Pith. sign in

REVIEW 3 major objections 4 minor 43 references

Visual Relocalization from Sparse Views in Aliased and Low-Texture Environments via Novel View Synthesis

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that geometric supervision of the Gaussian Splatting map — not photometric fidelity — is the decisive factor for single-image relocalization in low-texture, aliased rover terrain.

desk verdict A sensible combination of geometric losses for 3DGS relocalization, but the headline ablation is confounded across pipeline, iterations, and coordinate frame; the idea deserves a properly controlled follow-up. read the letter →

arxiv 2607.22147 v1 pith:3CM7WSX2 submitted 2026-07-24 cs.CV

Maria Peribañez , Javier Civera , Rudolph Triebel , Riccardo Giubilato This is my paper · ORCID
classification cs.CV
keywords visualrelocalization3DGaussianSplattingmulti-viewstereoLiDARsupervisionlow-textureenvironmentsplanetaryroverposeestimationnovelviewsynthesis
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 tries to establish that, for visual relocalization in planetary-like environments, the geometric accuracy of the 3D scene representation matters more than its photometric fidelity. It proposes training a 3D Gaussian Splatting map with additional depth and surface-normal losses from multi-view stereo and a Chamfer loss against LiDAR point clouds, alongside the standard photometric loss. The authors argue this combination produces a map on which a fixed single-image pose estimator succeeds where a photometric-only map fails. If true, the practical consequence is that relocalization in aliased, low-texture terrain can be engineered through map-level geometric supervision rather than better feature matching.

What carries the argument

The load-bearing mechanism is a structure-preserving 3D Gaussian Splatting objective that stacks three supervisory signals on the standard photometric loss: an L1 depth alignment and cosine normal loss against zero-shot multi-view stereo estimates, and a symmetric Chamfer distance anchoring rendered depth maps to aggregated LiDAR point clouds. Because the map is differentiable, all these losses backpropagate into the Gaussian parameters. The paper's argument is that this combination resolves depth ambiguity in low-parallax, forward-motion observations and pins the map to metric scale, which in turn gives the fixed downstream pose estimator (a feature-based ray–pixel registration against the

What would settle it

Retrain the photometric-only map under exactly the same training procedure, iteration count, normalization, and metric-frame settings as the geometry-aware variants (removing only the geometric losses), and measure pose recall at the relaxed threshold. If recall rises substantially above 6.25%, the central attribution to geometric supervision is weakened. A second check: train a variant with MVS supervision but no LiDAR and another with LiDAR but no MVS, all settings otherwise identical, and confirm the claimed complementarity holds.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that adding geometric losses to 3D Gaussian Splatting fitting — specifically MVS-derived depth and normal alignment plus a LiDAR-guided Chamfer distance — yields a map with 74% lower reconstruction error and lifts single-image relocalization pose recall from 6.25% to 43.20% at a relaxed (10 m, 15°) threshold, while a strict (2 m, 10°) threshold improves from 0% to 6.80%. The pose estimator is held fixed; the only intentional change is the supervision used to build the map. The paper interprets this as evidence that metric consistency of the representation, not photometric quality, is the load-bearing factor for 6-DoF pose recovery.

Load-bearing premise

The load-bearing premise is that the only difference between the photometric-only baseline and the geometry-aware variants is the geometric supervision; in fact the baseline uses a different training pipeline, 30,000 iterations, and enabled pose normalization and scene centering, while the variants use a different pipeline, 20,000 iterations, and metric-frame optimization — so the reported recall gain may not isolate the geometric losses.

Editorial extensions

If this is right

  • If the central claim is correct, robot relocalization in planetary-like terrain can be improved by adding geometric supervision to the map representation, independent of the pose estimator.
  • The combination of MVS and LiDAR supervision is shown to be complementary for the first time, implying that multi-modal geometric priors should be used jointly rather than as alternatives.
  • Geometry-aware 3DGS maps may enable reliable loop closure and drift correction for rovers operating in GNSS-denied environments.
  • The approach suggests that novel-view-synthesis representations can serve as a practical alternative to correspondence-based relocalization where features are aliased or sparse.
  • Pose error analysis indicates geometric supervision primarily reduces rotation errors, with median yaw errors below 25°, so the map's metric consistency directly stabilizes orientation recovery.

