Pith. sign in

REVIEW 4 major objections 5 minor 23 references

CuSfM claims that pose-graph priors make offline structure-from-motion six times faster and more accurate than the standard incremental baseline.

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

2026-08-04 09:24 UTC pith:LURIPG67

load-bearing objection A capable trajectory-refinement and mapping system, but the headline COLMAP comparison is unfair and the 'order-of-magnitude' speedup is contradicted by the paper's own Table 2. the 4 major comments →

arxiv 2510.15271 v1 pith:LURIPG67 submitted 2025-10-17 cs.CV cs.RO

CuSfM: CUDA-Accelerated Structure-from-Motion

Jingrui Yu , Jun Liu , Kefei Ren , Joydeep Biswas , Rurui Ye , Keqiang Wu , Chirag Majithia , Di Zeng This is my paper
classification cs.CV cs.RO MSC 68T4568U10
keywords structure-from-motionGPU accelerationpose graphview graphbundle adjustmentloop closurestereo relative poseextrinsic calibration
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 proposes CuSfM, an offline structure-from-motion system that treats reconstruction as trajectory refinement: it takes rough initial camera poses (typically supplied by a visual odometry or SLAM front-end) plus a video sequence, and outputs refined poses, a sparse 3D map, and refined extrinsic parameters for multi-camera rigs. Its key claim is that building the view graph from pose-graph edges — sequential links, loop-closure links, and extrinsic links between synchronized cameras — removes redundant feature matching and lets a GPU-accelerated pipeline finish the whole job in about 17% of the time of the standard incremental SfM pipeline, while in most test sequences also reducing absolute trajectory error. The authors further claim that a new stereo relative-pose routine recovers translation scale from 2D correspondences alone, avoiding triangulation-and-PnP heuristics, and that the system supports map updating and localization in addition to batch reconstruction. Read sympathetically, this is a claim about where SfM effort should be spent: refining already-good trajectories rather than reconstructing from scratch.

Core claim

The paper's central discovery is that pose-graph priors provide a minimal, non-redundant set of image associations for SfM, and that this selection beats both exhaustive and radius-based matching in speed and accuracy. On the algorithmic side, it introduces a stereo relative-pose method that recovers the absolute translation scale by composing two essential-matrix poses with a known calibrated baseline, then refines the six-degree-of-freedom pose by minimizing Sampson distance across both stereo pairs. Experiments on driving and simulated multi-camera data show the view-graph strategy consistently outperforms a radius-based strategy, and that enabling extrinsic refinement further reduces abs

What carries the argument

The load-bearing object is the pose-graph view graph: a graph whose nodes are camera (or vehicle-rig) poses and whose edges are only sequential links, loop-closure links, and extrinsic links between synchronized cameras. This replaces the exhaustive or radius-based candidate-pair selection that typically dominates feature-matching time in conventional SfM, supplying exactly the associations needed for robust triangulation while avoiding redundant ones. A second piece is the three-view translation-scale estimation, which composes two essential-matrix pose hypotheses with a known calibrated stereo baseline to fix scale using only 2D points, then refines the result by minimizing Sampson distanc

Load-bearing premise

The system requires a prior trajectory and a time-continuous image sequence as input; if such priors are unavailable, inaccurate, or already encode the scene structure, the claimed accuracy and speed advantages over prior-free SfM systems do not follow.

What would settle it

Run CuSfM on an unordered image collection with no initial poses (or with identity/random priors) and compare against the same baseline SfM systems on identical inputs. If accuracy collapses or runtime balloons past the baseline, the claim that cuSfM is a generally faster and more accurate SfM system fails; it survives only as a trajectory-refinement method.

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

If this is right

  • Offline SfM can be repositioned as a trajectory-refinement stage for any visual odometry or SLAM front-end, cutting wall-clock time by roughly 83% on the tested driving benchmark.
  • Multi-camera vehicle rigs can jointly optimize poses and inter-camera extrinsics, with reported trajectory-error reductions of 3–32% when extrinsic refinement is enabled.
  • The non-redundant view-graph principle suggests that exhaustive frame matching is wasteful; other reconstruction pipelines could adopt pose-graph edges as a cheap candidate-selection pre-filter.
  • The stereo scale-recovery routine could replace triangulation-and-PnP in stereo front-ends, removing a known source of heuristic error in two-view pose estimation.
  • Map-updating and localization modes allow incremental, crowdsourced mapping with either fixed or adjustable existing maps.

