Pith. sign in

REVIEW 3 major objections 5 minor 54 references

CA-I2P: Channel-Adaptive Registration Network with Global Optimal Selection

T0 review · 3 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 CA-I2P claims state-of-the-art image-to-point cloud registration by suppressing mismatched feature channels and replacing top-k selection with global optimal transport.

desk verdict Test-set-tuned hyperparameters threaten the SOTA claim, but the modules are sensible and the paper deserves a major-revision review. read the letter →

arxiv 2506.21364 v1 pith:VN4UEHWW submitted 2025-06-26 cs.CV cs.AI

Zhixin Cheng , Jiacheng Deng , Xinjun Li , Xiaotian Yin , Bohao Liao , Baoqun Yin , Wenfei Yang , Tianzhu Zhang This is my paper
classification cs.CVcs.AI
keywords image-to-pointcloudregistrationdetection-freematchingchanneladaptiveadjustmentcross-modalfilteringoptimaltransportselectionSinkhorniterationsRGB-DScenesV27-Scenes
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

The paper targets image-to-point cloud registration, the task of finding the rigid transformation that aligns a 2D image with a 3D point cloud of the same scene. It argues that two failure modes dominate current detection-free methods: channel-level differences in how image and point-cloud features attend to content, and redundant matches caused by repetitive structures under top-k selection. To counter these, CA-I2P adds a Channel Adaptive Adjustment Module that enhances and then filters features along the channel dimension, and a Global Optimal Selection Module that replaces top-k with optimal transport. On the RGB-D Scenes V2 and 7-Scenes benchmarks the method reports registration recall of 63.3% and 79.5%, beating the previous best by 6.9 and 3.7 percentage points respectively.

What carries the argument

Two modules carry the argument. The Cross-Modal Channel Filtering Stage (CMCF) computes covariance matrices $\hat{V}_I^x$ and $\hat{V}_P^x$ for instance-normalized image and point-cloud features, derives a cross-modal covariance $\mathrm{Cov}^x$ that scores how sensitive each channel pair is to modality variation, and gates the largest-variance entries with a learned mask $M^x$; the filtered features are fused back with the originals to preserve information. The Global Optimal Selection Module (GOS) replaces top-k patch selection with an optimal transport plan $T^*$ minimizing $\mathrm{Tr}(T^\top(1-S)) - \epsilon H(T)$ under uniform marginals, solved by roughly 10 Sinkhorn iterations. The covariance mask is the load-bearing novelty, while the transport plan is the mechanism that removes redundant matches.

What would settle it

Train the same network but mask a random set of channels instead of the high-covariance ones, or remove the mask and keep only instance normalization; if registration recall does not drop, the covariance criterion is not the cause of the gain. A direct version is to compare the masked channel pairs against actual pixel-point overlap maps and check whether high-variance channels are genuinely concentrated on non-overlap regions.

Watch

Extended reading notes

Core claim

The central claim is that cross-modal matching in image-to-point-cloud registration improves when features are aligned at the channel level before matching and when patch-level correspondences are chosen by global optimization instead of per-patch top-k. The Channel Adaptive Adjustment Module first enhances each modality separately, using rotation-and-residual branch attention for images and channel self-attention for point clouds, then computes per-sample channel covariance matrices, measures how much each channel pair varies across modalities, and applies a learned mask that suppresses the high-variance components. The Global Optimal Selection Module then treats correspondence selection as an optimal transport problem solved with Sinkhorn iterations, eliminating many-to-one matches. The paper reports that the full model reaches 63.3% registration recall on RGB-D Scenes V2 and 79.5% on 7-Scenes, exceeding 2D3D-MATR by 6.9 and 3.7 percentage points respectively.

Load-bearing premise

The method assumes that channels whose statistics vary most between image and point cloud samples are harmful attention to non-overlapping regions and should be suppressed, but the paper does not directly show those channels are mismatches rather than useful signal.

