS-SLAM. (a) Qualitative comparison of geometric reconstruction. GeoGS-SLAM recovers much cleaner scene geometry compared to the baseline method. (b) Per-primitive parameter efficiency. By retaining only the geometr-related parameters (µ,q,s,α), GeoGS remains only 17% of the parameters required by a standard 3DGS primitive. (c) Model ef… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GeoGS mapping framework. GeoGS primitives are initialized using a local-plane driven strategy and optimized through single-view [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between a standard 3DGS primitive and the proposed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-view supervision in GeoGS mapping. The rendered depth ˆ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Local-plane driven initialization for GeoGS primitives. For each [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the proposed GeoGS-SLAM system. Given a monocular RGB sequence, the frontend estimates camera poses and dense depth maps with [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of scene reconstruction on the Replica dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of geometry reconstruction quality on the [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of scene reconstruction on the Scannet dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Robustness of GeoGS-SLAM to illumination changes. (a) and (b) show two representative scene regions under different lighting conditions, with [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Geometry convergence curves on the DTU Scan55 sequence [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 11
Figure 11. Figure 11: Geometry convergence performance on the DTU Scan55 sequence. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Geometry convergence performance on the Replica Office0 sequence. [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative ablation of local-plane driven initialization. PCA-aligned [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Ablation study of the proposed local-plane driven initialization on the [PITH_FULL_IMAGE:figures/full_fig_p014_15.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

80 extracted references · 80 canonical work pages · 9 internal anchors

  1. [1]

    Slam handbook: From localization and mapping to spatial intelligence,

    L. Carlone, A. Kim, T. Barfoot, D. Cremers, and F. Dellaert, “Slam handbook: From localization and mapping to spatial intelligence,” 2025

    work page 2025

  2. [2]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakiset al., “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    work page 2023

  3. [3]

    What is the best 3d scene representation for robotics? from geometric to foundation models,

    T. Deng, Y . Pan, S. Yuan, D. Li, C. Wang, M. Li, L. Chen, L. Xie, D. Wang, J. Wanget al., “What is the best 3d scene representation for robotics? from geometric to foundation models,”arXiv preprint arXiv:2512.03422, 2025

    work page arXiv 2025

  4. [4]

    Mip-splatting: Alias-free 3d gaussian splatting,

    Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip-splatting: Alias-free 3d gaussian splatting,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 447–19 456

    work page 2024

  5. [5]

    Mini-splatting: Representing scenes with a constrained number of gaussians,

    G. Fang and B. Wang, “Mini-splatting: Representing scenes with a constrained number of gaussians,” inEuropean conference on computer vision. Springer, 2024, pp. 165–181

    work page 2024

  6. [6]

    Taming 3dgs: High-quality radiance fields with limited resources,

    S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V . Carrasco, and F. De La Torre, “Taming 3dgs: High-quality radiance fields with limited resources,” inSIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1– 11

    work page 2024

  7. [7]

    Gaussianpro: 3d gaussian splatting with progressive propa- gation,

    K. Cheng, X. Long, K. Yang, Y . Yao, W. Yin, Y . Ma, W. Wang, and X. Chen, “Gaussianpro: 3d gaussian splatting with progressive propa- gation,” inForty-first International Conference on Machine Learning, 2024

    work page 2024

  8. [8]

    Dn-splatter: Depth and normal priors for gaussian splatting and meshing,

    M. Turkulainen, X. Ren, I. Melekhov, O. Seiskari, E. Rahtu, and J. Kannala, “Dn-splatter: Depth and normal priors for gaussian splatting and meshing,” in2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2025, pp. 2421–2431

    work page 2025

  9. [9]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 9, pp. 6100–6111, 2024

    work page 2024

  10. [10]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inACM SIGGRAPH 2024 conference papers, 2024, pp. 1–11

    work page 2024

  11. [11]

    Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives,

    Z. Zhang, B. Huang, H. Jiang, L. Zhou, X. Xiang, and S. Shen, “Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 28 260–28 270

    work page 2025

  12. [12]

    Evolving high-quality rendering and reconstruction in a unified framework with contribution-adaptive regularization,

    Y . Shen, Z. Zhang, X. Li, Y . Qu, Y . Lin, S. Zhang, and L. Cao, “Evolving high-quality rendering and reconstruction in a unified framework with contribution-adaptive regularization,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16 346–16 355

    work page 2025

  13. [13]

