Pith. sign in

REVIEW 4 major objections 5 minor 23 references

This paper claims that a 3D Gaussian Splatting SLAM system can overcome the GPU memory bottleneck by storing inactive spatial chunks on disk, enabling multi-kilometer photorealistic mapping on a 24 GB GPU and making it the only one of four

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

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

T0 review · deepseek-v4-flash

2026-08-03 19:38 UTC pith:UDUYULNJ

load-bearing objection Useful memory-management mechanism, but the 'uniquely completes all 11 KITTI' headlining comparison rests on author-configured baselines and needs audit before the strong claims stick. the 4 major comments →

arxiv 2511.23030 v2 pith:UDUYULNJ submitted 2025-11-28 cs.RO cs.CV

DiskChunGS: Large-Scale 3D Gaussian SLAM Through Chunk-Based Memory Management

Casimir Feldmann , Maximum Wilder-Smith , Vaishakh Patil , Michael Oechsle , Michael Niemeyer , Keisuke Tateno , Marco Hutter This is my paper
classification cs.RO cs.CV
keywords 3D Gaussian SplattingSLAMout-of-core mappingchunk-based memory managementlarge-scale reconstructionloop closureGPU memoryreal-time mapping
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 tries to establish that 3D Gaussian Splatting SLAM can scale to multi-kilometer environments on a standard GPU by treating the map as an out-of-core database: the scene is split into spatial chunks that are swapped between disk and VRAM as the camera moves. If the mechanism holds, the VRAM ceiling that currently confines 3DGS SLAM to rooms and small labs is removed, and photorealistic mapping scales with disk rather than GPU memory. The paper demonstrates this on indoor and outdoor datasets and claims to be the only system in its experiments to finish all 11 sequences of a standard urban driving benchmark without memory failures or tracking loss, while achieving better perceptual quality than the baselines that do complete the route.

Core claim

The paper's central claim is that the GPU-memory bottleneck that limits 3D Gaussian Splatting SLAM to small scenes can be eliminated algorithmically: partition the map into fixed-size cubic chunks, keep only the chunks visible from the current keyframe in VRAM, evict least-recently-used chunks to disk, and restrict keyframe selection to a coarse spatial grid so that consecutive optimization steps touch overlapping chunk sets. With a 1.5-million-Gaussian budget and a 400-keyframe budget, the system reports flat VRAM usage while the map grows, and it is the only method in its comparison to process all 11 sequences of an outdoor driving benchmark without memory crashes or tracking failure, with

What carries the argument

The chunk: a 10-meter cube of Gaussians with a 64-bit packed ID for O(1) lookups. Hierarchical frustum culling selects visible chunks per keyframe; an LRU policy writes inactive chunks to disk and preloads them on demand; a 200-meter keyframe-selection grid keeps chunk overlap high and I/O low; depth-supervised direct Gaussian placement (based on Laplacian-of-Gaussian image content) replaces iterative densification.

Load-bearing premise

The headline claim that the method is the only one to finish all 11 benchmark sequences depends on baseline competitors being run in configurations their own authors would endorse, and the paper itself notes in its implementation details that it used default hyperparameters for one method, fine-tuned another, and omitted an offline refinement stage, so those baseline results may understate the competitors; a secondary premise is that the tested intake rates count as online de

What would settle it

Check the mechanism by measuring VRAM usage on a synthetic route with repeated revisits and a 20-km extent: if memory rises linearly instead of plateauing at the Gaussian budget, the LRU eviction claim fails. Check the comparative claim by re-running each baseline with its own preferred hyperparameters, available VRAM, and any published offline refinement; if one then completes all 11 sequences with comparable or better perceptual quality, the uniqueness claim collapses.

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

If this is right

  • Multi-kilometer outdoor mapping becomes feasible on a 24 GB GPU with a flat memory footprint that does not grow with map size.
  • Incremental persistence: individual chunks can be serialized to disk and reloaded, enabling persistent maps and targeted updates.
  • Loop closure can be applied by transforming only visible Gaussians in affected chunks, avoiding global re-optimization.
  • The architecture runs on resource-constrained edge hardware with modest quality loss, suggesting mobile robotic deployment.
  • I/O, not VRAM, becomes the next scaling bottleneck: disk traffic costs 5–20% of processing time depending on the speed-accuracy trade-off.