Editorial extensions

If this is right

  • If the reported gains hold, channel-level alignment becomes a standard first step in detection-free image-to-point-cloud pipelines, with the largest benefits on scenes with repetitive geometry.
  • The optimal-transport selection step should transfer to any coarse-to-fine cross-modal matcher that currently uses top-k, removing many-to-one correspondences without retraining the backbone.
  • The small parameter increase, from 31.05M to 31.25M for the enhancement stages, indicates the accuracy gain comes from the matching strategy rather than added model capacity.
  • The improvement on the challenging Heads and Stairs scenes of 7-Scenes suggests the method is most effective where close-range depth error and repetitive patterns break earlier matchers.

Reading between the lines

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

  • An implicit testable extension is to apply the same channel-variance masking to outdoor image-to-point-cloud pairs, where the overlap region is much smaller; the paper argues indoor and outdoor methods do not transfer but does not run this experiment.
  • The covariance criterion predicts that channels with high cross-modal variance concentrate attention on non-overlapping or occluded regions; this prediction could be checked by visualizing per-channel attention maps, which the paper does not do.
  • The optimal transport formulation is generic enough to replace mutual nearest neighbour selection in other detection-free matchers, with the Sinkhorn temperature acting as a tunable precision-recall knob.
  • Because the paper evaluates only two indoor benchmarks, the generalization claim rests on the covariance statistic rather than on dataset diversity; a held-out indoor scene or a synthetic-to-real transfer test would bound the claim more tightly.
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 / 5 minor

Summary. The manuscript presents CA-I2P, a detection-free image-to-point cloud registration network built on the coarse-to-fine architecture of 2D3D-MATR. It contributes two modules: the Channel Adaptive Adjustment module, comprising an Intra-Modal Enhancement stage (ICE for images, PCE for point clouds) and a Cross-Modal Channel Filtering stage that masks high-variance channel pairs, and a Global Optimal Selection module that replaces top-k matching with Sinkhorn optimal transport. Experiments on RGB-D Scenes V2 and 7-Scenes report registration recall of 63.3% and 79.5%, exceeding 2D3D-MATR by 6.9 and 3.7 percentage points respectively, with ablations attributing gains to each component.

Significance. If the reported numbers withstand scrutiny, the paper makes a useful empirical contribution: it combines channel-wise feature alignment with a global matching formulation and shows consistent gains on two indoor benchmarks. The ablation study is detailed, and the design is modular and easy to build on. However, the headline gains are currently weakened by the absence of a definition for the new PIR metric, by an apparent test-set hyperparameter selection, and by a dimensional error in the optimal transport formulation. The paper does not provide code or pretrained models, so reproducibility rests entirely on the text.

