Pith. sign in

REVIEW 2 major objections 6 minor 45 references

A lightweight plug-in can refine the descriptors already used by real-time visual SLAM systems, raising matching accuracy and cutting trajectory error without rewriting the front end.

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 · grok-4.5

2026-07-14 07:03 UTC pith:5I2FBNR2

load-bearing objection Practical plug-and-play enhancer that beats FeatureBooster and lifts four real SLAM systems at ~5 ms cost; hybrid Mamba-AFT is the real piece of novelty. the 2 major comments →

arxiv 2607.11099 v1 pith:5I2FBNR2 submitted 2026-07-13 cs.RO cs.CV

Desc++: Efficient Descriptor Enhancement for Data Association in Existing Visual SLAM Systems

Ting-Wei Ou , Huang-Ting Lin , Kuu-Young Young This is my paper
classification cs.RO cs.CV
keywords visual SLAMdescriptor enhancementdata associationMambaattention-free transformerreal-time roboticsfeature matching
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.

Most deployed visual SLAM systems still match handcrafted local descriptors. Those descriptors fail under large viewpoint or lighting changes, yet swapping them for a full learned front end is expensive and breaks existing pipelines. This paper claims that a small enhancement network can keep the original detector and matching interface while still producing more discriminative descriptors. The network fuses each descriptor with its keypoint geometry, then mixes order-agnostic global context with geometry-aware sequential modeling along a Z-order keypoint sequence, all in linear time. The same module is dropped, unchanged, into four different SLAM systems and improves trajectory accuracy on most sequences of EuRoC, KITTI and Hilti while adding only a few milliseconds and a few hundred megabytes of GPU memory. The practical point is that better data association can be bought without redesigning the systems already running on robots.

Core claim

Desc++ shows that a hybrid linear-time context aggregator—pairing attention-free global attention with a Mamba branch over Z-order-serialized keypoints—can refine existing descriptors inside their original format so that matching accuracy exceeds the prior enhancement method and the same module measurably lowers absolute trajectory error when inserted into four heterogeneous real-time V-SLAM systems.

What carries the argument

The Mamba-AFTs block: a parallel hybrid that fuses order-agnostic global context from an attention-free transformer with geometry-aware sequential state-space modeling along a Z-order keypoint sequence, gated per keypoint and running in linear time.

Load-bearing premise

The original feature extractor must still produce a usable set of keypoints; if the detector fails, no amount of later refinement can recover the missing information.

What would settle it

Insert Desc++ into the same four SLAM systems on the same EuRoC, KITTI and Hilti sequences; if matching accuracy on HPatches and absolute trajectory error do not improve relative to both the unenhanced baselines and FeatureBooster on the majority of sequences, the central claim fails.

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

If this is right

  • Deployed stereo, stereo-inertial, visual-LiDAR and multi-camera V-SLAM systems can raise localization accuracy by swapping only the descriptor vectors, without changing tracking or mapping code.
  • The largest relative gains appear under large cross-camera viewpoint changes, so multi-camera industrial and warehouse setups stand to benefit most.
  • Because the enhanced descriptors keep their original dimension and metric, existing Hamming or Euclidean matchers continue to work without modification.
  • The added latency stays near 5 ms and GPU memory near 0.5 GB, keeping real-time rates on both desktop and embedded Jetson-class hardware.

Where Pith is reading between the lines

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

  • A natural next step is a joint enhancement-and-confidence head that lets the SLAM back-end down-weight unreliable associations rather than treating every refined descriptor equally.
  • The same linear hybrid block may transfer to other sparse correspondence tasks that already own a detector but need better descriptors under viewpoint change.
  • Because the method is bounded by extraction quality, pairing it with a more resilient detector under motion blur or extreme texture loss would expand its operating regime without abandoning the plug-and-play design.

