Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Stereo 3D Gaussian Splatting SLAM for Outdoor Urban Scenes

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Passive stereo RGB pairs plus a pretrained stereo network can drive 3D Gaussian Splatting SLAM outdoors and surpass RGB-D and LiDAR-guided baselines.

desk verdict A useful stereo-prior 3DGS-SLAM system with strong rendering numbers, but the evaluation compares full sequences against 300-frame baselines and the 'superior tracking' claim never isolates ORB-SLAM2, so the paper needs a fair comparison before its headline claims hold. read the letter →

arxiv 2507.23677 v1 pith:VDUH4V6L submitted 2025-07-31 cs.RO

Xiaohan Li , Ziren Gong , Fabio Tosi , Matteo Poggi , Stefano Mattoccia , Dong Liu , Jun Wu This is my paper
classification cs.RO
keywords 3DGaussianSplattingSLAMstereodepthestimationoutdoorscenesKITTIKITTI-360neuralrenderingautonomousdriving
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

BGS-SLAM presents itself as the first 3D Gaussian Splatting (3DGS) SLAM system for outdoor scenes that uses only passive stereo RGB images, with no LiDAR or RGB-D sensor; in 3DGS, a scene is represented as colored 3D Gaussians rendered by fast rasterization. The core proposal is to let a pretrained deep stereo network produce dense metric depth maps from synchronized left-right pairs, then use those maps, together with RGB, normal, and smoothness losses, to supervise where and how 3D Gaussians are placed. Tracking is handled by an external feature-based stereo tracker, so pose estimation stays stable while the Gaussian map is optimized. On KITTI and KITTI-360 the paper reports higher rendering quality and lower depth error than the RGB-D 3DGS baselines it compares against, and it can process full kilometer-scale sequences where one baseline runs out of memory. If correct, this means an affordable stereo camera plus a good stereo network can substitute for expensive active depth sensors in outdoor mapping.

What carries the argument

The load-bearing mechanism is stereo depth supervision of a 3D Gaussian map. A pretrained stereo network, MonSter-M by default, computes a dense disparity map from a synchronized left-right image pair, and the map is converted to metric depth by $\tilde{D} = f\cdot b / d$, where $f$ is the focal length and $b$ the stereo baseline. In the mapping thread, a weighted depth loss samples about 25% of pixels on a regular grid to reduce the influence of locally correlated stereo errors, an RGB-gradient weight $g_{rgb} = \exp(-\nabla I)$ downweights textureless regions, and normal-consistency and smoothness losses enforce geometric coherence. External ORB-SLAM2-style tracking supplies poses, and keyframing inserts frames when ORB keypoint IoU or relative motion exceeds thresholds. This combination lets a map of anisotropic 3D Gaussians be driven by dense but imperfect depth estimates instead of exact LiDAR points.

What would settle it

Run the full pipeline on a held-out outdoor stereo sequence with ground-truth LiDAR while replacing MonSter-M with a broken depth source, such as a constant-disparity map, keeping all other components identical; if rendering and depth quality remain high, the RGB and regularization losses are doing the work rather than the stereo depth prior, and if quality collapses, the claimed role of deep stereo supervision is confirmed.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that approximate metric depth from deep stereo matching is enough to anchor a 3D Gaussian map outdoors. The pipeline runs two threads: a tracking thread based on ORB feature matching and local bundle adjustment estimates camera poses, while a mapping thread optimizes 3D Gaussians under a weighted, 25%-sampled depth loss derived from the stereo network's disparity map, plus normal-consistency and smoothness losses and optional sky segmentation. The authors report that this stereo-supervised variant reaches an average PSNR of 21.54 dB on KITTI and 23.66 dB on KITTI-360, substantially above SplaTAM, MonoGS, and Point-SLAM, which were evaluated only on the first 300 frames per sequence, and that it matches the tracking accuracy of its own LiDAR-guided variant while using no active sensor. Memory use stays within 48 GB for full sequences where SplaTAM fails, because the system keeps only keyframes and removes low-opacity Gaussians. The authors state that the system is not real-time, with 0.24 s per frame for tracking and 1.37 s per frame for mapping.

Load-bearing premise

The whole approach assumes the pretrained stereo network returns dependable metric depth on the scenes being mapped; biased or failed depth estimates would degrade the Gaussian map no matter how the losses are weighted.