major comments (3)
  1. [Sec. 4.1, Sec. 4.3, Fig. 6] The hyperparameters x (number of CMCF iterations) and λ (number of Sinkhorn iterations) are selected from Fig. 6(b,c), which reports RR on RGB-D Scenes V2, and the selected values x=1, λ=10 reproduce the headline RR of 63.3 in Table 1. Section 4.1 defines Scenes 9–10 as validation, but the text never states whether Fig. 6 was produced on that validation split or on the test split. If the latter, the claimed 6.9 pp improvement over 2D3D-MATR is a test-set-tuned number rather than the performance of a fixed model, and the comparison is not unbiased. Please report the sweeps on validation, or state clearly that they are on test and treat the headline claim accordingly.
  2. [Sec. 4.3, Tables 4/5/8] The new metric PIR is introduced in Sec. 4.3 as "the proportion of correctly matched points among all correspondences", but it is never formally defined. There is no threshold, no statement of which correspondence set (patch-level, dense, or after PnP-RANSAC) is used, and no formula analogous to Eq. (22). Because PIR is used in three ablation tables to support the CMCF module (e.g., "M5 achieving a 10.7pp PIR boost"), the metric must be defined before those claims can be evaluated.
  3. [Sec. 3.3, Eqs. (17)–(18)] The optimal transport formulation is dimensionally inconsistent. Eq. (17) states that T* ∈ R^{(H×W)×N}, while Eq. (18) constrains T to R^{N×N} with T1 = (1/N)1 and T^T1 = (1/N)1. These marginals are only feasible when the two sets have the same cardinality, which is not guaranteed because the image patch grid has H×W elements and the point cloud has N points. Please restate the search space and marginals consistently, e.g., T ∈ R^{(H×W)×N} with row and column marginals 1/(H×W) and 1/N, or explain any pooling step that reconciles the sizes.
minor comments (5)
  1. [Sec. 4.3, Table 8] In the paragraph following Table 8, the text says F2 produces "a slight increase in PIR to 83.4", but the table lists PIR 83.9 for F1 and 83.4 for F2, i.e., a decrease; it also says F5 yields "a 1.5 percentage point gain in RR over the baseline", while the table shows 79.5 versus 75.8, a 3.7 pp gain. Please correct these figures.
  2. [Sec. 4.3, Table 5] In the discussion of Table 5, the text reports that adding FI1 alone raises IR to 32.3, but the table reports 33.3; please reconcile the number.
  3. [Sec. 3.2.2, Eqs. (13)–(14)] Equations (13)–(14) compute an elementwise variance of per-sample covariance matrices across the batch, not a covariance matrix between channels; calling the result Cov^x a covariance matrix is misleading. Please rename or clarify the terminology.
  4. [Fig. 6(a)] Sub-figure 6(a) uses labels BL, M1, M2, M3, Full, which do not correspond to the M1–M8 notation of Table 4 or the F1–F5 notation of Table 8; please align the labels with the tables or explain the mapping.
  5. [Sec. 4, appendices] References to Appendices E, I, and J are made in Section 4, but the submitted manuscript contains no appendix material; please include the appendices or remove the references.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline RR is selected as the peak of hyperparameter sweeps on the same test benchmark; reported SOTA gain is a test-set-tuned number, not a fixed-model prediction.

  1. fitted input called prediction [Section 4.2, Figures 6(b)-(c) and Table 1 ('Visualization of ablation studies' and 'Evaluation results on RGB-D Scenes V2')]
    "We also conducted experiments to explore the impact of the number of CMCF iterations x and the Sinkhorn iterations λ on RGB-D Scenes V2. These studies aim to understand the impact of these parameters on model performance and to optimize their settings for better results. ... It is observed that the model achieves the highest registration recall (RR) of 63.3 when the number of CMCF iterations is set to 1. ... RR exhibits a noticeable peak at λ = 10, reaching 63.3."

    The final configuration (x=1, λ=10) is chosen because it maximizes RR on RGB-D Scenes V2, and the reported headline RR (63.3 in Table 1, giving the 6.9 pp gain over 2D3D-MATR) is exactly that maximum. The sweep curves are reported on RGB-D Scenes V2 as a whole, with no validation-only basis stated; the peak equals the test-set number. The method's numerical claim is therefore a selected optimum of the evaluation metric, not an unbiased prediction of a fixed model; the 'prediction' is by construction the argument maximum of the RR curve.

full rationale

The paper's modules (CAA/IME/CMCF and GOS/OT) are described empirically; there is no derivation chain that defines outputs in terms of inputs. CMCF's covariance-based masking is a design choice borrowed from external work [55], not a self-citation, and its effectiveness is supported by ablations (Tables 4 and 8). GOS is explicitly identified as optimal transport with Sinkhorn (Eqs. 17-18), so the renaming is not hidden. The one substantive circularity is the test-set hyperparameter selection: Fig. 6(b)-(c) sweeps x and λ on RGB-D Scenes V2 and reports RR=63.3 at exactly the adopted x=1, λ=10; Table 1 reports the same 63.3 as the SOTA result. This makes the headline 6.9 pp gain a test-set-tuned optimum rather than a fixed-model prediction. Because the modules' contributions still have independent empirical content, the circularity is partial; no self-citation chain or definitional equivalence is present. Score 6.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim rests on standard deep learning machinery plus several data-dependent choices: learned fusion weights, loss weighting, kernel size, Sinkhorn and CMCF iteration counts tuned against the benchmark, and a variance-grouping heuristic for the CMCF mask. No new physical or mathematical entity is introduced; the PIR metric is an ad hoc evaluation invention that is never formally specified.

