Pith. sign in

REVIEW 4 major objections 4 minor 51 references

RARE: Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning

T0 review · 4 major / 4 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 Depth diffusion features from a pretrained image diffusion model can refine any point cloud registration method without training, raising GeoTransformer's 3DMatch recall from 92.0% to 94.5%.

desk verdict A sensible zero-shot refinement for point cloud registration that works on standard benchmarks, but the 'refine any registration' claim outruns the evidence and the results lack error bars and a ZeroMatch comparison. read the letter →

arxiv 2507.19950 v1 pith:R7V4HMAB submitted 2025-07-26 cs.CV cs.AI

Chengyu Zheng , Jin Huang , Honghua Chen , Mingqiang Wei This is my paper
classification cs.CVcs.AI
keywords pointcloudregistrationzero-shotrefinementdiffusionfeaturesdepthmapprojectionControlNetStableinlieraggregation3DMatch
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 claims that point cloud registration, a core step in robotics and 3D reconstruction, can be refined without any training data by borrowing the internal features of a pretrained image diffusion model. The idea is to use the initial transform produced by an existing registration method to project both point clouds into depth maps with small parallax, extract depth diffusion features with ControlNet/Stable Diffusion, fuse these with the original 3D geometric features, and re-estimate the rigid transform from the enriched correspondences. The method is a wrapper around any existing registration method, and the experiments report consistent gains on 3DMatch, 3DLoMatch, and KITTI, including raising GeoTransformer's registration recall on 3DMatch from 92.0% to 94.5% while reducing rotation and translation errors. The practical interest is that this turns a pretrained 2D generative model into an alignment aid without needing paired 3D training data.

What carries the argument

The load-bearing object is the depth diffusion feature: a multi-layer feature map extracted from the decoder of a pretrained Stable Diffusion U-Net under ControlNet's depth conditioning. The method samples denoising steps from pure noise, aggregates features from U-Net layers 0, 3, and 6, applies PCA for dimensionality reduction, and upsamples the maps to a common resolution. These 2D features are attached back to the point clouds by a pixel-to-point Feature Transfer module that finds common points by Euclidean nearest-neighbor matching under a threshold, then concatenates L2-normalized 3D and diffusion features. An Inlier Aggregation Module builds five correspondence sets, computes a candidate rigid transform for each by weighted SVD, selects the transform with the most inlier support, and iterates the reweighting. The mechanism works because projecting both clouds through the initial transform produces small-parallax depth views, the situation in which the paper argues diffusion features are more consistent and more distinctive.

What would settle it

Take 3DMatch pairs and deliberately perturb the initial transformation by increasing rotation angles such as 2°, 5°, 10°, and 15° before running RARE. The paper's small-parallax mechanism predicts that the refinement gain over the base method shrinks monotonically and eventually becomes negative as the perturbation grows; if RARE still improves the perturbed pairs, or if depth diffusion correspondences remain accurate under large parallax, the stated mechanism is not the one doing the work.

Watch

Extended reading notes

Core claim

The central discovery the paper is trying to establish is that diffusion features extracted from depth maps become strongly discriminative for point-cloud correspondence once the two views are pre-aligned by an initial transformation, even though the same features struggle on large-parallax depth pairs. On this basis, RARE takes any existing registration method's 3D features and initial transform, projects the source and reference clouds into depth maps from both the reference and source viewpoints, extracts multi-scale ControlNet diffusion features, reduces them with PCA, and transfers them onto 3D points by nearest-neighbor matching under a distance threshold. Five correspondence sets (reference-common, source-common, reference-diffusion, source-diffusion, and 3D) are then pooled through a weighted-SVD inlier aggregation module that iteratively reweights matches and selects the transformation admitting the most inliers. The paper reports that this training-free wrapper improves registration recall and lowers rotation and translation error relative to each base method it wraps, with the largest gains coming from the inlier aggregation rather than the diffusion features alone.

Load-bearing premise

The method depends on the initial alignment being close enough that the two projected depth images look nearly the same from the two viewpoints; if that starting alignment is too far off, the diffusion features extracted from the depth images can no longer be trusted to find correct correspondences.