Where Pith is reading between the lines

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

  • If the flat-VRAM mechanism is real, the same chunk-eviction design could be applied to other dense SLAM representations (voxel grids, hash maps, meshes), generalizing the out-of-core principle beyond Gaussian splatting.
  • The 'uniquely completes all 11 sequences' claim is only as strong as the fairness of the baseline runs; the paper itself notes in its implementation details that it used default hyperparameters for one method, fine-tuned another, and excluded a published offline refinement stage, so a reader should treat the superiority claim as provisional until baselines are run under their own preferred setting
  • A sharp testable prediction follows: VRAM usage should stay flat on arbitrarily long trajectories, so extending to a 10 km+ route would confirm or break the mechanism.
  • The reported processing speeds (1–6 frames per second) are below the 10 Hz capture rate of the driving benchmark, so the 'online' claim should be read as 'operates during capture with back-pressure' rather than true real-time; the adaptive speed/quality trade-off is the paper's answer to this.

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

4 major / 5 minor

Summary. The paper introduces DiskChunGS, a 3D Gaussian Splatting SLAM system that addresses GPU memory limits by partitioning the scene into spatial chunks, keeping only active chunks in VRAM, and evicting inactive chunks to disk via an LRU policy. The system uses ORB-SLAM3 for tracking, keyframe-driven optimization with a spatial grid for keyframe selection, direct Gaussian sampling from image content, and chunk-aware loop closure. Experiments on Replica, TUM-RGBD, KITTI, and a Jetson Orin report that the method completes all 11 KITTI sequences without memory failures, achieves superior visual quality versus CaRtGS, On-The-Fly, and GigaSLAM, and maintains a flat VRAM footprint as scene size grows. The central claims are: (1) a mechanism claim that chunked out-of-core management bounds VRAM regardless of map size, and (2) a comparative claim that the method is the only one to finish all KITTI sequences with better perceptual quality.

Significance. The chunk-based memory management mechanism is a credible and practically relevant contribution. The VRAM plateau in Fig. 5, the ablation in Table V (6 of 11 sequences fail without chunking, 2 fail without grid keyframe selection), and the successful completion of all KITTI sequences provide direct evidence that the mechanism works. The additional validation on Jetson hardware and integration with ROS strengthen the practical value. However, the paper's headline comparisons are less secure: the baseline runs are configured asymmetrically by the authors, the stereo-vs-mono comparison is not controlled, and the reported processing FPS does not reach the KITTI capture rate, undermining the 'online' claim. If the comparative claims are hardened, this would be a solid systems contribution; as presented, the mechanism is well supported but the superlative claims overreach.

major comments (4)
  1. [Sec. IV, Implementation Details; Tables I-III] The central comparative claim—'uniquely completes all 11 KITTI sequences' and 'superior visual quality'—rests entirely on author-run baseline evaluations with asymmetric configurations. CaRtGS is 'fine-tuned' by the authors (although originally intended for indoor scenes), On-The-Fly is run with default hyperparameters, and GigaSLAM is evaluated in 'online-only' mode, explicitly excluding the offline stage that the paper states can achieve higher quality. No configuration files, commit hashes, or code are provided. This makes the 'uniquely' and 'superior' claims unverifiable and potentially dependent on evaluation-protocol choices rather than algorithmic merit. The authors should provide exact baseline configurations and either run baselines under their intended protocols or justify any deviations with sensitivity analyses.
  2. [Sec. IV, Table II; Abstract and Contributions] The abstract and contributions describe 'efficient online processing,' but the processing FPS reported in Table II is 0.59–1.66 FPS for Ours-5 and 2.26–6.59 FPS for Ours-20, both below the KITTI capture rate of 10 FPS. The system therefore does not keep up with the input stream in the tested configurations. The term 'online' needs to be qualified: either demonstrate a configuration that meets or exceeds the sensor frame rate, or revise the claim to 'offline processing with incremental mapping.' This is load-bearing for the practical deployment claim.
  3. [Sec. IV, Results; Tables I-III] The evaluation compares a stereo-input system (DiskChunGS, CaRtGS) against mono-input baselines (On-The-Fly, GigaSLAM). The paper states that modality 'is secondary to the fundamental scalability challenges,' but provides no control or evidence for this assertion. Since the method uses stereo depth estimation (Fast-ACVNet) for dense supervision (Sec. III), the visual-quality advantage could be partly attributable to the richer input rather than the chunking mechanism. A mono variant of DiskChunGS or an ablation with identical depth supervision across modalities is needed to support the comparative quality claim.
  4. [Sec. IV, Implementation Details; Reproducibility] For an empirical systems paper whose headline is a comparative superlative, the absence of released code, configuration files, and even commit hashes for the baseline runs is a serious gap. The text says the baselines are 'rerun' using publicly available code, but does not specify versions or command-line settings. This prevents independent verification of the central claim. At minimum, the authors should include full configuration details, and ideally release the artifact.
minor comments (5)
  1. [References] Reference [15] contains a typo: 'hierachical' should be 'hierarchical.'
  2. [Eq. (5)] The floor notation in Eq. (5) is ambiguous for a vector argument. Specify that the floor is applied element-wise to each coordinate.
  3. [Sec. III, Losses and Implementation Details] The loss weights lambda_s and lambda_depth are introduced in Eqs. (2)-(3) but their values are not given. Provide these hyperparameters for reproducibility.
  4. [Fig. 5] The legend label 'Keyframes / Gaussians' is unclear. Use separate legend entries or axis labels for active Gaussians and active keyframes, and clarify that the y-axis is VRAM allocation.
  5. [Sec. IV, Dataset descriptions] The sentence 'KITTI is a comprehensive multi-kilometer outdoor dataset collected from a moving vehicle' is fine, but the later phrase 'datasets with lower capture rates (e.g., KITTI at 10 FPS) naturally yield lower processing FPS values' is slightly confusing because processing FPS is wall-clock based; clarify that lower capture rate means fewer frames per second to process, not a lower speed requirement.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with external baselines and ablations; central claims do not reduce to their inputs.