Editorial extensions

If this is right

  • A calibrated stereo pair plus a pretrained stereo network can replace LiDAR or RGB-D sensors for outdoor 3DGS mapping: the geometry needed to place Gaussians comes from dense stereo depth rather than active sensing.
  • Dense stereo supervision yields much better rendering than sparse LiDAR supervision in this system: average PSNR on KITTI is 21.54 dB for the stereo variant versus 9.30 dB for the LiDAR-guided variant.
  • The system can process kilometer-scale KITTI sequences within 48 GB of GPU memory, while the SplaTAM baseline exhausts memory after a few hundred frames, so the keyframing and Gaussian-pruning strategy scales better.
  • Pose accuracy is decoupled from mapping quality: ATE matches the LiDAR-guided variant because poses come from the ORB-SLAM2-style tracker, not from the Gaussian optimization.
  • The method is not real-time, with 0.24 s per frame for tracking and 1.37 s per frame for mapping, plus stereo network inference time, so practical SLAM use awaits further acceleration.

Reading between the lines

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

  • If the stereo depth prior is the active ingredient, the system's quality should track the accuracy of whatever stereo network is plugged in; a useful test is to run the pipeline with an intentionally biased or weak disparity source and measure how quickly Depth L1 degrades.
  • Because tracking is ORB-SLAM2-style feature matching, the system inherits classical feature-based failure modes, such as drift in low-texture or highly repetitive urban scenes, even when stereo depth is still informative; coupling pose optimization to Gaussian rendering might close that gap.
  • The 25% regular-grid depth sampling suggests that dense, locally correlated stereo errors hurt optimization, so an adaptive sampling strategy guided by stereo confidence or RGB gradient could improve both accuracy and efficiency.
  • The fixed-baseline car setup leaves open whether the approach transfers to handheld or aerial stereo rigs with shorter baselines and more aggressive motion, which is a testable next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes BGS-SLAM, a stereo-only 3D Gaussian Splatting SLAM system aimed at large-scale outdoor urban scenes. The system uses ORB-SLAM2 for camera tracking, a pretrained deep stereo network (MonSter-M) to generate dense depth maps, and a keyframing module to select frames. The mapping thread optimizes a collection of 3D Gaussians with a combined RGB, weighted depth, normal-consistency, and smoothness loss. Experiments on KITTI and KITTI-360 report large improvements in rendering quality and depth accuracy over Point-SLAM, SplaTAM, and MonoGS, as well as a LiDAR-supervised variant of the same pipeline. The paper also reports an ablation of the loss components and of stereo backbones, and a memory-consumption analysis.

Significance. If the claims are supported, the paper makes a useful contribution: it shows that a passive stereo pair plus a pretrained stereo network can drive dense 3D Gaussian mapping in large outdoor scenes, replacing LiDAR or RGB-D supervision. The system design is constructive and the multi-loss formulation is well motivated. The ablation study is valuable and the limitations paragraph is honest about the lack of real-time operation. However, the headline claims of 'superior tracking accuracy and mapping performance' and of being 'the first binocular 3DGS-SLAM system for outdoor scenarios' go beyond what the current experimental protocol can establish, because the baselines are evaluated on only the first 300 frames while BGS-SLAM runs full sequences, and because the tracking contribution is not isolated. The paper is likely to be acceptable after a substantial revision of the evaluation and a precise scoping of the novelty claims.