Editorial extensions

If this is right

  • Wrapping GeoTransformer, Predator, CoFiNet, or FCGF with RARE raises 3DMatch registration recall over the base method by 2.5 to 8.3 percentage points without any training.
  • On low-overlap 3DLoMatch pairs, RARE reaches 76.7% recall while posting the best rotation error (2.55°) and translation error (0.074) among the compared refinement pipelines.
  • On KITTI, where registration recall is already saturated at 99.8%, RARE still reduces translation error from 0.068 to 0.064 without hurting recall.
  • Ablations show that the inlier aggregation module contributes the largest share of the final gain, and removing the feature transfer step drops 3DMatch recall from 94.5% to 93.1%.

Reading between the lines

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

  • If the paper is right, the same wrapper should apply to sparse-keypoint or object-level registration methods as well, since the diffusion features are computed from dense depth projections rather than from any specific keypoint detector.
  • A natural testable extension is to condition the diffusion branch on color or RGB-D data in addition to depth; the paper's mechanism does not require depth to be the only input, so multi-modal conditioning may sharpen correspondences further.
  • The reliance on the initial transform suggests that at very low overlap a poor initialization could turn the diffusion branch into a liability; adding a confidence gate that falls back to the base method when the refined transform does not improve inlier counts would be a cheap safeguard.
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 / 4 minor

Summary. The paper proposes RARE, a zero-shot refinement module for pairwise point cloud registration. Given an initial transformation Tinit from an existing registration method, RARE projects the reference and source point clouds into depth maps from both viewpoints, extracts diffusion features from a pretrained ControlNet/Stable Diffusion model, transfers these features onto 3D points via coordinate-space nearest-neighbor matching, and then uses an inlier aggregation module with weighted SVD and transformation selection to produce a refined rigid transformation. Experiments on 3DMatch, 3DLoMatch, and KITTI report improved registration recall and reduced rotation/translation errors over several baselines initialized with GeoTransformer, and ablations on 3DMatch show that the feature transfer, PCA, and inlier aggregation components each contribute to the final result.

Significance. If the central claims hold, the work is valuable: it offers a training-free, plug-in refinement that exploits powerful pretrained 2D diffusion features for 3D registration, releases code, and reports consistent improvements across indoor and outdoor benchmarks. The absence of learned parameters in the refinement pipeline is a genuine strength, as is the ablation evidence that each module contributes. However, the paper's headline claim that RARE refines 'any' registration method is not currently supported by the experiments: all main benchmark tables use GeoTransformer as the initialization, the generalization study is limited to 3DMatch, and the method's sensitivity to initial transformation quality is never measured. The lack of a comparison with the most closely related zero-shot method, ZeroMatch, and the absence of error bars or significance tests further limit the strength of the empirical conclusions. The underlying idea is promising and the reported gains are plausible, but the evidence as presented is incomplete.