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

2 major / 6 minor

Summary. The paper proposes Desc++, a lightweight plug-and-play descriptor enhancement module for existing feature-based V-SLAM systems. It fuses original descriptors with keypoint geometry (via MLPs and Learnable Fourier Features), then aggregates context with a parallel hybrid Mamba–AFTs block: an order-agnostic AFT-Simple branch for global context and a Mamba branch over Z-order-serialized keypoints for geometry-aware sequential modeling, fused by per-keypoint gating, all in linear time. Enhanced descriptors retain original dimensionality and matching interface. Training uses a FastAP-based composite loss on MegaDepth. Evaluation spans HPatches MMA (vs original descriptors and FeatureBooster), Tracked MapPoints analysis, and system-level ATE RMSE (mean±std over 5 runs) when the same module is inserted without pipeline changes into ORB-SLAM2, ORB-SLAM3, RGB-L, and MAVIS-SLAM on EuRoC, KITTI, and Hilti 2023, plus a cost comparison against a SuperPoint+LightGlue front-end (Rover-SLAM).

Significance. If the results hold, Desc++ is a practically useful contribution: it improves data association in mature real-time V-SLAM pipelines without replacing detectors, matchers, or tracking logic, at modest overhead (~5 ms, ~544 MiB) versus full learned front-ends. Multi-system, multi-benchmark validation with reported variance, architecture ablations (Table VII–IX), open code/weights, and explicit discussion of the extractor-failure boundary (§IV-G) are genuine strengths. The work sits in a useful middle ground between handcrafted front-ends and replacement-based learned systems, and is of clear interest to the robotics/SLAM community.

major comments (2)
  1. Tables III–V (system ATE): The central claim that gains “translate into more accurate and stable trajectory estimation” rests on majority-positive sequences, but several reported improvements (and the two small regressions on EuRoC MH05/V101 and the KITTI-03 regression) lie within or near one standard deviation of the baseline. The paper should either (a) apply a simple paired significance test or report win/tie/loss counts under a clear margin (e.g., mean improvement > max(std)), or (b) explicitly qualify “more accurate” as “lower mean ATE on a majority of sequences, with reduced variance on hard open-loop runs.” Without this, the system-level claim is slightly overstated relative to the reported uncertainty.
  2. §III-E, Eq. (8) and training protocol: The composite loss includes a booster term Lboost that encourages improvement over the original descriptor AP. There is no ablation of Lmatch alone vs Lmatch+Lboost, nor of the 3 px / 16 px positive/negative radii. Because the HPatches gains over FeatureBooster (Fig. 3) and the TMP gains (Table II) are the main evidence that the architecture (not just the loss) is responsible, a short loss-component ablation (or a statement that Lboost was ablated and found necessary/neutral) would make the training design load-bearing rather than inherited from prior work.
minor comments (6)
  1. Fig. 3 table: “RG” (relative gain at 5 px) is defined in the caption but not in the main text; a one-line definition would help readers scanning the figure alone.
  2. Table I: ALIKE uses 6 CA layers “to maintain model capacity comparable to FeatureBooster.” State the parameter counts of Desc++ vs FeatureBooster for ALIKE explicitly so the fairness claim is checkable.
  3. §III-B / §IV-F: Z-order is motivated and ablated (Table IX), but the modest MMA lift (0.453→0.459 @3) could be briefly related to typical keypoint counts in SLAM (1k–2k) so readers know when ordering matters.
  4. Table VI: Latency is measured on a desktop RTX 4090; the Jetson Orin NX numbers in Table VII are module-only. A single end-to-end FPS number for ORB-SLAM3+Desc++ on the Jetson (even if approximate) would strengthen the embedded-deployment claim in §IV-G.
  5. Typos / notation: “MA VIS-SLAM” / “MAVIS-SLAM” spacing is inconsistent; “EuRoC MA V” has a stray space; Eq. (5) uses YAFT,i while the surrounding text uses Y_AFT—pick one style.
  6. §IV-C: The observation that TMP gains concentrate on Hilti side cameras (large viewpoint change) while Camera-L is flat or slightly down is insightful; a short sentence linking this to the HPatches viewpoint subset would tighten the narrative.