    Scaffold- gs: Structured 3d gaussians for view-adaptive rendering,

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai, “Scaffold- gs: Structured 3d gaussians for view-adaptive rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 654–20 664

    work page 2024

  14. [14]

    Stopthepop: Sorted gaussian splatting for view-consistent real- time rendering,

    L. Radl, M. Steiner, M. Parger, A. Weinrauch, B. Kerbl, and M. Stein- berger, “Stopthepop: Sorted gaussian splatting for view-consistent real- time rendering,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–17, 2024

    work page 2024

  15. [15]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,

    A. Gu ´edon and V . Lepetit, “Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 5354–5363

    work page 2024

  16. [16]

    Gaussian opacity fields: Efficient adap- tive surface reconstruction in unbounded scenes,

    Z. Yu, T. Sattler, and A. Geiger, “Gaussian opacity fields: Efficient adap- tive surface reconstruction in unbounded scenes,”ACM Transactions on Graphics (ToG), vol. 43, no. 6, pp. 1–13, 2024

    work page 2024

  17. [17]

    Rade-gs: Rasterizing depth in gaussian splatting,

    B. Zhang, C. Fang, R. Shrestha, Y . Liang, X.-X. Long, and P. Tan, “Rade-gs: Rasterizing depth in gaussian splatting,”ACM Transactions on Graphics, vol. 45, no. 2, pp. 1–14, 2026

    work page 2026

  18. [18]

    Gsdf: 3dgs meets sdf for improved neural rendering and reconstruction,

    M. Yu, T. Lu, L. Xu, L. Jiang, Y . Xiangli, and B. Dai, “Gsdf: 3dgs meets sdf for improved neural rendering and reconstruction,”Advances in Neural Information Processing Systems, vol. 37, pp. 129 507–129 530, 2024

    work page 2024

  19. [19]

    A Mixed-Primitive-based Gaussian Splatting Method for Surface Reconstruction

    H. Qu, Y . Cai, H. Rahmani, A. Kumar, J. Yuan, and J. Liu, “A mixed- primitive-based gaussian splatting method for surface reconstruction,” arXiv preprint arXiv:2507.11321, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    Geometry-grounded gaussian splatting,

    B. Zhang, C. Jiang, H. Li, S. Shen, and P. Tan, “Geometry-grounded gaussian splatting,”arXiv preprint arXiv:2601.17835, 2026

    work page arXiv 2026

  21. [21]

    Dtam: Dense tracking and mapping in real-time,

    R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “Dtam: Dense tracking and mapping in real-time,” in2011 international conference on computer vision. IEEE, 2011, pp. 2320–2327

    work page 2011

  22. [22]

    Kinectfusion: Real-time dense surface mapping and tracking,

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, “Kinectfusion: Real-time dense surface mapping and tracking,” in2011 10th IEEE international symposium on mixed and augmented reality. Ieee, 2011, pp. 127–136

    work page 2011

  23. [23]

    Elasticfusion: Real-time dense slam and light source estimation,

    T. Whelan, R. F. Salas-Moreno, B. Glocker, A. J. Davison, and S. Leutenegger, “Elasticfusion: Real-time dense slam and light source estimation,”The International Journal of Robotics Research, vol. 35, no. 14, pp. 1697–1716, 2016

    work page 2016

  24. [24]

    Bundle- fusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration,

    A. Dai, M. Nießner, M. Zollh ¨ofer, S. Izadi, and C. Theobalt, “Bundle- fusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration,”ACM Transactions on Graphics (ToG), vol. 36, no. 4, p. 1, 2017

    work page 2017

  25. [25]

    Lsd-slam: Large-scale direct monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale direct monocular slam,” inEuropean conference on computer vision. Springer, 2014, pp. 834–849

    work page 2014

  26. [26]

    Cnn-slam: Real-time dense monocular slam with learned depth prediction,

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

    work page 2017

  27. [27]

    Codeslam—learning a compact, optimisable representation for dense visual slam,

    M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison, “Codeslam—learning a compact, optimisable representation for dense visual slam,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2560–2568

    work page 2018

  28. [28]

    Raft: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” inEuropean conference on computer vision. Springer, 2020, pp. 402–419

    work page 2020

  29. [29]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    ——, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,”Advances in neural information processing systems, vol. 34, pp. 16 558–16 569, 2021

    work page 2021

  30. [30]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    work page 2021

  31. [31]

    imap: Implicit mapping and positioning in real-time,

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

    work page 2021

  32. [32]