Where Pith is reading between the lines

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

  • The paper's headline speed claim ('order of magnitude') is not fully supported by its own runtime table, which shows a 5.9x total speedup; the 10x claim appears to apply only to the mapping stage. A careful reader should take the total measured speedup as the honest number.
  • The accuracy comparisons are not apples-to-apples: CuSfM receives initial poses (in one experiment, ground truth corrupted by noise) while the comparison SfM systems receive only raw monocular images. Part of the reported accuracy gap may be inherited from the initializer rather than produced by the refinement itself; a fair test would give identical priors to all systems or remove priors from CuS
  • The view-graph non-redundancy idea is separable from the GPU backend and could be tested as a drop-in replacement for candidate-pair selection in other SfM systems.
  • The extrinsic-refinement gains suggest that modeling a multi-camera vehicle as a single rigid body with shared pose nodes is a cheap way to stabilize long trajectories, a trick that may generalize beyond this system.

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 presents cuSfM, a CUDA-accelerated Structure-from-Motion system that, unlike traditional SfM, requires initial pose estimates and time-continuous image sequences as input (Section 2). Its pipeline consists of GPU feature extraction and matching (ALIKED, SuperPoint, SIFT, LightGlue), bag-of-words loop detection, pose-graph-based non-redundant view-graph construction, iterative triangulation with bundle adjustment, and optional extrinsic refinement. The authors report speed and accuracy comparisons against COLMAP and GLOMAP on KITTI and a simulated SDG dataset, and they release an open-source wrapper, PyCuSfM. The paper positions cuSfM as a pose-optimization and mapping backend for trajectory refinement, prior-map localization, and multi-camera calibration.

Significance. If the headline claims were fully supported, cuSfM would be a valuable contribution: it demonstrates a modular, GPU-accelerated SfM pipeline with open-source code and a public dataset, and its trajectory-refinement experiments (Tables 4 and 5) show consistent ATE reductions relative to PyCuVSLAM and ORB-SLAM2 initializations. However, the paper's central claims — that cuSfM achieves 'significantly improved accuracy and processing speed' and 'an order-of-magnitude runtime improvement' over COLMAP — are not supported by the presented evidence. The system is honestly a trajectory-refinement backend, not a drop-in SfM replacement, and the comparison to COLMAP/GLOMAP is asymmetric because cuSfM receives pose priors that the baselines do not. The credible portion is the refinement capability; the overreach is in the head-to-head framing. The open-source release and modular architecture are strengths, but the claims need to be either re-run with equal priors or substantially reframed.

major comments (4)
  1. [Abstract & Section 1 vs. Table 2] The abstract and introduction claim an order-of-magnitude runtime improvement over COLMAP, but the paper's own Table 2 shows total times of 58.824 s vs. 346.162 s per 100 frames — a 5.9× speedup, not 10×. Moreover, the feature-matching stage is ~10× slower than COLMAP (37.370 s vs. 3.690 s in view-graph mode; 76.600 s in radius mode). The 20× mapping-phase reduction (340.467 s → 16.458 s) is a stage-level result that does not translate to the claimed overall order-of-magnitude speedup. Please correct the text or present end-to-end runtime results that substantiate the claim.
  2. [Section 6.3.1, Table 3] The accuracy comparison against COLMAP and GLOMAP is asymmetric. Section 2 states that cuSfM 'requires initial poses,' and Section 6.3.1 initializes cuSfM from ground-truth poses corrupted by Gaussian noise, while COLMAP and GLOMAP receive only raw monocular images. This makes Section 6.3.1 a trajectory-refinement study, not a fair head-to-head SfM benchmark. The claim of 'significantly improved accuracy' is also not uniform: in Sequence 05, CuSfM's RMSE is 0.907 m, worse than COLMAP's 0.839 m and GLOMAP's 0.723 m. Please either rerun the comparison with the same priors provided to all methods (e.g., initializing COLMAP/GLOMAP with the same noisy poses) or explicitly reframe the paper as a pose-refinement system.
  3. [Equations (11), (12), Section 3.4] There is a mild circularity in the accuracy evaluation. The same prior trajectories are used twice: first to construct the non-redundant view graph (Section 3.4) and second as absolute-pose constraints in the optimization loss, Eq. (12) with e_a from Eq. (11). The reported improvements over the input trajectories are therefore partly a consequence of pulling the output toward the input. To demonstrate that the refinement is not trivially caused by these constraints, please include ablations with the absolute-pose term removed, or vary the strength λ_a and show how the ATE changes. Without such analysis, the refinement gains over PyCuVSLAM/ORB-SLAM2 are confounded by the prior serving as the optimization target.
  4. [Section 3.4, Table 2] The view-graph strategy is a key contribution, but the speed comparison in Table 2 conflates it with the feature-matching bottleneck. The paper reports that view-graph mode reduces matching time to 37.370 s from COLMAP's 3.690 s, yet the 'order-of-magnitude' claim is based on mapping time alone. Please provide a breakdown of the view-graph construction overhead (BoW dictionary, search tree) and a direct comparison with COLMAP's sequential matcher under identical input conditions, so that the actual end-to-end advantage is transparent.
minor comments (5)
  1. [Section 6.1] Typo: 'GLMAP-GPU' should be 'GLOMAP' (and similarly in Figure 6 caption, 'COLAMP' should be 'COLMAP'). Please proofread.
  2. [Equation (1)] The notation in Eq. (1) is unclear: T_map_left and T_right_map are composed as T_right_left = T_map_left · T_right_map, but the subscripts/superscripts do not make the frame order unambiguous. Please clarify with explicit coordinate frames and check that the composition order is consistent with the definitions elsewhere.
  3. [Section 4.1] The stereo relative pose algorithm is presented as novel, but it closely follows the SOFT2 epipolar-line metric and three-view scale recovery (Cvišić et al.). Please clarify what is new beyond the cited SOFT2 approach and add a quantitative comparison to SOFT2 or an ablation that isolates this module's contribution.
  4. [Section 6.2] The paper says all methods use Ceres with SPARSE_NORMAL_CHOLESKY, but COLMAP and GLOMAP have their own internal solvers; please clarify whether this applies only to cuSfM or to all methods, and whether the GLOMAP results use the same BA settings. This affects the interpretation of the mapping-time comparison.
  5. [Section 6.3.1] In Table 3, the reported min/max values for Sequence 05 show large asymmetry (min 0.075 m vs max 3.324 m). Please discuss this outlier case; the current text ('superior performance across most sequences') obscures the fact that cuSfM is worse than both baselines on this sequence.