Circularity Check

0 steps flagged

No significant circularity: purely empirical architecture + training on MegaDepth, evaluated on independent public benchmarks with no fitted quantity re-presented as a prediction.

full rationale

Desc++ is an engineering paper whose central claims are empirical (higher MMA than FeatureBooster on HPatches; higher TMP; lower ATE RMSE when the same module is dropped into four heterogeneous V-SLAM systems). The training objective (FastAP + booster term on MegaDepth correspondences) is independent of the evaluation sets (HPatches, EuRoC, KITTI, Hilti). No parameter is fitted to a subset of the reported metrics and then re-labeled a prediction; no uniqueness theorem or ansatz is imported via self-citation; the hybrid Mamba–AFT design is presented as an architectural choice, not a first-principles derivation. Self-citations are limited to standard SLAM baselines (ORB-SLAM2/3, etc.) that are externally available and not load-bearing for the novelty claim. The acknowledged boundary case (extractor failure on EuRoC V203) is stated openly rather than hidden by circular reasoning. Consequently the derivation chain contains no step that reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The paper is an empirical systems contribution. Its claims rest on standard deep-learning and SLAM assumptions plus a small set of architectural and training choices that function as free parameters. No new physical entities are postulated.

free parameters (4)
  • Number of Context Aggregation layers = 2 or 6
    Set to 2 for ORB/SIFT/SuperPoint and 6 for ALIKE to match FeatureBooster capacity (Table I); chosen by hand for fairness rather than derived.
  • LFF frequency groups and latent dimension = [64,32,32] or [64,64] → 128-D
    Geometry attributes partitioned into groups projected to 128-D via Learnable Fourier Features; group sizes listed in Table I are design choices.
  • Training hyperparameters (lr, epochs, batch, warmup) = lr=1e-3, 50 epochs, bs=16
    AdamW, cosine schedule from 1e-3 after 500-step warmup, 50 epochs, batch 16; standard but free choices that affect final weights.
  • Positive/negative pixel radii for FastAP supervision = 3 px / 16 px
    3 px positives, 16 px negatives, intermediate excluded; conventional but arbitrary thresholds that define the training signal.
axioms (4)
  • domain assumption Z-order (Morton) serialization of 2-D keypoints sufficiently preserves local geometric neighborhood structure for a causal SSM to exploit.
    Invoked in Section III-B and validated only by the modest ablation gain in Table IX; not proven in general.
  • domain assumption AFT-Simple provides a sufficiently expressive order-agnostic global context in linear time to complement the Mamba branch.
    Taken from the AFT paper and used as a black-box component (Eq. 5); performance relies on this property holding for sparse keypoint sets.
  • domain assumption Nearest-neighbor matching with the original distance metric (Hamming or Euclidean) remains the correct interface after enhancement.
    Core plug-and-play premise stated in the problem formulation (Eq. 1) and throughout the system experiments.
  • domain assumption Standard SLAM back-ends (pose-graph optimization, RANSAC, etc.) will convert improved correspondences into lower trajectory error.
    Implicit in the claim that higher TMP and MMA translate to better ATE; observed empirically but not guaranteed for every sequence.
invented entities (1)
  • Mamba-AFTs Block (parallel gated hybrid of Mamba and AFT-Simple) no independent evidence
    purpose: Aggregate both sequential geometric and order-agnostic global context among keypoints in linear time.
    The specific parallel architecture with content-dependent gating (Eq. 7) is introduced by the paper; independent evidence is limited to the ablations and downstream SLAM gains reported here.

pith-pipeline@v1.1.0-grok45 · 22936 in / 2789 out tokens · 32817 ms · 2026-07-14T07:03:59.099808+00:00 · methodology

0 comments
read the original abstract

Reliable visual data association is fundamental to visual SLAM (V-SLAM), as it directly determines the quality of the camera pose estimation and map consistency. However, the handcrafted descriptors used by most mature real-time systems degrade under illumination and viewpoint changes, while learning-based front-ends that address this weakness typically require replacing the extraction-and-matching pipeline and introduce substantial computational overhead. Descriptor enhancement offers a compromise by refining existing descriptors within their original format, yet current methods rely on simplified attention mechanisms whose limited contextual modeling constrains the achievable matching quality. To resolve this trade-off between contextual expressiveness and efficiency, we propose Desc++, a lightweight enhancement module that jointly encodes descriptor representations and keypoint geometry and aggregates spatial context through a hybrid architecture that combines order-agnostic global attention with geometry-aware sequential modeling in linear time. The enhanced descriptors retain their original dimensionality and matching interface, enabling integration into deployed V-SLAM systems without modifying the pipeline. Experiments across descriptor matching, correspondence analysis, and system-level benchmarks with four different V-SLAM systems demonstrate that Desc++ improves matching accuracy over the state-of-the-art enhancement method, translates these gains into more accurate and stable trajectory estimation, and achieves a favorable balance between accuracy and efficiency for practical integration into existing real-time V-SLAM pipelines.