    Nice-slam: Neural implicit scalable encoding for slam,

    Z. Zhu, S. Peng, V . Larsson, W. Xu, H. Bao, Z. Cui, M. R. Oswald, and M. Pollefeys, “Nice-slam: Neural implicit scalable encoding for slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 786–12 796

    work page 2022

  33. [33]

    Nicer-slam: Neural implicit scene encoding for rgb slam,

    Z. Zhu, S. Peng, V . Larsson, Z. Cui, M. R. Oswald, A. Geiger, and M. Pollefeys, “Nicer-slam: Neural implicit scene encoding for rgb slam,” in2024 International Conference on 3D Vision (3DV). IEEE, 2024, pp. 42–52

    work page 2024

  34. [34]

    Eslam: Efficient dense slam system based on hybrid representation of signed distance fields,

    M. M. Johari, C. Carta, and F. Fleuret, “Eslam: Efficient dense slam system based on hybrid representation of signed distance fields,” in 16 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17 408–17 419

    work page 2023

  35. [35]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM transactions on graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022

    work page 2022

  36. [36]

    Go-slam: Global optimization for consistent 3d instant reconstruction,

    Y . Zhang, F. Tosi, S. Mattoccia, and M. Poggi, “Go-slam: Global optimization for consistent 3d instant reconstruction,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3727–3737

    work page 2023

  37. [37]

    Hi-slam: Monocular real-time dense mapping with hybrid implicit fields,

    W. Zhang, T. Sun, S. Wang, Q. Cheng, and N. Haala, “Hi-slam: Monocular real-time dense mapping with hybrid implicit fields,”IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1548–1555, 2023

    work page 2023

  38. [38]

    GlORIE-SLAM: Globally Optimized RGB-only Implicit Encoding Point Cloud SLAM

    G. Zhang, E. Sandstr ¨om, Y . Zhang, M. Patel, L. Van Gool, and M. R. Oswald, “Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam,”arXiv preprint arXiv:2403.19549, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Grounding image matching in 3d with mast3r,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with mast3r,” inEuropean conference on computer vision. Springer, 2024, pp. 71–91

    work page 2024

  40. [40]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294– 5306

    work page 2025

  41. [41]

    Depth Anything 3: Recovering the Visual Space from Any Views

    H. Lin, S. Chen, J. Liew, D. Y . Chen, Z. Li, G. Shi, J. Feng, and B. Kang, “Depth anything 3: Recovering the visual space from any views,”arXiv preprint arXiv:2511.10647, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  42. [42]

    Vggt-slam: Dense rgb slam optimized on the sl (4) manifold,

    D. Maggio, H. Lim, and L. Carlone, “Vggt-slam: Dense rgb slam optimized on the sl (4) manifold,”Advances in Neural Information Processing Systems, vol. 38, pp. 129 839–129 867, 2026

    work page 2026

  43. [43]

    VGGT-Long: Chunk it, Loop it, Align it -- Pushing VGGT's Limits on Kilometer-scale Long RGB Sequences

    K. Deng, Z. Ti, J. Xu, J. Yang, and J. Xie, “Vggt-long: Chunk it, loop it, align it–pushing vggt’s limits on kilometer-scale long rgb sequences,” arXiv preprint arXiv:2507.16443, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Vggt-motion: Motion- aware calibration-free monocular slam for long-range consistency,

    Z. Xiong, C. Zhang, Q. Xu, and W. Tao, “Vggt-motion: Motion- aware calibration-free monocular slam for long-range consistency,” arXiv preprint arXiv:2602.05508, 2026

    work page arXiv 2026

  45. [45]

    Hyvggt-vo: Tightly coupled hy- brid dense visual odometry with feed-forward models,

    J. Pang, L. Zhou, and B. Chen, “Hyvggt-vo: Tightly coupled hy- brid dense visual odometry with feed-forward models,”arXiv preprint arXiv:2604.02107, 2026

    work page arXiv 2026

  46. [46]

    Gs- slam: Dense visual slam with 3d gaussian splatting,

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, and X. Li, “Gs- slam: Dense visual slam with 3d gaussian splatting,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 595–19 604

    work page 2024

  47. [47]