Reading between the lines

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

  • If the reported gains are real, the same geometric supervision recipe may transfer to other differentiable scene representations (e.g., neural radiance fields) and other sparse-view robotics settings, but the paper only tests 3DGS.
  • The 6.25%→43.2% jump might shrink if the photometric-only baseline were retrained under identical pipeline settings, because the reported comparison varies training iterations, normalization, and scene centering alongside geometric losses; a fair reader should treat that as an open confound.
  • A testable extension would be to replace LiDAR with a second RGB source of metric depth (e.g., stereo or structure from motion) to see whether the Chamfer anchor, not LiDAR per se, is what stabilizes the map.
  • The low absolute recall at the strict threshold (6.8%) suggests remaining local ambiguities; the paper's own failure cases indicate the bottleneck may be map resolution near submap boundaries rather than the supervision strategy.
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

3 major / 4 minor

Summary. The paper proposes a visual relocalization pipeline for challenging planetary-like environments, where low texture, aliasing, and sparse forward-motion viewpoints degrade classic correspondence-based methods. The map is represented by 3D Gaussian Splatting (3DGS) submaps, and the main contribution is a geometry-aware training objective that adds MVS-derived depth/normal losses and a LiDAR-guided Chamfer loss to the standard photometric loss. A coarse visual place recognition stage retrieves candidate submaps, and the 6DGS method estimates the 6-DoF pose against the fixed 3DGS representation. Experiments on a single sequence of the DLR S3LI Vulcano dataset report a pose-recall improvement from 6.25% (photometric-only 3DGS) to 43.2% (proposed method) under a (10 m, 15°) threshold, and lower Chamfer reconstruction errors, supporting the claim that geometric supervision is the key factor for relocalization in such environments.

Significance. If the central claim is established, the work would be a valuable demonstration that geometric supervision of 3DGS maps — especially combining zero-shot MVS and LiDAR — can dramatically improve single-image relocalization in perceptually difficult outdoor scenes, a setting where feature-based methods fail. The paper ships code, uses a real-world planetary-analog dataset, and evaluates with a fixed pose estimator (6DGS), which are strengths. However, the main controlled-ablation premise currently does not hold: the photometric baseline and geometry-aware variants differ in pipeline, iteration count, and coordinate-frame normalization, so the headline recall gain cannot be unambiguously attributed to the proposed losses. The evaluation is also restricted to a single sequence, which tempers the generality of the claims.