major comments (4)
  1. [Experiments, Tables 3 and 4] The headline claim of superior tracking accuracy is not supported by the current protocol. All three baselines are evaluated on only the first 300 frames per sequence (as indicated by the dagger marks), while BGS-SLAM is evaluated on the full sequences of 801 to 14,000 frames. ATE RMSE is sensitive to trajectory length, so this comparison conflates method quality with evaluation length. Moreover, BGS-SLAM (LiDAR) and BGS-SLAM (Ours) report identical ATE values in every sequence, which confirms that tracking is entirely determined by the external ORB-SLAM2 module; nevertheless, no standalone ORB-SLAM2 baseline is reported. A proper comparison should evaluate all methods on identical frame ranges and should report ORB-SLAM2 by itself. Given the statement that results are averaged over three runs, the absence of standard deviations further weakens the tracking comparison.
  2. [Related Work and Experiments] The claims of being the first binocular 3DGS-SLAM system for outdoor scenarios and of achieving superior mapping performance are not tested against the most relevant state of the art. The Related Work section cites Photo-SLAM, which explicitly supports monocular, stereo, and RGB-D cameras, yet Photo-SLAM is not included in the experiments. The evaluated baselines (Point-SLAM, SplaTAM, and MonoGS) are primarily indoor-oriented or RGB-D methods, and MonoGS is run in RGB-D mode. To support the 'first' and 'superior' claims, the authors should either compare directly with a stereo-capable 3DGS-SLAM system such as Photo-SLAM, or substantially scope the claims to the specific configuration used in the evaluation.
  3. [Implementation Details and Table 2] The depth supervision that drives the mapping thread comes from MonSter-M, and the backbone ablation in Table 2 distinguishes models 'fine-tuned on KITTI datasets' from models 'trained on a mix of datasets' without reporting the exact training sets. Since the evaluation is performed on KITTI and KITTI-360, the manuscript does not establish that MonSter-M's training data is disjoint from the test sequences. If MonSter-M was trained on KITTI or KITTI-360, the measured depth-based gains could partly reflect overlap rather than zero-shot stereo generalization. Please report the training-data composition of MonSter-M, or evaluate on a held-out outdoor dataset, before drawing the conclusion that pretrained stereo depth is sufficient for the reported mapping quality.
  4. [Table 5] The memory-efficiency claim is not supported by the presented data. Table 5 reports GPU memory consumption only for SplaTAM and BGS-SLAM, despite the text saying that the table shows memory consumption for all methods. MonoGS and Point-SLAM are absent, so the reader cannot compare their memory usage on the same sequences. Without memory measurements for all baselines on comparable frame counts and sequence lengths, the conclusion that BGS-SLAM is efficient in memory relative to other 3DGS-based solutions is untested.
minor comments (5)
  1. [Eq. (17) surrounding text] There is a typo: 'appropiate' should be 'appropriate'.
  2. [Eqs. (13) and (14)] The notation for the stereo depth maps is inconsistent: Eq. (13) uses `Dsi` and `ˆDsi`, while Eq. (14) uses `Dsi,j` and `ˆDsi,j`. Please define the subscript convention clearly.
  3. [Eq. (16)] The smoothness loss uses `i,j` and `r` without specifying the range of `r` or the treatment of image boundaries in the finite differences; please clarify.
  4. [Limitations] The limitations paragraph gives tracking and mapping times per frame but not the inference time of the deep stereo network. Since the paper states that this network adds computational overhead, please report the end-to-end per-frame time including stereo inference.
  5. [Table 5] The table caption and the accompanying text say 'all methods,' but the table only contains SplaTAM and BGS-SLAM entries; please either add the missing methods or revise the wording.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the system is a constructive combination of a pretrained stereo network, ORB-SLAM2 tracking, and 3DGS mapping losses, with reported accuracy measured against external LiDAR ground truth.

full rationale

BGS-SLAM's derivation chain is constructive and empirically grounded. The mapping loss (Eq. 17) combines RGB, sampled depth, normal, and smoothness terms; the depth term compares rendered depth to stereo-network depth (Eq. 13-14), while the reported Depth L1 is measured against LiDAR ground truth, an independent external signal. Tracking is performed by ORB-SLAM2 (Eq. 8-10) and is explicitly decoupled from mapping: Tables 3-4 show identical ATE for BGS-SLAM (LiDAR) and BGS-SLAM (Ours), so the tracking claim is inherited from a classical tracker, not derived from the Gaussian optimization, which is a composition choice rather than a circular reduction. The choice of MonSter-M and loss weights is informed by ablations on KITTI (Tables 1-2), which is benchmark tuning and not a self-referential derivation. Self-citations to Tosi et al. (2024, 2025) support background assertions about stereo matching and SLAM surveys and are not load-bearing. The asymmetric baseline protocol (300 frames for baselines vs full sequences for BGS-SLAM) and the non-real-time limitation are experimental-validity and correctness concerns, not circularity. No equation in the paper is equivalent to the headline result by construction.