Circularity Check

0 steps flagged

No circular derivation found: cuSfM's outputs are not equal to its inputs by construction; the main concerns are an asymmetric COLMAP comparison and a minor non-load-bearing self-citation.

full rationale

The paper is a system paper whose central operation is trajectory refinement from given initial poses. It does not claim to derive a pose from that pose by equation: Section 4.3's bundle adjustment combines reprojection errors with initial-pose priors (Eqs. 9-12), so the output is constrained by, but not identical to, its input. The trajectory-refinement comparisons in Tables 4 and 5 are legitimate before/after evaluations against the same initializers (PyCuVSLAM and ORB-SLAM2). One self-citation, [22] (Yu et al. 2022, overlapping first author), is used in Section 3.4 for the 'complete yet non-redundant information sets' principle underlying the view-graph construction; however, the same sentence independently invokes ORB-SLAM2's Essential Graph, and the paper empirically validates view-graph vs radius retrieval (Table 4), so this self-citation is not the sole load-bearing support and is not a circular step. The headline claim of 'significantly improved accuracy and processing speed compared to COLMAP' is not a circularity but an asymmetric-evaluation concern: cuSfM receives initial poses (Section 2; Section 6.3.1: 'cuSfM was initialized with ground truth poses corrupted by Gaussian noise') while COLMAP and GLOMAP receive only monocular images, and Table 2 shows a total speedup of 5.9x (346.162s vs 58.824s), not the claimed order-of-magnitude. These are overclaims affecting supportability, not reductions of claimed results to their inputs. No step in the derivation chain has been shown to be equivalent to its input by construction, so the circularity finding is 'no significant circularity' apart from the minor self-citation noted above.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical or mathematical entities are introduced; the system uses established components (features, pose graphs, BA). The central claim rests on domain assumptions about availability of reliable initial trajectories and calibrated stereo extrinsics, plus several unreported numerical hyperparameters. No parameter fitting to the evaluation targets is evident.