major comments (3)
  1. [Sec. V-C and V-E] The claim in Sec. V-E that 'only the training strategy of the underlying 3DGS map is varied' is contradicted by the training details in Sec. V-C. The photometric baseline (3DGS) uses the splatfacto pipeline for 30,000 iterations with Nerfstudio's built-in pose normalization and scene centering enabled, while all geometry-aware variants use regsplatfacto for 20,000 iterations with normalization and centering disabled and direct metric-frame optimization. These are not part of the proposed geometric losses, but they affect the optimization and coordinate frame. Therefore, the 6.25%→43.2% recall gain in Table II and the 74% Chamfer reduction in Table I may arise from pipeline/iteration/normalization changes instead of the MVS and LiDAR supervision. The paper must add a controlled ablation where the pipeline, iteration count, and normalization settings are held fixed and only the loss terms
  2. [Sec. V-E, Table II] All experimental results are reported on a single sequence ('moon lake') from the DLR S3LI Vulcano dataset. No results on other sequences, no repeated runs, and no statistical significance tests are provided. Given the variability of visual relocalization in such environments, the strong quantitative claims (e.g., 'substantial gains', 'first demonstration') are not yet supported as general findings. The authors should either evaluate on additional sequences from the dataset or explicitly restrict the claims to this sequence and provide measures of variance.
  3. [Sec. III-B and V-D] The geometric supervision relies on two external sources: MVS depth/normal predictions from MVSAnywhere [38] and submap poses from the LiDAR/VIO SLAM system [41]. The paper assumes these are accurate enough to serve as training targets, but it provides no quantitative validation of these supervision signals against the d-GNSS ground truth or any analysis of how errors in these signals affect the reconstructed map and downstream pose recall. Since low texture and sparse baselines are precisely the conditions that degrade MVS, this is a load-bearing assumption. A sensitivity test (e.g., perturbing the LiDAR poses, or comparing rendered depths to d-GNSS) would strengthen the causal interpretation of the reported gains.
minor comments (4)
  1. [Sec. II-B] Typo: 'In constrast' should be 'In contrast'.
  2. [Sec. II-A] Typo: 'retireving' should be 'retrieving'.
  3. [Eq. (3) and Eq. (7)] Notation inconsistency: Eq. (3) defines the normal loss as LMSVN, while Eq. (7) uses the weight λMVSN and the text refers to LMVSN. Please unify the superscript/subscript notation.
  4. [Table I] The row '3DGS + MVSA' shows a Chamfer distance of 4.03, worse than the baseline 3DGS at 3.36. The caption should explicitly state that lower Chamfer is better and that this row illustrates how MVSA alone can degrade geometric fidelity, to prevent misreading.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central result is an empirical comparison of training variants, and no equation reduces to a fitted target; the ablation-confounding issue is an experimental-control concern, not a self-referential reduction.

full rationale

The paper's central claim is that geometry-aware training of a 3DGS map improves single-image relocalization. This is established by an empirical comparison (Table II) of pose recall under different supervision strategies, not by a derivation. The geometric losses in Eqs. (2)-(6) are supervision objectives that constrain the map during training; the downstream pose estimator (6DGS) is fixed and is not fitted to the reported pose-recall numbers. No fitted parameter is renamed as a prediction, and no equation is defined in terms of the target result. The MVS and LiDAR depths are external supervision signals, not derived from the pose-recall metric. The self-citations to the dataset [7], submap SLAM [41], place-recognition pipeline [39], and MVSAnywhere [38] provide data and pretrained tools; they are not invoked as proof of the paper's central claim, and they are not uniqueness theorems or ansatz-smuggling devices. The most notable weakness is that the claimed controlled ablation ("only the training strategy of the underlying 3DGS map is varied") is not actually clean, because Sec. V-C shows the photometric baseline uses splatfacto for 30k iterations with pose normalization and scene centering enabled, while geometry-aware variants use regsplatfacto for 20k iterations with normalization disabled and direct metric-frame optimization. That is a confounding-variable problem affecting causal attribution, but it is not circularity: the reported recall values are not constructed to equal the loss terms, and the result would not become true by definition even if the ablation were clean. Therefore no significant circularity is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new physical objects or representations beyond 3DGS are introduced; the contribution is a combination of existing supervision signals. The main hand-set elements are the loss weights and schedules, captured as free parameters. The empirical claim rests on the accuracy of self-supplied LiDAR/VIO poses and zero-shot MVS depths.