free parameters (8)
  • alpha, beta, lambda, mu (feature fusion weights) = learned, not reported
    Eqs (8)-(9) weight original and enhanced features; values are learned during training.
  • lambda1, lambda2, lambda3 (loss weights) = not reported
    Eq (20) combines Lf, Lic, Lif; the values are hyperparameters not given in the paper.
  • k (ICE convolution kernel size) = 7
    Set in Sec 4.1; affects the ICE attention computation.
  • S (attention scaling factor) = C^{-0.5}
    Set in Sec 4.1; used in Eq (6) for PCE self-attention.
  • epsilon (entropy regularization weight in OT) = not reported
    Eq (17); not specified, only 'approximately 10 iterations'.
  • Sinkhorn iteration count = 10
    Selected because RR peaks at lambda=10 in Fig 6(c).
  • CMCF iteration count x = 1
    Selected because RR peaks at x=1 in Fig 6(b).
  • Variance-group thresholds for covariance mask = not specified
    Sec 3.2.2: covariance matrix divided into three groups by variance width; the grouping rule is not defined.
assumptions (5)
  • standard math Entropy-regularized optimal transport solved by Sinkhorn iterations converges to a valid transport plan with the stated marginal constraints (Eqs 17-18).
    Invoked in Sec 3.3 to justify GOS. However, Eq 18 states T in R^{N x N} with uniform margins, which conflicts with the rectangular T* in R^{(H*W) x N}; the discrepancy is unaddressed.
  • domain assumption The covariance matrix Covx in Eq 14 (computed as variance across batch samples of channel-pair statistics) identifies channels that are sensitive to variations across modalities and therefore should be suppressed.
    Sec 3.2.2 uses the highest-variance group to build mask Mx. The paper provides no independent evidence that these channels correspond to non-overlapping regions or mismatches; the assumption is borrowed from stereo matching domain generalization [55].
  • domain assumption Cosine similarity between flattened image and point cloud features is a valid proxy for patch-level matching probability.
    Used to form score map S before OT in Sec 3.3. This is inherited from 2D3D-MATR [24].
  • domain assumption Instance normalization generalizes better than batch normalization for cross-modal feature alignment in this task.
    Stated in Sec 3.2.2 without ablations comparing IN vs BN specifically.
  • ad hoc to paper The new metric PIR (proportion of correctly matched points among all correspondences) is well-defined and is a valid measure of matching quality.
    Introduced in Sec 4.3 with no formal definition, making comparisons with PIR non-reproducible.
invented entities (1)
  • PIR (Patch Inlier Ratio)
    purpose: New metric to quantify proportion of correctly matched points among all correspondences; used in ablations (Tables 4, 5, 8).
    Defined only in prose in Sec 4.3; no formula or threshold given, so it cannot be computed by readers. It is invented for this paper's ablation story.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CA-I2P: Channel-Adaptive Registration Network with Global Optimal Selection." pith.science (2026). https://pith.science/paper/VN4UEHWW

@misc{pith2026250621364,
  author       = {Pith},
  title        = {Pith review of: CA-I2P: Channel-Adaptive Registration Network with Global Optimal Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VN4UEHWW}},
  note         = {Machine review of arXiv:2506.21364}
}
read the original abstract

Detection-free methods typically follow a coarse-to-fine pipeline, extracting image and point cloud features for patch-level matching and refining dense pixel-to-point correspondences. However, differences in feature channel attention between images and point clouds may lead to degraded matching results, ultimately impairing registration accuracy. Furthermore, similar structures in the scene could lead to redundant correspondences in cross-modal matching. To address these issues, we propose Channel Adaptive Adjustment Module (CAA) and Global Optimal Selection Module (GOS). CAA enhances intra-modal features and suppresses cross-modal sensitivity, while GOS replaces local selection with global optimization. Experiments on RGB-D Scenes V2 and 7-Scenes demonstrate the superiority of our method, achieving state-of-the-art performance in image-to-point cloud registration.