    Splatam: Splat track & map 3d gaussians for dense rgb-d slam,

    N. Keetha, J. Karhade, K. M. Jatavallabhula, G. Yang, S. Scherer, D. Ramanan, and J. Luiten, “Splatam: Splat track & map 3d gaussians for dense rgb-d slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 357–21 366

    work page 2024

  48. [48]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 18 039–18 048

    work page 2024

  49. [49]

    Towards next-generation slam: A survey on 3dgs-slam focusing on performance, robustness, and future directions,

    L. Wang, R. Gong, Y . Han, L. Yang, L. Yang, Y . Li, B. Xu, H. Liu, and R. Fu, “Towards next-generation slam: A survey on 3dgs-slam focusing on performance, robustness, and future directions,”arXiv preprint arXiv:2602.04251, 2026

    work page arXiv 2026

  50. [50]

    Gaussian-SLAM: Photo-realistic Dense SLAM with Gaussian Splatting

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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  51. [51]

    Photo-slam: Real- time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,

    H. Huang, L. Li, H. Cheng, and S.-K. Yeung, “Photo-slam: Real- time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 584–21 593

    work page 2024

  52. [52]

    Rtg- slam: Real-time 3d reconstruction at scale using gaussian splatting,

    Z. Peng, T. Shao, Y . Liu, J. Zhou, Y . Yang, J. Wang, and K. Zhou, “Rtg- slam: Real-time 3d reconstruction at scale using gaussian splatting,” in ACM SIGGRAPH 2024 conference papers, 2024, pp. 1–11

    work page 2024

  53. [53]

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

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

    work page 2024

  54. [54]

    Fgs- slam: Fourier-based gaussian splatting for real-time slam with sparse and dense map fusion,

    Y . Xu, J. Li, W. Zhang, S. Chen, S. Zhang, Y . Leng, and W. Zhou, “Fgs- slam: Fourier-based gaussian splatting for real-time slam with sparse and dense map fusion,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 5355–5362

    work page 2025

  55. [55]

    Mg- slam: Structure gaussian splatting slam with manhattan world hypothe- sis,

    S. Liu, T. Deng, H. Zhou, L. Li, H. Wang, D. Wang, and M. Li, “Mg- slam: Structure gaussian splatting slam with manhattan world hypothe- sis,”IEEE Transactions on Automation Science and Engineering, 2025

    work page 2025

  56. [56]

    SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding,

    Y . F. Tianci Wen, Zhiang Liu, “SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding,” inProceedings of the IEEE/CVF Conference on Computer Vision, 2025

    work page 2025

  57. [57]

    Mgs-slam: Monocular sparse tracking and gaussian mapping with depth smooth regularization,

    P. Zhu, Y . Zhuang, B. Chen, L. Li, C. Wu, and Z. Liu, “Mgs-slam: Monocular sparse tracking and gaussian mapping with depth smooth regularization,”IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9486–9493, 2024

    work page 2024

  58. [58]

    Splat-slam: Globally optimized rgb-only slam with 3d gaussians,

    E. Sandstr ¨om, G. Zhang, K. Tateno, M. Oechsle, M. Niemeyer, Y . Zhang, M. Patel, L. Van Gool, M. Oswald, and F. Tombari, “Splat-slam: Globally optimized rgb-only slam with 3d gaussians,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1680–1691

    work page 2025

  59. [59]

    Droid-splat combining end- to-end slam with 3d gaussian splatting,

    C. Homeyer, L. Begiristain, and C. Schn ¨orr, “Droid-splat combining end- to-end slam with 3d gaussian splatting,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 2767–2777

    work page 2025

  60. [60]

    Hi-slam2: Geometry-aware gaussian slam for fast monocular scene reconstruction,

    W. Zhang, Q. Cheng, D. Skuddis, N. Zeller, D. Cremers, and N. Haala, “Hi-slam2: Geometry-aware gaussian slam for fast monocular scene reconstruction,”IEEE Transactions on Robotics, vol. 41, pp. 6478–6493, 2025

    work page 2025

  61. [61]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 3, pp. 611–625, 2017

    work page 2017

  62. [62]

    Mgso: Monocular real-time photometric slam with efficient 3d gaussian splatting,

    Y . S. Hu, N. Abboud, M. Q. Ali, A. S. Yang, I. Elhajj, D. Asmar, Y . Chen, and J. S. Zelek, “Mgso: Monocular real-time photometric slam with efficient 3d gaussian splatting,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 11 061–11 067

    work page 2025

  63. [63]