free parameters (5)
  • λ_MVSD (MVS depth loss weight) = 0.05
    Weight in Eq. (7); chosen by hand, no sensitivity analysis reported.
  • λ_MVSN (MVS normal loss weight) = 0.1
    Weight in Eq. (7); chosen by hand.
  • λ_Ch (LiDAR Chamfer loss weight) = 5×10^-5
    Final weight in Eq. (7); ramped linearly from 0 between iterations 2,000 and 8,000.
  • λ_comp (Chamfer completeness weight) = unspecified
    Appears in Eq. (4) but no numerical value is reported; balances L_acc and L_comp.
  • Training schedule (iteration counts and ramp) = 30k/20k; ramp 2k-8k
    Photometric baseline trained 30k steps, geometry variants 20k; Chamfer activated at 2k. This schedule affects the comparison and is treated as fixed by the authors.
assumptions (3)
  • domain assumption Submap LiDAR point clouds and keyframe poses from the authors' multi-modal SLAM [41] are metrically accurate and mutually aligned; geometry-aware 3DGS is optimized directly in metric scale without pose refinement.
    Sec. V-A: keyframe poses fixed, LiDAR aggregated from local VIO estimates. If VIO drift or extrinsics are wrong, the depth/Chamfer targets inject bias into the map and pose results.
  • domain assumption MVSAnywhere zero-shot depth/normal predictions are sufficiently accurate on the planetary-analog terrain to serve as supervision.
    Sec. III-B: depths/normals used in Eqs. (2)-(3) with no reported accuracy check on this domain; the authors themselves note MVS degrades under sparse parallax.
  • ad hoc to paper The change from splatfacto to regsplatfacto, from 30k to 20k iterations, and from enabled to disabled pose normalization/scene centering has no material effect on downstream pose recall, so recall differences isolate the geometric losses.
    Sec. V-C: required for the ablation logic in Sec. V-E; contradicted by the paper's own description of different training settings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Relocalization from Sparse Views in Aliased and Low-Texture Environments via Novel View Synthesis." pith.science (2026). https://pith.science/paper/3CM7WSX2

@misc{pith2026260722147,
  author       = {Pith},
  title        = {Pith review of: Visual Relocalization from Sparse Views in Aliased and Low-Texture Environments via Novel View Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CM7WSX2}},
  note         = {Machine review of arXiv:2607.22147}
}
read the original abstract

Visual localization becomes extremely challenging in planetary-like terrains characterized by low texture, perceptual aliasing, harsh illumination, and sparse, weakly overlapping viewpoints induced by forward rover motion and unconstrained driving directions. Under these conditions, state-of-the-art image-to-image and image-to-map matching pipelines suffer significant performance degradation. In this work, we propose a visual relocalization method that departs from classical correspondence-based pipelines by directly estimating camera poses against a differentiable map representation built with 3D Gaussian Splatting (3DGS). Our key contribution is a geometry-aware training strategy that combines photometric and geometric losses, where the geometric supervision is provided for the first time by combining multi-view stereo (MVS) and LiDAR depths. We show that this joint optimization produces a 3DGS model that better fits the underlying scene geometry, leading to improved photometric and geometric consistency and more robust, accurate single-image 6-DoF pose estimation. Extensive experiments on data acquired in planetary-analog environments validate the effectiveness of our approach, showing substantial gains in relocalization accuracy under challenging conditions. Code is available at https://github.com/DLR-RM/multimodal-gsplat-relocalization.

Figures

Figures reproduced from arXiv: 2607.22147 by the authors.

Figure 1
Figure 1. Illustration of relocalization challenges in extreme, planetary￾like environments. Despite the three images at the bottom of the figure overlap, the large viewpoint change and weak, aliased textures pose sub￾stantial challenges for visual relocalization, i.e., estimating the six-degrees￾of-freedom camera pose within the map shown at the top of the figure. image-to-map manner against dense 3D scene representa￾tions w… view at source ↗
Figure 2
Figure 2. Overview of our relocalization pipeline, based on structure-preserving 3DGS. RGB images and LiDAR readings are grouped into submaps, and used to train a 3DGS with strong geometric supervision, in addition to the standard photometric supervision. At inference, the query image is used to retrieve top-k candidate submaps, and camera pose is estimated by minimizing the rendering error against the 3DGS representations. I… view at source ↗
Figure 3
Figure 3. Top-down view of the moon lake sequence used in this work, from the DLR S3LI Vulcano dataset. Impressions from the sequence are presented in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of Gaussian Splatting variants on two representative submaps. Rows alternate RGB renderings and depth maps (blue denotes closer regions and red denotes farther ones), while columns compare different supervision strategies. The rectangle overlaid …
Figure 6
Figure 6. Figure 6: Illustrative example of SuperPoint and SuperGlue feature matching between a query and a retrieved image. Several incorrect correspondences are observed, particularly at close range, caused by aliasing and significant viewpoint changes. TABLE II POSE RECALL COMPARISON B…
Figure 7
Figure 7. Figure 7: Translation and rotation estimation error distribution for relocalization baselines and ours. spatial alignments. The proposed geometry-aware represen￾tation substantially reduces rotation errors at all ranks, with median values consistently below 25◦ , while maintaini…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 2 linked inside Pith