Figures

Figures reproduced from arXiv: 2506.21364 by the authors.

Figure 1
Figure 1. (a) Visualization of feature variations across channels. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of CA-I2P. We use a feature extraction backbone to obtain features from images and point clouds. The image and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The architecture of Intra-Modal Enhancement Stage. It includes the Image Channel Enhancement Unit (ICE) and the Point [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Indoor and outdoor scenes with images, point clouds, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of performance comparison of models on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization of ablation studies. These results confirm that IME, CMCF, and GOS each play a crucial role in enhancing image-to-point cloud matching. IME strengthens intra-modal feature quality, CMCF filters out noisy and redundant features, and GOS optimizes global ma…
Figure 7
Figure 7. Figure 7: Visualization of modality differences. With training, the [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: Visualzation of intra-model enhancement stage. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 8
Figure 8. Figure 8: The visualization result of CA-I2P [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 43 canonical work pages

  1. [55]

    Revisiting domain gener- alized stereo matching networks from a feature consistency perspective

    J Zhang, X Wang, X Bai, et al. Revisiting domain gener- alized stereo matching networks from a feature consistency perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13001– 13011, 2022. 2, 6

    work page 2022

  2. [29]

    Experimental implementation of extended multi- variable ppf control on an active structure

    SO Reza Moheimani, Benjamin JG Vautier, and Bharath Bhikkaji. Experimental implementation of extended multi- variable ppf control on an active structure. IEEE Transac- tions on Control Systems Technology, 14(3):443–455, 2006. 2

    work page 2006

  3. [32]

    Sift: Predicting amino acid changes that affect protein function

    Pauline C Ng and Steven Henikoff. Sift: Predicting amino acid changes that affect protein function. Nucleic acids re- search, 31(13):3812–3814, 2003. 2, 3

    work page 2003

  4. [1]

    Andres Barroso-Laguna, Edgar Riba, Daniel Ponsa, et al. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 5836–5844, 2019. 3

    work page 2019

  5. [2]

    Reproducing kernel Hilbert spaces in probability and statistics

    Alain Berlinet and Christine Thomas-Agnan. Reproducing kernel Hilbert spaces in probability and statistics . Springer Science & Business Media, 2011. 7

    work page 2011

  6. [3]

    Comparison of methods generalizing max-and average- pooling

    Florian Bieder, Robin Sandk ¨uhler, and Philippe C Cat- tin. Comparison of methods generalizing max-and average- pooling. arXiv preprint arXiv:2103.01746, 2021. 3

    arXiv 2021

  7. [4]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Ad- vances in Neural Information Processing Systems, 33:9912– 9924, 2020. 6

    work page 2020

  8. [5]

    Fully convolutional geometric features

    Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Proceedings of the IEEE/CVF international conference on computer vision , pages 8958–8966, 2019. 8, 9

    work page 2019

Show all 54 references
  1. [6]

    Joint distribution optimal transportation for domain adaptation

    Nicolas Courty, R ´emi Flamary, Amaury Habrard, et al. Joint distribution optimal transportation for domain adaptation. Advances in Neural Information Processing Systems , 30,

  2. [7]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in Neural Information Pro- cessing Systems, 26:3–8, 2013. 2

    2013

  3. [8]

    Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence

    Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2023

  4. [9]

    Unsuper- vised template-assisted point cloud shape correspondence network

    Jiacheng Deng, Jiahao Lu, and Tianzhu Zhang. Unsuper- vised template-assisted point cloud shape correspondence network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5250– 5259, 2024. 1

    2024

  5. [10]