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

The paper introduces no new physical or mathematical entities. Its load-bearing elements are external components, MonSter-M and ORB-SLAM2, plus hand-set loss weights and sampling parameters. The central performance claim therefore rests on the unstated validity of the stereo depth prior on the test split and on the external tracker's pose accuracy.

free parameters (4)
  • Mapping loss weights = λrgb=0.8, λssim=0.2, λgeo=0.1, λn=0.1, λs=0.5
    Chosen by hand in Eq. 17 with no sensitivity study; the balance between color, depth, normal, and smoothness terms drives performance.
  • Depth sampling ratio = approximately 25% of pixels
    Selected from empirical observation on KITTI and central to the sampled depth loss in Eq. 14; the paper does not analyze sensitivity to this ratio.
  • Keyframe motion multiplier = 1.5x previous motion
    Adaptive threshold in Eq. 11; no norm is defined for the matrix comparison, so the effective threshold is ambiguous in practice.
  • 3DGS optimization hyperparameters = learning rates and opacity threshold listed in Implementation Details
    Learning rates for color, rotation, scale, opacity, and an opacity removal threshold of 5e-3 are stated but not justified or ablated.
assumptions (4)
  • domain assumption The pretrained stereo network produces metrically accurate depth maps in outdoor scenes.
    Relied on in Eqs. 5-7 and in every depth or normal loss; if the network is trained on the test distribution or fails in certain lighting, the geometric supervision is invalid.
  • domain assumption ORB-SLAM2 stereo tracking provides accurate camera poses throughout each long sequence.
    The tracking thread is an external component and all reported ATE values come from it; the paper does not validate poses independently or compare with ORB-SLAM2 alone.
  • domain assumption The 3DGS optimization with the chosen losses converges to a good scene representation.
    There is no convergence argument; the quality depends on initial poses, depth priors, keyframing, and hyperparameters.
  • domain assumption LiDAR ground truth is correctly aligned and masked for Depth L1 evaluation.
    Depth L1 is computed against LiDAR measurements, but the alignment, occlusion handling, and depth-range masking are not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stereo 3D Gaussian Splatting SLAM for Outdoor Urban Scenes." pith.science (2026). https://pith.science/paper/VDUH4V6L

@misc{pith2026250723677,
  author       = {Pith},
  title        = {Pith review of: Stereo 3D Gaussian Splatting SLAM for Outdoor Urban Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VDUH4V6L}},
  note         = {Machine review of arXiv:2507.23677}
}
read the original abstract

3D Gaussian Splatting (3DGS) has recently gained popularity in SLAM applications due to its fast rendering and high-fidelity representation. However, existing 3DGS-SLAM systems have predominantly focused on indoor environments and relied on active depth sensors, leaving a gap for large-scale outdoor applications. We present BGS-SLAM, the first binocular 3D Gaussian Splatting SLAM system designed for outdoor scenarios. Our approach uses only RGB stereo pairs without requiring LiDAR or active sensors. BGS-SLAM leverages depth estimates from pre-trained deep stereo networks to guide 3D Gaussian optimization with a multi-loss strategy enhancing both geometric consistency and visual quality. Experiments on multiple datasets demonstrate that BGS-SLAM achieves superior tracking accuracy and mapping performance compared to other 3DGS-based solutions in complex outdoor environments.

Figures

Figures reproduced from arXiv: 2507.23677 by the authors.

Figure 1
Figure 1. Comparison of Rendering and Depth Estimation. The top row shows RGB renderings generated by SplaTAM (Keetha et al. 2024), our BGS-SLAM method trained with LiDAR depth points, and our approach using only stereo RGB pairs with depth maps from deep stereo networks for supervision. The bottom row presents the corresponding depth renderings. Abstract 3D Gaussian Splatting (3DGS) has recently gained popular￾ity in SLAM ap… view at source ↗
Figure 2
Figure 2. Framework Overview. BGS-SLAM uses stereo images to reconstruct outdoor environments using 3D Gaussians. A pre-trained stereo network extracts dense depth maps from the stereo pairs, with optional sky masking to improve reconstruc￾tion. The tracking thread estimates camera poses through feature matching and local bundle adjustment, while the keyframing thread maintains a buffer of key observations. In the mapping thr… view at source ↗
Figure 3
Figure 3. Visualization of rendering quality on KITTI. pared to existing radiance-field SLAM approaches without requiring expensive LiDAR sensors. Limitations. BGS-SLAM does not yet operate in real￾time, with average tracking and mapping times of 0.24 s and 1.37 s per frame, respectively—posing a limitation for practical SLAM applications. The computational overhead is further increased by the inference time of the deep stere… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MAGiSt3R: Multi-Agent Feed-forward 3D Reconstruction from Monocular RGB Videos

    cs.CV 2026-07 conditional novelty 6.0 of 10

    MAGiSt3R is a multi-agent feed-forward 3D reconstruction system using a learned submap-merging model (MAGMA) and pose graph optimization to align local maps from multiple monocular RGB cameras into one consistent map ...