full rationale

This is an empirical systems paper rather than a derivation-driven one. Equations (1)-(8) define sampling probability, chunk indexing, overlap, and keyframe selection; they are operational definitions, not predictions obtained by fitting parameters to target outputs. The flat VRAM curve in Fig. 5 is the intended consequence of the explicitly configured Gaussian/keyframe budgets, and the paper does not present it as an emergent prediction; the load-bearing validations are the completion of all 11 KITTI sequences, the component ablations in Table V, and comparisons on public datasets against external baselines. The paper contains no self-citations: references [5], [14]-[16], [18], and [19] are all external works. Concerns about baseline configuration—such as GigaSLAM's excluded offline refinement stage, On-The-Fly's default hyperparameters, CaRtGS's fine-tuning, and mono-vs-stereo modality differences—are evaluation-fairness and correctness risks, not circular reductions. No claim in the paper is equivalent by construction to its inputs, and no fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper's central mechanism is an engineering configuration rather than a derivation, so its real burden is six hand-set parameters (chunk size, VRAM budget, keyframe budget, grid resolution, loss weights, refinement iterations), none with sensitivity analysis, plus four domain assumptions about camera locality, depth quality, ORB-SLAM3 robustness, and modality comparability. No invented entities. The mechanism claim (bounded VRAM, completion of all 11 KITTI sequences) is externally benchmarked against standard datasets and ablations, which is the paper's strongest independent grounding.

free parameters (6)
  • chunk size s = 10 m
    Set to 10 meters for all datasets with no sensitivity analysis; controls the trade-off between I/O granularity and renderable set size (§IV Implementation Details).
  • Gaussian VRAM budget = 1.5M Gaussians
    Configurable budget that defines when LRU eviction kicks in; hardware-dependent, and no study of how its value affects quality or completion (§III Optimization).
  • keyframe memory budget = 400 keyframes
    LRU queue cap for keyframes; hand-set, no sensitivity analysis (§III).
  • keyframe grid resolution g = 200 m
    Spatial grid for keyframe candidate selection, coarse relative to the 10 m chunk size; no ablation of the value itself (§III Keyframe Selection).
  • loss weights lambda_s, lambda_depth = not reported
    Weights in Eqs. 2-3 for the SSIM and depth terms; values never stated, so the tuning is unexposed.
  • loop-closure refinement iterations = 1k iterations (~10 s)
    Fixed refinement budget for junction regions after loop closure (§III Loop Closure).
axioms (4)
  • domain assumption The camera path has spatial locality; LRU chunk eviction and visibility-based loading will not thrash.
    The whole design (Eqs. 5-8 and the LRU policy) assumes a coherent trajectory. A teleporting or rapidly switching camera across distant chunks would cycle chunks per keyframe; the paper only validates on driving/handheld trajectories.
  • domain assumption Fused depth from Depth-Anything-2/Fast-ACVNet, aligned to triangulated ORB-SLAM3 matches, is accurate enough to place and supervise Gaussians.
    §III Sampling adopts On-The-Fly's direct sampling; Gaussian placement quality inherits the accuracy of predicted+aligned depth, which is not independently validated here.
  • domain assumption ORB-SLAM3 completes all 11 sequences without tracking loss.
    The 'no tracking loss' claim is inherited from ORB-SLAM3; the paper's own Table I shows ATE up to 16.41 m (seq 01), so 'success' means completion, not accuracy (Tables I-II).
  • domain assumption Mono-vs-stereo baseline comparison is acceptable because modality 'is secondary to the fundamental scalability challenges.'
    §IV Results asserts modality is secondary; this premise underpins the 'uniquely completes all 11' framing against mono baselines and is not tested.