    Gso-slam: Bidirectionally coupled gaussian splatting and direct visual odometry,

    J. Yeon, S. Ha, and H. Yu, “Gso-slam: Bidirectionally coupled gaussian splatting and direct visual odometry,”IEEE Robotics and Automation Letters, 2026

    work page 2026

  64. [64]

    Gaussianflow slam: Monocular gaussian splatting slam guided by gaussianflow,

    D.-U. Seo, J. Jeon, E. M. Lee, and H. Myung, “Gaussianflow slam: Monocular gaussian splatting slam guided by gaussianflow,”IEEE Robotics and Automation Letters, 2026

    work page 2026

  65. [65]

    Outdoor monocular slam with global scale-consistent 3d gaussian pointmaps,

    C. Cheng, S. Yu, Z. Wang, Y . Zhou, and H. Wang, “Outdoor monocular slam with global scale-consistent 3d gaussian pointmaps,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 26 035–26 044

    work page 2025

  66. [66]

    Artdeco: Towards efficient and high-fidelity on-the-fly 3d reconstruction with structured scene representation,

    G. Li, K. Ren, L. Xu, Z. Zheng, C. Jiang, X. Gao, B. Dai, J. Pu, M. Yu, and J. Pang, “Artdeco: Towards efficient and high-fidelity on-the-fly 3d reconstruction with structured scene representation,”arXiv preprint arXiv:2510.08551, 2025

    work page arXiv 2025

  67. [67]

    Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAM

    Z. Zhang, K. Wu, X. Meng, K. Liu, J. Zhao, and W. Ding, “Flash- mono: Feed-forward accelerated gaussian splatting monocular slam,” arXiv preprint arXiv:2604.03092, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  68. [68]

    Vings-mono: Visual-inertial gaussian splatting monocular slam in large scenes,

    K. Wu, Z. Zhang, M. Tie, Z. Ai, Z. Gan, and W. Ding, “Vings-mono: Visual-inertial gaussian splatting monocular slam in large scenes,”IEEE Transactions on Robotics, 2025

    work page 2025

  69. [69]

    GLC-SLAM: Gaussian Splatting SLAM with Efficient Loop Closure

    Z. Xu, Q. Li, C. Chen, X. Liu, and J. Niu, “Glc-slam: Gaussian splatting slam with efficient loop closure,”arXiv preprint arXiv:2409.10982, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  70. [70]

    Vigs-slam: visual inertial gaussian splatting slam,

    Z. Zhu, W. Zhang, M. Li, N. Haala, M. Pollefeys, and D. Barath, “Vigs-slam: visual inertial gaussian splatting slam,”arXiv preprint arXiv:2512.02293, 2025

    work page arXiv 2025

  71. [71]

    Globally consistent rgb-d slam with 2d gaussian splatting,

    X. Zhong, Y . Pan, L. Jin, M. Popovi ´c, J. Behley, and C. Stachniss, “Globally consistent rgb-d slam with 2d gaussian splatting,”IEEE Transactions on Robotics, 2026

    work page 2026

  72. [72]

    Ewa splatting,

    M. Zwicker, H. Pfister, J. Van Baar, and M. Gross, “Ewa splatting,”IEEE Transactions on Visualization & Computer Graphics, vol. 8, no. 03, pp. 223–238, 2002

    work page 2002

  73. [73]

    Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,

    A. Eftekhar, A. Sax, J. Malik, and A. Zamir, “Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 786–10 796

    work page 2021

  74. [74]

    Truncated signed distance function: experiments on voxel size,

    D. Werner, A. Al-Hamadi, and P. Werner, “Truncated signed distance function: experiments on voxel size,” inInternational conference image analysis and recognition. Springer, 2014, pp. 357–364

    work page 2014

  75. [75]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  76. [76]

    The Replica Dataset: A Digital Replica of Indoor Spaces

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

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  77. [77]

    Scannet++: A high- fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai, “Scannet++: A high- fidelity dataset of 3d indoor scenes,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12–22

    work page 2023

  78. [78]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5828–5839. 17

    work page 2017

  79. [79]

    Large scale multi-view stereopsis evaluation,

    R. Jensen, A. Dahl, G. V ogiatzis, E. Tola, and H. Aanæs, “Large scale multi-view stereopsis evaluation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 406–413

    work page 2014

  80. [80]

    evo: Python package for the evaluation of odometry and slam,

    M. Grupp, “evo: Python package for the evaluation of odometry and slam,” 2017

    work page 2017