  2. Pocket-SLAM: Rendering-Area-Aware Pruning for Memory-Efficient 3DGS-SLAM

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    Pocket-SLAM introduces rendering-area-aware pruning for 3DGS-SLAM, claiming over 60% memory reduction and 2x FPS gain on EuRoC and KITTI while keeping localization and mapping accuracy.

Reference graph

Works this paper leans on

59 extracted references · 44 canonical work pages · cited by 2 Pith papers

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bartolomei, L.; Tosi, F.; Poggi, M.; and Mattoccia, S. 2025. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail. In Proceedings of the Computer Vision and Pattern Recognition Conference, 1013--1027

    work page 2025

  4. [4]

    Campos, C.; Elvira, R.; Rodr \' guez, J. J. G.; Montiel, J. M.; and Tard \'o s, J. D. 2021. Orb-slam3: An accurate open-source library for visual, visual--inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874--1890

    work page 2021

  5. [5]

    Chen, G.; and Wang, W. 2024. A Survey on 3D Gaussian Splatting. arXiv preprint arXiv:2401.03890

    arXiv 2024

  6. [6]

    Chen, Z.; Long, W.; Yao, H.; Zhang, Y.; Wang, B.; Qin, Y.; and Wu, J. 2024. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27768--27777

    work page 2024

  7. [7]

    Cheng, J.; Liu, L.; Xu, G.; Wang, X.; Zhang, Z.; Deng, Y.; Zang, J.; Chen, Y.; Cai, Z.; and Yang, X. 2025. Monster: Marry monodepth to stereo unleashes power. In Proceedings of the Computer Vision and Pattern Recognition Conference, 6273--6282

    work page 2025

  8. [8]

    Geiger, A.; Lenz, P.; and Urtasun, R. 2012 a . Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, 3354--3361. IEEE

    work page 2012

Show all 59 references
  1. [9]

    Geiger, A.; Lenz, P.; and Urtasun, R. 2012 b . Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR)

    2012

  2. [10]

    Hong, S.; and et al. 2024. LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real-time 3D Radiance Field Map Rendering. IEEE Robotics and Automation Letters

    2024

  3. [11]

    Hu, J.; Chen, X.; Feng, B.; Li, G.; Yang, L.; Bao, H.; Zhang, G.; and Cui, Z. 2024. CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field. In European Conference on Computer Vision (ECCV)

    2024

  4. [12]

    Huang, H.; Li, L.; Cheng, H.; and Yeung, S.-K. 2023. Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular, Stereo, and RGB-D Cameras. arXiv preprint arXiv:2311.16728

    2023 arXiv

  5. [13]

    Jiang, H.; Lou, Z.; Ding, L.; Xu, R.; Tan, M.; Jiang, W.; and Huang, R. 2025. Defom-stereo: Depth foundation model based stereo matching. In Proceedings of the Computer Vision and Pattern Recognition Conference, 21857--21867

    2025

  6. [14]

    C.; and Schindler, K

    Ke, B.; Obukhov, A.; Huang, S.; Metzger, N.; Daudt, R. C.; and Schindler, K. 2024. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9492--9502

    2024

  7. [15]

    M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J

    Keetha, N.; Karhade, J.; Jatavallabhula, K. M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J. 2024. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21357--21366

    2024

  8. [16]

    Kendall, A.; Martirosyan, H.; Dasgupta, S.; Henry, P.; Kennedy, R.; Bachrach, A.; and Bry, A. 2017. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international conference on computer vision, 66--75

    2017

  9. [17]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4)

    2023

  10. [18]

    Li, H.; Gu, X.; Yuan, W.; Yang, L.; Dong, Z.; and Tan, P. 2023. Dense RGB SLAM With Neural Implicit Maps. In Proceedings of the International Conference on Learning Representations

    2023

  11. [19]

    Li, R.-W.; Ke, W.; Li, D.; Tian, L.; and Barsoum, E. 2024. MonoGS ++: Fast and Accurate Monocular RGB Gaussian SLAM. In British Conference on Machine Vision (BMVC)