Figures

Figures reproduced from arXiv: 2607.11099 by Huang-Ting Lin, Kuu-Young Young, Ting-Wei Ou.

Figure 1
Figure 1. Figure 1: The overall pipeline of Desc++. First, unstructured keypoints extracted from an image are spatially organized via z-ordering. In the Local Fusion Stage, the geometry and original descriptor of each keypoint are fused to enrich the local representation with geometric priors. These features are then processed by the Context Aggregation Stage, where the hybrid Mamba–AFTs Block performs relation-aware modeling… view at source ↗
Figure 2
Figure 2. Figure 2: Detailed diagrams of (a) AFT-Simple [16] vs. (b) Mamba–AFTs [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MMA curves and quantitative results on the HPatches dataset. We compare Desc++ against the original descriptors and FeatureBooster. The table [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of mutual nearest neighbor matching under illumination and viewpoint variations. Green and red lines denote correct (dist. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative feature correspondences on the Hilti sequences. Desc++ preserves more valid feature-to-map associations under large cross-camera [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Robot configuration (left) and qualitative tracking comparisons in a textureless corridor (middle) and a room with varying illumination (right). [PITH_FULL_IMAGE:figures/full_fig_p011_6.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

45 extracted references · 3 linked inside Pith

  1. [1]

    Rublee, V

    E. Rublee, V . Rabaud, K. Konolige, and G. Bradski, ”ORB: An efficient alternative to SIFT or SURF,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., pp. 2564–2571, 2011

    2011

  2. [2]

    Mur-Artal and J

    R. Mur-Artal and J. D. Tard ´os, ”ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D Cameras,” IEEE Trans. Robot., vol. 33, no. 5, pp. 1255–1262, 2017

    2017

  3. [3]

    Campos, R

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. M. Montiel, and J. D. Tard´os, ”ORB-SLAM3: An accurate open-source library for visual, visual–inertial, and multimap SLAM,” IEEE Trans. Robot. , vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  4. [4]

    Sauerbeck, B

    F. Sauerbeck, B. Obermeier, M. Rudolph and J. Betz, ”RGB-L: Enhancing indirect visual SLAM using LiDAR-based dense depth maps,” in Proc. Int. Conf. Comput., Control Robot., pp. 95–100, 2023

    2023

  5. [5]

    Mikolajczyk and C

    K. Mikolajczyk and C. Schmid, ”A performance evaluation of local descriptors,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 27, no. 10, pp. 1615–1630, 2005

    2005

  6. [6]

    J. L. Sch ¨onberger, E. Zheng, M. Pollefeys, and J. -M. Frahm, ”Pixel- wise view selection for unstructured multi-view stereo,” in Proc. Eur. Conf. Comput. Vis., pp. 501–518, 2016

    2016

  7. [7]

    Wang, et al., ”FeatureBooster: Boosting feature descriptors with a lightweight neural network,” in Proc

    X. Wang, et al., ”FeatureBooster: Boosting feature descriptors with a lightweight neural network,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 7630–7639, 2023

    2023

  8. [8]

    DeTone, T

    D. DeTone, T. Malisiewicz, and A. Rabinovich, ”Superpoint: Self-supervised interest point detection and description,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops, pp. 224– 236, 2018

    2018

  9. [9]

    Zhao, et al., ”ALIKE: Accurate and lightweight keypoint detection and descriptor extraction,” IEEE Trans

    X. Zhao, et al., ”ALIKE: Accurate and lightweight keypoint detection and descriptor extraction,” IEEE Trans. Multimedia, vol. 25, pp. 3101– 3112, 2022

    2022

  10. [10]

    Tyszkiewicz, P

    M. Tyszkiewicz, P. Fua, and E. Trulls, ”DISK: Learning local features with policy gradient,” in Proc. Adv. Neural Inf. Process. Syst., vol. 33, pp. 14254–14265, 2020