  1. [38]

    Mvsanywhere: Zero-shot multi-view stereo,

    S. Izquierdo, M. Sayed, M. Firman, G. Garcia-Hernando, D. Tur- mukhambetov, J. Civera, O. Mac Aodha, G. Brostow, and J. Watson, “Mvsanywhere: Zero-shot multi-view stereo,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11 493–11 504

    2025

  2. [41]

    Multi-modal loop closing in unstructured planetary environments with visually enriched submaps,

    R. Giubilato, M. Vayugundla, W. St ¨urzl, M. J. Schuster, A. Wedler, and R. Triebel, “Multi-modal loop closing in unstructured planetary environments with visually enriched submaps,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 8758–8765

    2021

  3. [1]

    Present and future of slam in extreme environments: The darpa subt challenge,

    K. Ebadi, L. Bernreiter, H. Biggie, G. Catt, Y . Chang, A. Chatterjee, C. E. Denniston, S.-P. Desch ˆenes, K. Harlow, S. Khattaket al., “Present and future of slam in extreme environments: The darpa subt challenge,”IEEE Transactions on Robotics, vol. 40, pp. 936–959, 2023

    2023

  4. [2]

    Slam for autonomous planetary rovers with global localization,

    D. Geromichalos, M. Azkarate, E. Tsardoulias, L. Gerdes, L. Petrou, and C. Perez Del Pulgar, “Slam for autonomous planetary rovers with global localization,”Journal of Field Robotics, vol. 37, no. 5, pp. 830– 847, 2020

    2020

  5. [3]

    Challenges of slam in extremely unstructured environments: the dlr planetary stereo, solid- state lidar, inertial dataset,

    R. Giubilato, W. St ¨urzl, A. Wedler, and R. Triebel, “Challenges of slam in extremely unstructured environments: the dlr planetary stereo, solid- state lidar, inertial dataset,”IEEE Robotics and Automation Letters, pp. 1–8, 2022

    2022

  6. [4]

    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

    arXiv 2025

  7. [5]

    How nerfs and 3d gaussian splatting are reshaping slam: a survey,

    F. Tosi, Y . Zhang, Z. Gong, S. Mattoccia, M. R. Oswald, E. Sandstrom, and M. Poggi, “How nerfs and 3d gaussian splatting are reshaping slam: a survey,”IEEE Transactions on Robotics, 2026

    2026

  8. [6]

    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

    2023

Show all 43 references
  1. [7]

    The s3li vulcano dataset: A dataset for multi-modal slam in unstructured planetary environments,

    R. Giubilato, M. G. M ¨uller, M. Sewtz, L. A. E. Gonzalez, J. Folkesson, and R. Triebel, “The s3li vulcano dataset: A dataset for multi-modal slam in unstructured planetary environments,”2026 IEEE Aerospace Conference, 2026

    2026

  2. [8]

    Visual place recognition: A tutorial [tutorial],

    S. Schubert, P. Neubert, S. Garg, M. Milford, and T. Fischer, “Visual place recognition: A tutorial [tutorial],”IEEE Robotics & Automation Magazine, vol. 31, no. 3, pp. 139–153, 2023

    2023

  3. [9]

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

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE transactions on robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  4. [10]

    Bags of binary words for fast place recognition in image sequences,

    D. G ´alvez-L´opez and J. D. Tardos, “Bags of binary words for fast place recognition in image sequences,”IEEE Transactions on robotics, vol. 28, no. 5, pp. 1188–1197, 2012

    2012

  5. [11]