    Quantity- quality enhanced self-training network for weakly super- vised point cloud semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Jiacheng Deng, Jiahao Lu, and Tianzhu Zhang. Quantity- quality enhanced self-training network for weakly super- vised point cloud semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1

    2025

  6. [11]

    Simultaneous local- ization and mapping: part i

    Hugh Durrant-Whyte and Tim Bailey. Simultaneous local- ization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 1

    2006

  7. [12]

    2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud

    Mengdan Feng, Sixing Hu, Marcelo H Ang, and Gim Hee Lee. 2d3d-matchnet: Learning to match keypoints across 2d image and 3d point cloud. In 2019 International Confer- ence on Robotics and Automation (ICRA), pages 4790–4796. IEEE, 2019. 1, 3

    2019

  8. [13]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395, 1981. 2, 3, 6

    1981

  9. [14]

    Ota: Optimal transport assignment for object de- tection

    Zhiqiang Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object de- tection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 303–312,

  10. [15]

    Real-time rgb-d camera relocalization

    Ben Glocker, Shahram Izadi, Jamie Shotton, and Antonio Criminisi. Real-time rgb-d camera relocalization. In 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 173–179. IEEE, 2013. 7

    2013

  11. [16]

    Hierarchical shape-consistent trans- former for unsupervised point cloud shape correspondence

    Jianfeng He, Jiacheng Deng, Tianzhu Zhang, Zhe Zhang, and Yongdong Zhang. Hierarchical shape-consistent trans- former for unsupervised point cloud shape correspondence. IEEE Transactions on Image Processing , 32:2734–2748,

  12. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 3

    2016

  13. [18]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132–7141, 2018. 2

    2018

  14. [19]

    Predator: Registration of 3d point clouds with low overlap

    Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4267–4276, 2021. 8, 9

    2021

  15. [20]

    Cofii2p: Coarse-to-fine correspondences for image-to-point cloud registration

    Shuhao Kang, Youqi Liao, Jianping Li, Fuxun Liang, Yuhao Li, Fangning Li, Zhen Dong, and Bisheng Yang. Cofii2p: Coarse-to-fine correspondences for image-to-point cloud registration. arXiv preprint arXiv:2309.14660, 2023. 2

    2023 arXiv

  16. [21]

    Transformers in vision: A survey

    Salman Khan, Muzammal Naseer, Munawar Hayat, et al. Transformers in vision: A survey. ACM Computing Surveys (CSUR), 54(10s):1–41, 2022. 4

    2022

  17. [22]

    Unsupervised fea- ture learning for 3d scene labeling

    Kevin Lai, Liefeng Bo, and Dieter Fox. Unsupervised fea- ture learning for 3d scene labeling. In 2014 IEEE Inter- national Conference on Robotics and Automation (ICRA) , pages 3050–3057. IEEE, 2014. 7

    2014

  18. [23]

    Ep n p: An accurate o (n) solution to the p n p problem

    Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p n p problem. International journal of computer vision, 81:155–166, 2009. 2, 3, 6

    2009

  19. [24]

    2d3d-matr: 2d-3d matching transformer for detection-free registration between images and point clouds

    Minhao Li, Zheng Qin, Zhirui Gao, Renjiao Yi, Chenyang Zhu, Yulan Guo, and Kai Xu. 2d3d-matr: 2d-3d matching transformer for detection-free registration between images and point clouds. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 14128–14138,

  20. [25]

    Feature pyra- mid networks for object detection

    Tsung-Yi Lin, Piotr Doll ´ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyra- mid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 2117–2125, 2017. 3

    2017

  21. [26]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 3

    2021

  22. [27]

    Misunderstanding analysis of covariance

    Gerald A Miller and Jean P Chapman. Misunderstanding analysis of covariance. Journal of Abnormal Psychology , 110(1):40, 2001. 5

    2001

  23. [28]