free parameters (4)
  • Radius-search thresholds (20 m, 90 degrees) = 20 m, 90 degrees
    Hand-set thresholds for candidate pair selection in KITTI experiments; the view-graph vs radius comparison in Table 4 depends on them, and no sensitivity analysis is given.
  • BA loss weights lambda_c, lambda_a = Not reported
    Eq. 12 combines reprojection error with temporal and absolute pose constraints weighted by lambda_c and lambda_a; the values are unspecified, so the objective's actual balance is unknown.
  • Two-stage BA outlier thresholds and robust loss parameters = Not reported
    Section 4.3 says stage 1 uses relaxed thresholds and robust loss functions and stage 2 uses stricter criteria, but exact thresholds and loss choices are not given.
  • BoW vocabulary tree parameters (K, depth) = Not reported
    Section 3.2 builds a hierarchical vocabulary with recursive K-means, but cluster count K and tree depth are unspecified, affecting loop closure recall.
axioms (4)
  • domain assumption Input trajectories are reliable enough to define the pose graph and candidate matching pairs
    Section 2 states the system 'requires initial poses, time-continuous image sequences'; Section 3.4 uses pose graph priors to construct the view graph. If the priors are bad, non-redundant matching may omit correct edges.
  • domain assumption Calibrated stereo baseline is available and accurate for scale recovery
    Section 4.1 recovers translation scale by composing relative poses and using known left-right calibration in Eq. 1; this fails without calibrated stereo or known extrinsics.
  • standard math Epipolar geometry, pose graph optimization, and bundle adjustment work as assumed
    Used throughout Sections 3.3, 4.1-4.3; no proofs are provided, but these are standard results in the field.
  • domain assumption ALIKED+LightGlue performance on a two-view relative pose benchmark transfers to full SfM
    Section 3.1 selects components based on the Roma two-view benchmark AUC; no evidence is provided that this choice is optimal for large-scale multi-view bundle adjustment.

reviewed 2026-08-04 · how reviews work

0 comments
read the original abstract

Efficient and accurate camera pose estimation forms the foundational requirement for dense reconstruction in autonomous navigation, robotic perception, and virtual simulation systems. This paper addresses the challenge via cuSfM, a CUDA-accelerated offline Structure-from-Motion system that leverages GPU parallelization to efficiently employ computationally intensive yet highly accurate feature extractors, generating comprehensive and non-redundant data associations for precise camera pose estimation and globally consistent mapping. The system supports pose optimization, mapping, prior-map localization, and extrinsic refinement. It is designed for offline processing, where computational resources can be fully utilized to maximize accuracy. Experimental results demonstrate that cuSfM achieves significantly improved accuracy and processing speed compared to the widely used COLMAP method across various testing scenarios, while maintaining the high precision and global consistency essential for offline SfM applications. The system is released as an open-source Python wrapper implementation, PyCuSfM, available at https://github.com/nvidia-isaac/pyCuSFM, to facilitate research and applications in computer vision and robotics.

Figures

Figures reproduced from arXiv: 2510.15271 by Chirag Majithia, Di Zeng, Jingrui Yu, Joydeep Biswas, Jun Liu, Kefei Ren, Keqiang Wu, Rurui Ye.

Figure 1
Figure 1. Figure 1: High-quality 3D reconstruction achieved by cuSfM: A dense point cloud of an indoor meeting room [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System architecture of cuSfM. The pipeline flows from left to right, comprising system inputs, non [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the vocabulary tree construction and loop detection process. The left green panel shows [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the redundancy issue in feature matching. When Frame A matches with Frame B, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the pose graph structure. Camera poses are represented as triangular nodes, with [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trajectory comparison on KITTI 00 sequence. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of map integration capabilities in cuSfM. Left: New trajectory data. Middle: Previous map. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 3 linked inside Pith

  1. [1]

    Scene coordinate reconstruction: Posing of image collections via incremental learning of a relocalizer

    Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Áron Monszpart, Daniyar Turmukham- betov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Posing of image collections via incremental learning of a relocalizer. InProceedings of the European Conference on Computer Vision (ECCV), pages 421–440. Springer, 2024. 1

    2024

  2. [2]

    Brief: Binary robust independent elementary features

    Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. InProceedings of the European Conference on Computer Vision (ECCV), pages 778–792. Springer, 2010. 4

    2010

  3. [3]

    Soft2: Stereo visual odometry for road vehicles based on a point-to-epipolar-line metric.IEEE Transactions on Robotics, 39(1):273–288, 2023

    Igor Cvišić, Ivan Marković, and Ivan Petrović. Soft2: Stereo visual odometry for road vehicles based on a point-to-epipolar-line metric.IEEE Transactions on Robotics, 39(1):273–288, 2023. 10

    2023

  4. [4]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops ( CVPRW), pages 337–33712. IEEE/CVF, 2018. 3

    2018

  5. [5]

    Mast3r-sfm: A fully-integrated solution for unconstrained structure-from-motion

    Bardienus Pieter Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: A fully-integrated solution for unconstrained structure-from-motion. InProceedings of the International Conference on 3D Vision, pages 1–10. IEEE, 2025. 1