    2024

  12. [20]

    Li, X.; Liu, D.; and Wu, J. 2024. CTO-SLAM: contour tracking for object-level robust 4D SLAM. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 10323--10331

    2024

  13. [21]

    X.; Taylor, R

    Li, Z.; Liu, X.; Drenkow, N.; Ding, A.; Creighton, F. X.; Taylor, R. H.; and Unberath, M. 2021. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 6197--6206

    2021

  14. [22]

    Liao, Y.; Xie, J.; and Geiger, A. 2022. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3): 3292--3310

    2022

  15. [23]

    Lipson, L.; Teed, Z.; and Deng, J. 2021. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), 218--227. IEEE

    2021

  16. [24]

    H.; and Davison, A

    Matsuki, H.; Murai, R.; Kelly, P. H.; and Davison, A. J. 2024 a . Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18039--18048

    2024

  17. [25]

    Matsuki, H.; Murai, R.; Kelly, P. H. J.; and Davison, A. J. 2024 b . G aussian S platting SLAM . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  18. [26]

    Matsuki, H.; Sucar, E.; Laidow, T.; Wada, K.; Scona, R.; and Davison, A. J. 2023. iMODE: Real-Time Incremental Monocular Dense Mapping Using Neural Field. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 4171--4177. IEEE

    2023

  19. [27]

    Mayer, N.; Ilg, E.; Hausser, P.; Fischer, P.; Cremers, D.; Dosovitskiy, A.; and Brox, T. 2016. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recogniti...

    2016

  20. [28]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  21. [29]

    Mur-Artal, R.; Montiel, J. M. M.; and Tardos, J. D. 2015. ORB-SLAM: a versatile and accurate monocular SLAM system. IEEE transactions on robotics, 31(5): 1147--1163

    2015

  22. [30]

    Mur-Artal, R.; and Tard \'o s, J. D. 2017. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5): 1255--1262

    2017

  23. [31]

    Naumann, J.; Xu, B.; Leutenegger, S.; and Zuo, X. 2024. NeRF-VO: Real-Time Sparse Visual Odometry With Neural Radiance Fields. IEEE Robotics and Automation Letters

    2024

  24. [32]

    Peng, Z.; Shao, T.; Liu, Y.; Zhou, J.; Yang, Y.; Wang, J.; and Zhou, K. 2024. Rtg-slam: Real-time 3d reconstruction at scale using gaussian splatting. In ACM SIGGRAPH 2024 Conference Papers, 1--11

    2024

  25. [33]

    Oswald, M

    Sandström, E.; Li, Y.; Van Gool, L.; and R. Oswald, M. 2023. Point-SLAM: Dense Neural Point Cloud-based SLAM. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    2023

  26. [34]

    Sucar, E.; Liu, S.; Ortiz, J.; and Davison, A. J. 2021. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 6229--6238

    2021

  27. [35]

    C.; Bhatt, N

    Sun, L. C.; Bhatt, N. P.; Liu, J. C.; Fan, Z.; Wang, Z.; Humphreys, T. E.; and Topcu, U. 2024 a . MM3DGS SLAM: Multi-modal 3D Gaussian Splatting for SLAM Using Vision, Depth, and Inertial Measurements. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    2024

  28. [36]

    J.; and Magnusson, M

    Sun, S.; Mielle, M.; Lilienthal, A. J.; and Magnusson, M. 2024 b . High-Fidelity SLAM Using Gaussian Splatting with Rendering-Guided Densification and Regularized Optimization. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE

    2024

  29. [37]

    Teed, Z.; Lipson, L.; and Deng, J. 2022. Deep patch visual odometry. arXiv preprint arXiv:2208.04726

    2022 arXiv

  30. [38]

    Tosi, F.; Bartolomei, L.; and Poggi, M. 2025. A Survey on Deep Stereo Matching in the Twenties. International Journal of Computer Vision

    2025

  31. [39]

    R.; and Poggi, M

    Tosi, F.; Zhang, Y.; Gong, Z.; Sandstr \"o m, E.; Mattoccia, S.; Oswald, M. R.; and Poggi, M. 2024. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4: 1

    2024 arXiv

  32. [40]

    Wang, H.; Wang, J.; and Agapito, L. 2023. Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13293--13302