    Rotate to attend: Convolutional triplet attention module

    Diganta Misra, Tanishq Nalamada, Anoop Umapathi Arasa- nipalai, et al. Rotate to attend: Convolutional triplet attention module. In Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 3139–3148,

  24. [30]

    Real time localization and 3d reconstruction

    Etienne Mouragnon, Maxime Lhuillier, Michel Dhome, Fa- bien Dekeyser, and Patrick Sayd. Real time localization and 3d reconstruction. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), pages 363–370. IEEE, 2006. 1

    2006

  25. [31]

    Generalized max pooling

    Nicolas Murray and Florent Perronnin. Generalized max pooling. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , pages 2473–2480,

  26. [33]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  27. [35]

    Geometric transformer for fast and robust point cloud registration

    Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yux- ing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143–11152, 2022. 6

    2022

  28. [36]

    Geotrans- former: Fast and robust point cloud registration with geo- metric transformer

    Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotrans- former: Fast and robust point cloud registration with geo- metric transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9806–9821, 2023. 2

    2023

  29. [37]

    Corri2p: Deep image-to-point cloud registration via dense correspondence

    Siyu Ren, Yiming Zeng, Junhui Hou, and Xiaodong Chen. Corri2p: Deep image-to-point cloud registration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1198–1208, 2022. 1

    2022

  30. [38]

    Orb: An efficient alternative to sift or surf

    Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision , pages 2564–

    2011

  31. [39]

    Fast point feature histograms (fpfh) for 3d registration

    Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation , pages 3212–3217. IEEE, 2009. 2

    2009

  32. [40]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7

    2016

  33. [41]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2

    2020

  34. [42]

    Hand key- point detection in single images using multiview bootstrap- ping

    Tomas Simon, Hanbyul Joo, Iain Matthews, et al. Hand key- point detection in single images using multiview bootstrap- ping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1145–1153, 2017. 3

    2017

  35. [43]

    Comments on integral variants of iss

    Eduardo D Sontag. Comments on integral variants of iss. Systems & Control Letters, 34(1-2):93–100, 1998. 3

    1998

  36. [44]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8922–8931, 2021. 2

    2021

  37. [45]

    Circle loss: A unified perspective of pair similarity optimization

    Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398–6407, 2020. 6

    2020

  38. [46]

    Kpconv: Flexible and deformable convolution for point clouds

    Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3

    2019

  39. [47]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research , 9 (11), 2008. 7

    2008

  40. [48]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3

    2017

  41. [49]

    P2-net: Joint description and detection of local features for pixel and point matching

    Bing Wang, Changhao Chen, Zhaopeng Cui, Jie Qin, Chris Xiaoxuan Lu, Zhengdi Yu, Peijun Zhao, Zhen Dong, Fan Zhu, Niki Trigoni, et al. P2-net: Joint description and detection of local features for pixel and point matching. In Proceedings of the IEEE/CVF International Conference...

    2021

  42. [50]

    Freereg: Image- to-point cloud registration leveraging pretrained diffusion models and monocular depth estimators

    Hao Wang, Yifan Liu, Bo Wang, et al. Freereg: Image- to-point cloud registration leveraging pretrained diffusion models and monocular depth estimators. arXiv preprint arXiv:2310.03420, 2023. 3, 9

    2023 arXiv

  43. [51]

    Brightness perceiving for recursive low-light image enhancement

    H Wang, L Peng, Y Sun, et al. Brightness perceiving for recursive low-light image enhancement. IEEE Transactions on Artificial Intelligence, 5(6):3034–3045, 2023. 1

    2023

  44. [52]

    Efficient loftr: Semi-dense local feature matching with sparse-like speed

    Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21666–21675, 2024. 2

    2024

  45. [53]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, et al. Cbam: Convolutional block attention module. In Pro- ceedings of the European Conference on Computer Vision (ECCV), pages 3–19, 2018. 2, 3

    2018

  46. [54]

    Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration

    Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobo- dan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. Advances in Neural Infor- mation Processing Systems, 34:23872–23884, 2021. 2

    2021

Pith tools

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