pith-pipeline@v1.3.0-alltime-deepseek · 11347 in / 26190 out tokens · 244709 ms · 2026-08-03T19:38:26.197364+00:00 · methodology

0 comments
read the original abstract

Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated impressive results for novel view synthesis with real-time rendering capabilities. However, integrating 3DGS with SLAM systems faces a fundamental scalability limitation: methods are constrained by GPU memory capacity, restricting reconstruction to small-scale environments. We present DiskChunGS, a scalable 3DGS SLAM system that overcomes this bottleneck through an out-of-core approach that partitions scenes into spatial chunks and maintains only active regions in GPU memory while storing inactive areas on disk. Our architecture integrates seamlessly with existing SLAM frameworks for pose estimation and loop closure, enabling globally consistent reconstruction at scale. We validate DiskChunGS on indoor scenes (Replica, TUM-RGBD), urban driving scenarios (KITTI), and resource-constrained Nvidia Jetson platforms. Our method uniquely completes all 11 KITTI sequences without memory failures while achieving superior visual quality, demonstrating that algorithmic innovation can overcome the memory constraints that have limited previous 3DGS SLAM methods.

Figures

Figures reproduced from arXiv: 2511.23030 by Casimir Feldmann, Keisuke Tateno, Marco Hutter, Maximum Wilder-Smith, Michael Niemeyer, Michael Oechsle, Vaishakh Patil.

Figure 1
Figure 1. Figure 1: Pareto curves for KITTI [6] scenes. With more iterations and, as a consequence, more processing time, 3DGS SLAM methods can optimize for longer, achieving higher reconstruction quality. Our method achieves superior visual quality in less time than competing methods across all three scenes. II. RELATED WORK Neural Large-Scale Reconstructions Recent neural approaches have pursued large-scale scene representa… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of DiskChunGS. For each SLAM keyframe, we estimate depth and perform direct primitive placement based on image content analysis instead of iterative densification. For optimization of a keyframe, frustum culling is performed to identify visible chunks, which are loaded from disk into VRAM. On the other hand, old chunks are evicted from VRAM to disk to free up memory. The visible subset of Gaussian… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results on the KITTI [6] dataset. Reconstruction results on three scenes by all methods. On-The-Fly suffers from tracking drift and lacks loop closure. GigaSLAM’s neural approach fails without expensive post-processing. CaRtGS shows floating artifacts from missing depth supervision. Our method achieves superior quality through robust tracking, depth-supervised Gaussian placement, and efficient … view at source ↗
Figure 4
Figure 4. Figure 4: Zoomed qualitative results on the Replica [7] dataset (office0). TABLE IV COMPARISON OF METHODS ON THE TUM [8] AND REPLICA [7] DATASETS. ATE IS IN CM. BEST AND SECOND BEST RESULTS ARE HIGHLIGHTED. Metric CaRtGS On-The-Fly GigaSLAM Ours TUM ATE ↓ 0.85 7.12 22.02 1.88 PSNR ↑ 20.59 23.28 13.40 21.21 SSIM ↑ 0.71 0.84 0.53 0.75 LPIPS ↓ 0.22 0.14 0.68 0.22 FPS ↑ 24.14 51.88 10.99 25.09 Replica ATE ↓ 0.30 31.68 1… view at source ↗
Figure 5
Figure 5. Figure 5: Active Gaussians and Keyframes vs VRAM usage. Our Gaussian and Keyframe disk-based saving and loading system keeps memory usage steady as scene size increases. V. LIMITATIONS & FUTURE WORK Our DiskChunGS method provides an algorithmic solution for scaling 3DGS SLAM through efficient disk-based Gaussian storage. Like other 3DGS methods, it does not explicitly handle capture artifacts such as motion blur, le… 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

23 extracted references · 1 canonical work pages

  1. [1]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, July 2023. [Online]. Available: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    2023

  2. [2]

    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,” 2020. [Online]. Available: https://arxiv.org/abs/2003.08934

    Pith/arXiv arXiv 2020

  3. [3]