    2020

  11. [11]

    Luo et al., ”ContextDesc: Local descriptor augmentation with cross- modality context,” in Proc

    Z. Luo et al., ”ContextDesc: Local descriptor augmentation with cross- modality context,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 2527–2536, 2019

    2019

  12. [12]

    Sarlin et al., ”SuperGlue: Learning feature matching with graph neural networks,” in Proc

    P.-E. Sarlin et al., ”SuperGlue: Learning feature matching with graph neural networks,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 4938–4947, 2020

    2020

  13. [13]

    Lindenberger, P.-E

    P. Lindenberger, P.-E. Sarlin, and M. Pollefeys, ”LightGlue: Local feature matching at light speed,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., pp. 17627–17638, 2023

    2023

  14. [14]

    K. Ryoo, H. Lim and H. Myung, ”MambaGlue: Fast and Robust Local Feature Matching with Mamba,” in Proc. IEEE Int. Conf. Robot. Autom., pp. 5758–5765, 2025

    2025

  15. [15]

    Vaswani, et al

    A. Vaswani, et al. , ”Attention is all you need,” in Proc. Adv. Neural Inf. Process. Syst. , vol. 30, pp. 6000–6010, 2017

    2017

  16. [16]

    Zhai et al

    S. Zhai et al. , ”An attention-free transformer,” arXiv preprint arXiv:2105.14103, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    Gu, and T

    A. Gu, and T. Dao, ”Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Zhu et al., ”Vision Mamba: Efficient visual representation learning with bidirectional state space model,” in Proc

    L. Zhu et al., ”Vision Mamba: Efficient visual representation learning with bidirectional state space model,” in Proc. Int. Conf. Mach. Learn., pp. 62429–62442, 2024

    2024

  19. [19]

    Hatamizadeh and J

    A. Hatamizadeh and J. Kautz, ”MambaVision: A hybrid mamba- transformer vision backbone,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 25261–25270, 2025

    2025

  20. [20]

    Li et al., ”Learnable Fourier features for multi-dimensional spatial positional encoding,” in Proc

    Y . Li et al., ”Learnable Fourier features for multi-dimensional spatial positional encoding,” in Proc. Adv. Neural Inf. Process. Syst. , vol. 34, pp. 15816–15829, 2021

    2021

  21. [21]

    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, pp. 381–395, 1981

    1981

  22. [22]

    D. G. Lowe, ”Distinctive image features from scale-invariant key- points,” Int. J. Comput. Vis. , vol. 60, no. 2, pp. 91–110, 2004

    2004

  23. [23]

    Calonder, V

    M. Calonder, V . Lepetit, C. Strecha, and P. Fua, ”BRIEF: Binary robust independent elementary features,” in Proc. Eur. Conf. Comput. Vis. , vol. 6314, pp. 778–792, 2010

    2010

  24. [24]

    Rosten and T

    E. Rosten and T. Drummond, ”Machine learning for high-speed corner detection,” in Proc. Eur. Conf. Comput. Vis. , vol. 3951, pp. 430–443, 2006

    2006

  25. [25]

    Carion, et al., ”End-to-end object detection with Transformers,” in Proc

    N. Carion, et al., ”End-to-end object detection with Transformers,” in Proc. Eur. Conf. Comput. Vis. , vol. 12346, pp. 213–229, 2020

    2020

  26. [26]

    Dosovitskiy, et al., ”An image is worth 16×16 words: Transformers for image recognition at scale,” in Proc

    A. Dosovitskiy, et al., ”An image is worth 16×16 words: Transformers for image recognition at scale,” in Proc. Int. Conf. Learn. Represent. , 2021

    2021

  27. [27]

    Liu et al

    Y . Liu et al. , ”VMamba: Visual state space model,” in Proc. Adv. Neural Inf. Process. Syst. , vol. 37, pp. 103031–103063, 2024

    2024

  28. [28]

    G. M. Morton, ”A computer oriented geodetic data base; and a new technique in file sequencing,” Tech. Rep., IBM, 1966

    1966

  29. [29]

    Cakir, K

    F. Cakir, K. He, X. Xia, B. Kulis, and S. Sclaroff, ”Deep metric learning to rank,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 1861–1870, 2019