major comments (4)
  1. [sec. 4.2, Table 3] The paper's title and abstract claim that RARE refines 'any' or 'existing' registration methods, but the main benchmark comparisons in Tables 1, 2, and 4 all use GeoTransformer as the initialization, and the only multi-method generalization experiment (Table 3) is confined to 3DMatch. Moreover, the results in Table 3 show strong sensitivity to initialization: FCGF+Ours reaches only 76.2 RR versus 94.5 for GeoTransformer+Ours, and Predator+Ours and CoFiNet+Ours both underperform GeoTransformer+Ours on RR. This does not establish the claimed regime of applicability. Please add a systematic study that varies initial transformation quality while holding all other factors fixed (for example, by applying controlled rotation/translation perturbations to Tinit, or by using a range of weaker and stronger initializers), and report RR, RE, and TE as functions of the initial registration error on both 3DMatch and 3DLoMatch. Without such an experiment, the central 'refine any registration' claim is unproven.
  2. [sec. 3.2, Table 5; sec. 4.1] The diffusion layer set [0,3,6] is selected based on Table 5, which is evaluated on the 3DMatch test benchmark itself, and the same layer set is then used in the main 3DMatch evaluation (Tables 1 and 3). This constitutes test-set tuning and weakens the 'zero-shot' claim, since the design choice is informed by the benchmark on which the method is later reported. Please either select the layers on a validation split, or report the sensitivity of the main results to this choice. In addition, several key hyperparameters are not specified anywhere: tau_a in Eq. (5), the value of K (the number of correspondences used for the final SVD), the number of re-weighting iterations N_r, the PCA output dimension, the distance threshold for the nearest-neighbor common-point identification in Section 3.3, the number of diffusion sampling steps, and the depth-map resolution. These values are needed to reproduce the method and to judge how robust the reported gains are to hyperparameter choices.
  3. [sec. 2.3 and sec. 4.1] The related-work section cites ZeroMatch [20] as a method that 'uses the powerful image representations of SD to enhance geometric descriptors for robust registration,' which is the closest existing zero-shot approach to RARE, yet ZeroMatch is absent from all experimental comparisons. Without a head-to-head comparison on 3DMatch and 3DLoMatch (and KITTI, if the method supports it), the incremental contribution over this most relevant prior work cannot be assessed. Please add ZeroMatch to the comparison tables, or explicitly argue with quantitative evidence why it is not comparable (e.g., different input modalities or evaluation protocol).
  4. [Tables 1-4 and Table 6] All experimental tables report single point estimates without error bars, confidence intervals, or significance tests. Several of the reported gains are small (for example, RR 94.1 vs. 94.5 in Table 1 and TE 0.050 vs. 0.052 in Table 1 and Table 6), so it is unclear whether the differences are meaningful or within run-to-run variability. Please report the variance over multiple runs or bootstrap confidence intervals for the main metrics, and state the number of runs used. This is particularly important because the method is deterministic in its refinement steps but depends on the stochastic diffusion sampling, so repeated evaluations are feasible.
minor comments (4)
  1. [Eq. (3)] The notation 'F usedF{ref,src }' in Eq. (3) is difficult to parse; please clarify the subscripts and superscripts, for example by using a clearer notation such as F_p^{ref,fused}.
  2. [Fig. 1 caption] The caption states that '(f) presents the refined registration result by diffusion features,' but (f) in the figure is the rotated source-view depth map; the refined registration result appears to be labeled (h). Please correct the caption.
  3. [Sec. 4.4, Table 5] The text says 'we report the size of output feature maps of 13 layers,' but Table 5 only lists a handful of layers and layer combinations. Either present the full per-layer table or rephrase the description to match what is shown.
  4. [Sec. 4.1] The implementation details mention sampling 500 points for 3DMatch and 1000 points for KITTI, but no runtime or memory usage is reported. Since the method uses a full diffusion model on multiple depth maps, a runtime comparison with the baselines would help readers judge practical applicability.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: RARE's refinement combines externally pretrained diffusion features with the initial transformation as an input, and no derived quantity reduces to a fitted parameter by construction.

full rationale

The central derivation in Sections 3.2 and 3.3 treats the initial transformation Tinit and the 3D features Fp, Fq as inputs, then computes a refined transformation via feature transfer, inlier aggregation, weighted SVD (Eq. 4), and transformation selection (Eq. 5). The output is not defined as Tinit, nor does any equation make the final estimate equal to an input by construction. The depth maps are produced by projecting the point clouds using Tinit, and the diffusion features come from the externally pretrained Stable Diffusion/ControlNet models, so the refinement signal is not manufactured from the target metric. The only mild concern is that the diffusion layer choice [0,3,6] is selected after an ablation on 3DMatch (Table 5) and then used in the main 3DMatch evaluation (Table 1); this is benchmark-driven hyperparameter selection rather than a fitted quantity renamed as a prediction, and under the review rules it is not circularity. There is also one self-citation with author overlap: reference [5] (ImLoveNet) includes co-authors H. Chen and M. Wei, but it appears only in Related Work and is not load-bearing for the method. No uniqueness theorem or ansatz is imported from the authors' own prior work. The overbroad claim of refining 'any' registration method is a generalization concern about sensitivity to Tinit quality, not a circularity: the paper itself shows weaker gains with FCGF initialization (76.2 RR vs. 94.5 with GeoTransformer), which is an empirical robustness limitation, not a derivation that reduces to its own inputs. Overall, the paper's central empirical claim is self-contained and externally benchmarked, so circularity is minimal or absent.