    Epnp: An accurate o(n) solution to the pnp problem,

    V . Lepetit, F. Moreno-Noguer, and P. Fua, “Epnp: An accurate o(n) solution to the pnp problem,”International Journal of Computer Vision, vol. 81, no. 2, pp. 155–166, 2009

    2009

  6. [12]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,”Commun. ACM, vol. 24, no. 6, p. 381–395, 1981

    1981

  7. [13]

    Optimal transport aggregation for visual place recognition,

    S. Izquierdo and J. Civera, “Optimal transport aggregation for visual place recognition,” inProceedings of the ieee/cvf conference on computer vision and pattern recognition, 2024, pp. 17 658–17 668

    2024

  8. [14]

    Vggt: Visual geometry grounded transformer,

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

    2025

  9. [15]

    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,”arXiv preprint arXiv:2505.12549, 2025

    2025 arXiv

  10. [16]

    Bow3d: Bag of words for real-time loop closing in 3d lidar slam,

    Y . Cui, X. Chen, Y . Zhang, J. Dong, Q. Wu, and F. Zhu, “Bow3d: Bag of words for real-time loop closing in 3d lidar slam,”IEEE Robotics and Automation Letters, vol. 8, no. 5, pp. 2828–2835, 2022

    2022

  11. [17]

    Lcdnet: Deep loop closure de- tection and point cloud registration for lidar slam,

    D. Cattaneo, M. Vaghi, and A. Valada, “Lcdnet: Deep loop closure de- tection and point cloud registration for lidar slam,”IEEE Transactions on Robotics, vol. 38, no. 4, pp. 2074–2093, 2022

  12. [18]

    Fast and accurate deep loop closing and relocalization for reliable lidar slam,

    C. Shi, X. Chen, J. Xiao, B. Dai, and H. Lu, “Fast and accurate deep loop closing and relocalization for reliable lidar slam,”IEEE Transactions on Robotics, vol. 40, pp. 2620–2640, 2024

    2024

  13. [19]

    Convolutional networks for real-time 6-dof camera relocalization,

    A. Kendall, M. Grimes, and R. Cipolla, “Convolutional networks for real-time 6-dof camera relocalization,”CoRR, vol. abs/1505.07427, 2015

    2015 arXiv

  14. [20]

    Understanding the limitations of cnn-based absolute camera pose regression,

    T. Sattler, Q. Zhou, M. Pollefeys, and L. Leal-Taixe, “Understanding the limitations of cnn-based absolute camera pose regression,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3302–3312

    2019

  15. [21]

    Advances in neural rendering,

    A. Tewari, J. Thies, B. Mildenhall, P. Srinivasan, E. Tretschk, W. Yifan, C. Lassner, V . Sitzmann, R. Martin-Brualla, S. Lombardi, T. Simon, C. Theobalt, M. Nießner, J. T. Barron, G. Wetzstein, M. Zollh ¨ofer, and V . Golyanik, “Advances in neural rendering,”Computer Graphics...

    2022

  16. [22]

    View interpolation for image synthesis,

    S. E. Chen and L. Williams, “View interpolation for image synthesis,” inProceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, ser. SIGGRAPH ’93, 1993, p. 279–288

    1993

  17. [23]

    View morphing,

    S. M. Seitz and C. R. Dyer, “View morphing,” inProceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, ser. SIGGRAPH ’96, 1996, p. 21–30

    1996

  18. [24]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, 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

    2021

  19. [25]

    Mvsgaussian: Fast generalizable gaussian splatting reconstruc- tion from multi-view stereo,

    T. Liu, G. Wang, S. Hu, L. Shen, X. Ye, Y . Zang, Z. Cao, W. Li, and Z. Liu, “Mvsgaussian: Fast generalizable gaussian splatting reconstruc- tion from multi-view stereo,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 37–53

    2024

  20. [26]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T.-J. Cham, and J. Cai, “Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,” inEuropean conference on computer vision. Springer, 2024, pp. 370–386

    2024

  21. [27]

    Sparse2dgs: Sparse-view surface reconstruction using 2d gaussian splatting with dense point cloud,