    Plenoxels: Radiance fields without neural networks,

    A. Yu, S. Fridovich-Keil, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Radiance fields without neural networks,”

  4. [4]

    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, vol. 41, no. 4, p. 1–15, July 2022. [Online]. Available: http://dx.doi.org/10.1145/3528223.3530127

    arXiv 2022

  5. [5]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodriguez, J. M. M. Montiel, and J. D. Tardos, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE Transactions on Robotics, vol. 37, no. 6, p. 1874–1890, Dec. 2021. [Online]. Available: http://dx.doi.org/10.1109/TRO.2021.3075644

    arXiv 2021

  6. [6]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” inConference on Computer Vision and Pattern Recognition (CVPR), 2012

    2012

  7. [7]

    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. Verma, A. Clarkson, M. Yan, B. Budge, Y . Yan, X. Pan, J. Yon, Y . Zou, K. Leon, N. Carter, J. Briales, T. Gillingham, E. Mueggler, L. Pesqueira, M. Savva, D. Batra, H. M. Strasdat, R. D. Nardi, M. Goesele, S. Lovegrove, and R. Newcombe, “The replica dataset...

    Pith/arXiv arXiv 2019

  8. [8]

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

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” inProc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012

    2012

  9. [9]

    Block-nerf: Scalable large scene neural view synthesis,

    M. Tancik, V . Casser, X. Yan, S. Pradhan, B. Mildenhall, P. P. Srinivasan, J. T. Barron, and H. Kretzschmar, “Block-nerf: Scalable large scene neural view synthesis,” 2022. [Online]. Available: https://arxiv.org/abs/2202.05263

    Pith/arXiv arXiv 2022

  10. [10]

    Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,

    H. Turki, D. Ramanan, and M. Satyanarayanan, “Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2022, pp. 12 922–12 931

    2022

  11. [11]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,

    Y . Liu, H. Guan, C. Luo, L. Fan, N. Wang, J. Peng, and Z. Zhang, “Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,” 2024. [Online]. Available: https://arxiv.org/abs/2404.01133

    Pith/arXiv arXiv 2024

  12. [12]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2311.11700

    Pith/arXiv arXiv 2024

  13. [13]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2311.16728

    Pith/arXiv arXiv 2024

  14. [14]

    Cartgs: Computational alignment for real-time gaussian splatting slam,

    D. Feng, Z. Chen, Y . Yin, S. Zhong, Y . Qi, and H. Chen, “Cartgs: Computational alignment for real-time gaussian splatting slam,”IEEE Robotics and Automation Letters, vol. 10, no. 5, p. 4340–4347, May

  15. [15]

    Gigaslam: Large-scale monocular slam with hierachical gaussian splats,

    K. Deng, J. Yang, S. Wang, and J. Xie, “Gigaslam: Large-scale monocular slam with hierachical gaussian splats,” 2025. [Online]. Available: https://arxiv.org/abs/2503.08071

    Pith/arXiv arXiv 2025

  16. [16]

    On-the-fly reconstruction for large-scale novel view synthesis from unposed images,

    A. Meuleman, I. Shah, A. Lanvin, B. Kerbl, and G. Drettakis, “On-the-fly reconstruction for large-scale novel view synthesis from unposed images,”ACM Transactions on Graphics, vol. 44, no. 4, p. 1–14, July 2025. [Online]. Available: http://dx.doi.org/10.1145/3730913

    work page doi:10.1145/3730913 2025

  17. [17]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2312.10070

    Pith/arXiv arXiv 2024

  18. [18]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” 2024. [Online]. Available: https://arxiv.org/abs/2406.09414

    Pith/arXiv arXiv 2024

  19. [19]

    Accurate and efficient stereo matching via attention concatenation volume,

    G. Xu, Y . Wang, J. Cheng, J. Tang, and X. Yang, “Accurate and efficient stereo matching via attention concatenation volume,” 2023. [Online]. Available: https://arxiv.org/abs/2209.12699

    Pith/arXiv arXiv 2023

  20. [20]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,”

  21. [2018]

    Available: https://arxiv.org/abs/1801.03924

    [Online]. Available: https://arxiv.org/abs/1801.03924

    Pith/arXiv arXiv

  22. [2021]

    Available: https://arxiv.org/abs/2112.05131

    [Online]. Available: https://arxiv.org/abs/2112.05131

    Pith/arXiv arXiv

  23. [2025]

    Available: http://dx.doi.org/10.1109/LRA.2025.3544928

    [Online]. Available: http://dx.doi.org/10.1109/LRA.2025.3544928

    arXiv 2025