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

The method contributes no fitted weights, but its performance depends on hand-set choices such as diffusion layers, sampling counts, aggregation thresholds, and PCA dimensions, plus the premise that pretrained diffusion features on co-aligned depth maps are usable for point matching. The layer choice is selected on the same benchmark used for main results, which is the main ledger concern. No new physical or conceptual entities are introduced; the modules are compositions of existing operations.

free parameters (4)
  • Diffusion layers selected [0,3,6] = [0,3,6]
    Chosen from Table 5 ablation on 3DMatch and used in all main results, tying the hyperparameter selection to the benchmark being reported.
  • Sampling count = 500 points for 3DMatch/3DLoMatch, 1000 for KITTI
    Hand-set in Section 4.1; affects correspondences and runtime, with no sensitivity analysis.
  • Inlier aggregation hyperparameters = Not reported (tau_a, K, N_r)
    Used in Equations 4 and 5; no values are given, so the final transformation selection is under-specified.
  • Nearest-neighbor threshold and PCA dimension = Not reported
    The Feature Transfer and PCA steps depend on thresholds and output dimensions that are not specified in the paper.
assumptions (4)
  • domain assumption Pretrained Stable Diffusion and ControlNet features extracted from depth maps encode correspondences suitable for matching.
    Central premise for depth diffusion features, introduced in Section 3.2.
  • ad hoc to paper Reducing parallax by aligning depth maps with Tinit makes diffusion features more distinctive.
    Stated as an observation in the introduction and Figure 2, but not formally justified.
  • domain assumption Nearest-neighbor matching in 3D coordinates identifies common points between point clouds and projected diffusion feature maps.
    Used in Section 3.3 Feature Transfer; requires projection errors to be below the matching threshold.
  • standard math Weighted SVD with inlier counting yields the correct rigid transformation.
    Equation 4 is a standard technique, but the result depends on confidence weights that are not fully specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RARE: Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning." pith.science (2026). https://pith.science/paper/R7V4HMAB

@misc{pith2026250719950,
  author       = {Pith},
  title        = {Pith review of: RARE: Refine Any Registration of Pairwise Point Clouds via Zero-Shot Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7V4HMAB}},
  note         = {Machine review of arXiv:2507.19950}
}
read the original abstract

Recent research leveraging large-scale pretrained diffusion models has demonstrated the potential of using diffusion features to establish semantic correspondences in images. Inspired by advancements in diffusion-based techniques, we propose a novel zero-shot method for refining point cloud registration algorithms. Our approach leverages correspondences derived from depth images to enhance point feature representations, eliminating the need for a dedicated training dataset. Specifically, we first project the point cloud into depth maps from multiple perspectives and extract implicit knowledge from a pretrained diffusion network as depth diffusion features. These features are then integrated with geometric features obtained from existing methods to establish more accurate correspondences between point clouds. By leveraging these refined correspondences, our approach achieves significantly improved registration accuracy. Extensive experiments demonstrate that our method not only enhances the performance of existing point cloud registration techniques but also exhibits robust generalization capabilities across diverse datasets. Codes are available at https://github.com/zhengcy-lambo/RARE.git.

Figures

Figures reproduced from arXiv: 2507.19950 by the authors.

Figure 1
Figure 1. RARE refines arbitrary point cloud registration methods effectively in a zero-shot learning manner. (a) and (e) show the source [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of correspondence matching using pro [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Pipeline of our framework. Our framework begins by employing GeoTransformer [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Registration visualization results on 3DMatch. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Registration visualization results on 3DLoMatch. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 49 canonical work pages

  1. [20]

    Zero-shot rgb-d point cloud registration with pre- trained large vision model

    Haobo Jiang, Jin Xie, Jian Yang, Liang Yu, and Jianmin Zheng. Zero-shot rgb-d point cloud registration with pre- trained large vision model. In IEEE/CVF Computer Vision and Pattern Recognition Conference , pages 16943–16952,

    work page

  2. [1]