    N. Takama, S. Ito, K. Ito, H.-T. Chen, and T. Aoki, “Sparse2dgs: Sparse-view surface reconstruction using 2d gaussian splatting with dense point cloud,” in2025 IEEE International Conference on Image Processing (ICIP). IEEE, 2025, pp. 2844–2849

    2025

  22. [28]

    Drivinggaussian: Composite gaussian splatting for surrounding dy- namic autonomous driving scenes,

    X. Zhou, Z. Lin, X. Shan, Y . Wang, D. Sun, and M.-H. Yang, “Drivinggaussian: Composite gaussian splatting for surrounding dy- namic autonomous driving scenes,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 634–21 643

    2024

  23. [29]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting,

    Y . Yan, H. Lin, C. Zhou, W. Wang, H. Sun, K. Zhan, X. Lang, X. Zhou, and S. Peng, “Street gaussians: Modeling dynamic urban scenes with gaussian splatting,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 156–173

    2024

  24. [30]

    A constrained optimization approach for gaussian splatting from coarsely-posed images and noisy lidar point clouds,

    J. Peng, T. H. E. Tse, K. Xu, W. Gao, and A. Yao, “A constrained optimization approach for gaussian splatting from coarsely-posed images and noisy lidar point clouds,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 2961–2970

    2025

  25. [31]

    iNeRF: Inverting neural radiance fields for pose estimation,

    L. Yen-Chen, P. Florence, J. T. Barron, A. Rodriguez, P. Isola, and T.- Y . Lin, “iNeRF: Inverting neural radiance fields for pose estimation,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021

    2021

  26. [32]

    6dgs: 6d pose estimation from a single image and a 3d gaussian splatting model,

    B. Matteo, T. Tsesmelis, S. James, F. Poiesi, and A. Del Bue, “6dgs: 6d pose estimation from a single image and a 3d gaussian splatting model,” Springer, pp. 420–436, 2024

    2024

  27. [33]

    Gs-lidar: Generating realistic lidar point clouds with panoramic gaussian splatting,

    J. Jiang, C. Gu, Y . Chen, and L. Zhang, “Gs-lidar: Generating realistic lidar point clouds with panoramic gaussian splatting,” 2025

    2025

  28. [34]

    Splatad: Real-time lidar and camera rendering with 3d gaussian splatting for autonomous driving,

    G. Hess, C. Lindstr ¨om, M. Fatemi, C. Petersson, and L. Svensson, “Splatad: Real-time lidar and camera rendering with 3d gaussian splatting for autonomous driving,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11 982–11 992

    2025

  29. [35]

    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

    2024

  30. [36]

    Loopsplat: Loop closure by registering 3d gaussian splats,

    L. Zhu, Y . Li, E. Sandstr ¨om, S. Huang, K. Schindler, and I. Armeni, “Loopsplat: Loop closure by registering 3d gaussian splats,” inInter- national Conference on 3D Vision (3DV), 2025

    2025

  31. [37]

    Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements,

    L. C. Sun, N. P. Bhatt, J. C. Liu, Z. Fan, Z. Wang, T. E. Humphreys, and U. Topcu, “Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE,...

    2024

  32. [39]

    Multi- modal loop closure detection with foundation models in severely unstructured environments,

    L. A. E. Gonzalez, J. Folkesson, R. Triebel, and R. Giubilato, “Multi- modal loop closure detection with foundation models in severely unstructured environments,”2026 IEEE international conference on robotics and automation (ICRA), 2026

    2026

  33. [40]

    The faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazar ´e, M. Lomeli, L. Hosseini, and H. J ´egou, “The faiss library,”IEEE Transactions on Big Data, 2025

    2025

  34. [42]

    Superpoint: Self- supervised interest point detection and description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superpoint: Self- supervised interest point detection and description,” inProceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 224–236

    2018

  35. [43]

    Su- perglue: Learning feature matching with graph neural networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Su- perglue: Learning feature matching with graph neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4938–4947

    2020

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.