    2023

  33. [41]

    Wang, R.; Schworer, M.; and Cremers, D. 2017. Stereo DSO: Large-scale direct sparse visual odometry with stereo cameras. In Proceedings of the IEEE international conference on computer vision, 3903--3911

    2017

  34. [42]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600--612

    2004

  35. [43]

    Wen, B.; Trepte, M.; Aribido, J.; Kautz, J.; Gallo, O.; and Birchfield, S. 2025. Foundationstereo: Zero-shot stereo matching. In Proceedings of the Computer Vision and Pattern Recognition Conference, 5249--5260

    2025

  36. [44]

    Wu, C.; Duan, Y.; Zhang, X.; Sheng, Y.; Ji, J.; and Zhang, Y. 2024. MM-Gaussian: 3D Gaussian-based Multi-modal Fusion for Localization and Reconstruction in Unbounded Scenes. In RSJ International Conference on Intelligent Robots and Systems (IROS)

    2024

  37. [45]

    M.; and Luo, P

    Xie, E.; Wang, W.; Yu, Z.; Anandkumar, A.; Alvarez, J. M.; and Luo, P. 2021. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077--12090

    2021

  38. [46]

    Xu, G.; Wang, X.; Ding, X.; and Yang, X. 2023. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 21919--21928

    2023

  39. [47]

    Xu, G.; Wang, X.; Zhang, Z.; Cheng, J.; Liao, C.; and Yang, X. 2024. IGEV++: iterative multi-range geometry encoding volumes for stereo matching. arXiv preprint arXiv:2409.00638

    2024 arXiv

  40. [48]

    Yan, C.; Qu, D.; Xu, D.; Zhao, B.; Wang, Z.; Wang, D.; and Li, X. 2024. GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  41. [49]

    Yang, L.; Kang, B.; Huang, Z.; Xu, X.; Feng, J.; and Zhao, H. 2024 a . Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10371--10381

    2024

  42. [50]

    Yang, L.; Kang, B.; Huang, Z.; Zhao, Z.; Xu, X.; Feng, J.; and Zhao, H. 2024 b . Depth anything v2. Advances in Neural Information Processing Systems, 37: 21875--21911

    2024

  43. [51]

    Yang, X.; Li, H.; Zhai, H.; Ming, Y.; Liu, Y.; and Zhang, G. 2022. Vox-Fusion: Dense tracking and mapping with voxel-based neural implicit representation. In 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), 499--507. IEEE

    2022

  44. [52]

    Zeng, J.; Yao, C.; Wu, Y.; and Jia, Y. 2024. Temporally consistent stereo matching. In European Conference on Computer Vision, 341--359. Springer

    2024

  45. [53]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

    2018

  46. [54]

    Zhang, W.; Sun, T.; Wang, S.; Cheng, Q.; and Haala, N. 2023 a . Hi-slam: Monocular real-time dense mapping with hybrid implicit fields. IEEE Robotics and Automation Letters

    2023

  47. [55]

    Zhang, Y.; Tosi, F.; Mattoccia, S.; and Poggi, M. 2023 b . Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3727--3737

    2023

  48. [56]

    Zhou, H.; Guo, Z.; Ren, Y.; Liu, S.; Zhang, L.; Zhang, K.; and Li, M. 2024. MoD-SLAM: Monocular Dense Mapping for Unbounded 3D Scene Reconstruction

    2024

  49. [57]

    Zhu, P.; Zhuang, Y.; Chen, B.; Li, L.; Wu, C.; and Liu, Z. 2024 a . MGS-SLAM: Monocular Sparse Tracking and Gaussian Mapping with Depth Smooth Regularization. IEEE Robotics and Automation Letters

    2024

  50. [58]

    R.; Geiger, A.; and Pollefeys, M

    Zhu, Z.; Peng, S.; Larsson, V.; Cui, Z.; Oswald, M. R.; Geiger, A.; and Pollefeys, M. 2024 b . NICER-SLAM: Neural Implicit Scene Encoding for RGB SLAM. In International Conference on 3D Vision (3DV)

    2024

  51. [59]

    R.; and Pollefeys, M

    Zhu, Z.; Peng, S.; Larsson, V.; Xu, W.; Bao, H.; Cui, Z.; Oswald, M. R.; and Pollefeys, M. 2022. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12786--12796

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.