    2025

  6. [6]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR), pages 19790–19800. IEEE/CVF, 2024. 4

    2024

  7. [8]

    Are we ready for autonomous driving? the KITTI vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition ( CVPR), pages 3354–3361. IEEE, 2012. 15

    2012

  8. [9]

    evo: Python package for the evaluation of odometry and slam.https://github.com/ MichaelGrupp/evo, 2017

    Michael Grupp. evo: Python package for the evaluation of odometry and slam.https://github.com/ MichaelGrupp/evo, 2017. 16

    2017

  9. [10]

    Cambridge University Press, New York, NY, USA, 2 edition, 2003

    Richard Hartley and Andrew Zisserman.Multiple View Geometry in Computer Vision. Cambridge University Press, New York, NY, USA, 2 edition, 2003. 9

    2003

  10. [11]

    cuvslam: Cuda accelerated visual odometry.arXiv, 2506.04359,

    Alexander Korovko, Dmitry Slepichev, Alexander Efitorov, Aigul Dzhumamuratova, Viktor Kuznetsov, Hesam Rabeti, and Joydeep Biswas. cuvslam: Cuda accelerated visual odometry.arXiv, 2506.04359,

    Pith/arXiv arXiv

  11. [12]

    Triangulation: why optimize?arXiv, 1907.11917, 2019

    Seong Hun Lee and Javier Civera. Triangulation: why optimize?arXiv, 1907.11917, 2019. 12

    Pith/arXiv arXiv 1907

  12. [13]

    Grounding image matching in 3d with mast3r.arXiv, 2501.14914, 2025

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r.arXiv, 2501.14914, 2025. 4

    Pith/arXiv arXiv 2025

  13. [14]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17581–17592. IEEE/CVF, 2023. 3

    2023

  14. [15]

    Multi-session slam with differentiable wide-baseline pose optimization

    Lahav Lipson and Jia Deng. Multi-session slam with differentiable wide-baseline pose optimization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR), pages 19626–19635. IEEE/CVF, 2024. 4 23 CuSfM: CUDA-Accelerated Structure-from-Motion

    2024

  15. [16]

    Distinctive image features from scale-invariant keypoints.International Journal of Computer Vision, 60(2):91–110, 2004

    David G Lowe. Distinctive image features from scale-invariant keypoints.International Journal of Computer Vision, 60(2):91–110, 2004. 4

    2004

  16. [17]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE Transactions on Robotics, 33(5):1255–1262, 2017

    Raúl Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE Transactions on Robotics, 33(5):1255–1262, 2017. 4, 8, 15

    2017

  17. [18]

    Global structure-from-motion revisited

    Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schönberger. Global structure-from-motion revisited. InProceedings of the European Conference on Computer Vision (ECCV), pages 58–77. Springer,

  18. [19]

    Nascimento

    Guilherme Potje, Felipe Cadar, André Araujo, Renato Martins, and Erickson R. Nascimento. Xfeat: Accelerated features for lightweight image matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR), pages 2682–2691. IEEE/CVF, 2024. 4

    2024

  19. [20]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR), pages 4104–4113. IEEE/CVF,

  20. [21]

    Disk: Learning local features with policy gradient

    Michał J Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. InProceedings of the International Conference on Neural Information Processing Systems (NIPS), page 12. Curran Associates Inc., 2020. 4

    2020

  21. [22]

    Hierarchical multi-level information fusion for robust and consistent visual slam.IEEE Transactions on Vehicular Technology, 71(1):250–259, 2022

    Jingrui Yu, ZhenZhen Xiang, and Jianbo Su. Hierarchical multi-level information fusion for robust and consistent visual slam.IEEE Transactions on Vehicular Technology, 71(1):250–259, 2022. 8

    2022

  22. [23]

    Birch: an efficient data clustering method for very large databases.ACM SIGMOD Record, 25(2):103–114, 1996

    Tian Zhang, Raghu Ramakrishnan, and Miron Livny. Birch: an efficient data clustering method for very large databases.ACM SIGMOD Record, 25(2):103–114, 1996. 6

    1996

  23. [24]

    Xiaoming Zhao, Xingming Wu, Weihai Chen, Peter C. Y. Chen, Qingsong Xu, and Zhengguo Li. Aliked: A lighter keypoint and descriptor extraction network via deformable transformation.IEEE Transactions on Instrumentation & Measurement, 72:1–16, 2023. 3 24

    2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.