    2019

  30. [30]

    Li and N

    Z. Li and N. Snavely, ”MegaDepth: Learning single-view depth prediction from internet photos,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 2041–2050, 2018

    2041

  31. [31]

    Balntas, et al

    V . Balntas, et al. , ”HPatches: A benchmark and evaluation of hand- crafted and learned local descriptors,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , pp. 5173–5182, 2017

    2017

  32. [32]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, ”Vision meets robotics: The KITTI dataset,”Int. J. Robot. Res., vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  33. [33]

    Burri et al

    M. Burri et al. , ”The EuRoC micro aerial vehicle datasets,” Int. J. Robot. Res., vol. 35, no. 10, pp. 1157–1163, 2016

    2016

  34. [34]

    Grupp, ”evo: Python package for the evaluation of odom- etry and SLAM,” 2017

    M. Grupp, ”evo: Python package for the evaluation of odom- etry and SLAM,” 2017. Available: https://github.com/ MichaelGrupp/evo

    2017

  35. [35]

    Dusmanu et al

    M. Dusmanu et al. , ”D2-Net: A trainable CNN for joint description and detection of local features,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 8092–8101, 2019

    2019

  36. [36]

    Philbin, M

    J. Philbin, M. Isard, J. Sivic, and A. Zisserman, ”Descriptor learning for efficient retrieval,” in Proc. Eur. Conf. Comput. Vis., vol. 6316, pp. 677–691, 2010

    2010

  37. [37]

    J. L. Sch ¨onberger and J. -M. Frahm, ”Structure-from-motion revisited,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , pp. 4104– 4113, 2016

    2016

  38. [38]

    Legittimo, F

    M. Legittimo, F. Crocetti, M. L. Fravolini, G. Mollica and G. Costante, ”LF2SLAM: Learning-based Features For visual SLAM,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. , pp. 5648–5655, 2024

    2024

  39. [39]

    J. Tang, L. Ericson, J. Folkesson and P. Jensfelt, ”GCNv2: Efficient Correspondence Prediction for Real-Time SLAM,” IEEE Robot. Au- tom. Lett., vol. 4, no. 4, pp. 3505–3512, 2019

    2019

  40. [40]

    A. D. Nair, J. Kindle, P. Levchev, and D. Scaramuzza, ”Hilti SLAM Challenge 2023: Benchmarking Single + Multi-Session SLAM Across Sensor Constellations in Construction,” IEEE Robot. Autom. Lett., vol. 9, no. 8, pp. 7286–7293, 2024

    2023

  41. [41]

    Wang et al ., ”MA VIS: Multi-Camera Augmented Visual-Inertial SLAM using SE2(3) Based Exact IMU Pre-integration,” in Proc

    Y . Wang et al ., ”MA VIS: Multi-Camera Augmented Visual-Inertial SLAM using SE2(3) Based Exact IMU Pre-integration,” in Proc. IEEE Int. Conf. Robot. Autom. , pp. 1694–1700, 2024

    2024

  42. [42]

    Zhang et al ., ”A Real-Time, Robust, and Versatile Visual-SLAM Framework Based on Deep Learning Networks,” IEEE Trans

    X. Zhang et al ., ”A Real-Time, Robust, and Versatile Visual-SLAM Framework Based on Deep Learning Networks,” IEEE Trans. Instrum. Meas., vol. 74, pp. 1–13, 2025

    2025

  43. [43]

    Loshchilov and F

    I. Loshchilov and F. Hutter, ”Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  44. [44]

    J. Tang, J. Folkesson and P. Jensfelt, ”Geometric Correspondence Network for Camera Motion Estimation,” IEEE Robot. Autom. Lett. , vol. 3, no. 2, pp. 1010–1017, 2018

    2018

  45. [45]

    Cadena et al., ”Past, Present, and Future of Simultaneous Localiza- tion and Mapping: Toward the Robust Perception Age,” IEEE Trans

    C. Cadena et al., ”Past, Present, and Future of Simultaneous Localiza- tion and Mapping: Toward the Robust Perception Age,” IEEE Trans. Robot., vol. 32, no. 6, pp. 1309–1332, 2016

    2016