    BUFFER: balancing accuracy, efficiency, and gener- alizability in point cloud registration

    Sheng Ao, Qingyong Hu, Hanyun Wang, Kai Xu, and Yulan Guo. BUFFER: balancing accuracy, efficiency, and gener- alizability in point cloud registration. In IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1255–1264, 2023. 3

    work page 2023

  3. [2]

    Pointdsc: Robust point cloud registration using deep spatial consistency

    Xuyang Bai, Zixin Luo, Lei Zhou, Hongkai Chen, Lei Li, Zeyu Hu, Hongbo Fu, and Chiew-Lan Tai. Pointdsc: Robust point cloud registration using deep spatial consistency. In IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 15859–15869, 2021. 6, 7

    work page 2021

  4. [3]

    Bootstrap your own correspondences

    Mohamed El Banani and Justin Johnson. Bootstrap your own correspondences. In IEEE/CVF International Conference on Computer Vision, pages 6413–6422, 2021. 3

    work page 2021

  5. [4]

    Graph-cut RANSAC

    Daniel Barath and Jiri Matas. Graph-cut RANSAC. In IEEE Conference on Computer Vision and Pattern Recognition , pages 6733–6741, 2018. 3

    work page 2018

  6. [5]

    Imlovenet: Misaligned image-supported regis- tration network for low-overlap point cloud pairs

    Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported regis- tration network for low-overlap point cloud pairs. In ACM SIGGRAPH, pages 1–9, 2022. 3

    work page 2022

  7. [6]

    Diffu- siondet: Diffusion model for object detection

    Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffu- siondet: Diffusion model for object detection. In IEEE/CVF International Conference on Computer Vision, pages 19773– 19786, 2023. 3

    work page 2023

  8. [7]

    Hinton, and David J

    Ting Chen, Lala Li, Saurabh Saxena, Geoffrey E. Hinton, and David J. Fleet. A generalist framework for panoptic seg- mentation of images and videos. In IEEE/CVF International Conference on Computer Vision, pages 909–919, 2023. 3

    work page 2023

Show all 51 references
  1. [8]

    Sc 2-pcr: A second order spatial compatibility for efficient and robust point cloud registration

    Zhi Chen, Kun Sun, Fan Yang, and Wenbing Tao. Sc 2-pcr: A second order spatial compatibility for efficient and robust point cloud registration. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 13211–13221,

  2. [9]

    Sc2-pcr++: Rethinking the generation and selection for ef- ficient and robust point cloud registration

    Zhi Chen, Kun Sun, Fan Yang, Lin Guo, and Wenbing Tao. Sc2-pcr++: Rethinking the generation and selection for ef- ficient and robust point cloud registration. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 45(10): 12358–12376, 2023. 3, 6

    2023

  3. [10]

    Choy, Jaesik Park, and Vladlen Koltun

    Christopher B. Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In IEEE/CVF Interna- tional Conference on Computer Vision , pages 8957–8965,

  4. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780–8794, 2021. 3

    2021

  5. [12]

    Tam- ing transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bj ¨orn Ommer. Tam- ing transformers for high-resolution image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 4

    2021

  6. [13]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381–395, 1981. 1, 3, 6, 7

    1981

  7. [14]

    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. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3354–3361, 2012. 2, 6

    2012

  8. [15]

    3d object recognition in cluttered scenes with local surface features: A survey

    Yulan Guo, Mohammed Bennamoun, Ferdous Ahmed Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 36(11): 2270–2287, 2014. 1

    2014

  9. [16]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Informa- tion Processing Systems, 2020. 4

    2020

  10. [17]

    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 IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4267– 4276, 2021. 2, 3, 6, 7

    2021

  11. [18]

    Psreg: Prior-guided sparse mixture of experts for point cloud regis- tration

    Xiaoshui Huang, Zhou Huang, Yifan Zuo, Yongshun Gong, Chengdong Zhang, Deyang Liu, and Yuming Fang. Psreg: Prior-guided sparse mixture of experts for point cloud regis- tration. In AAAI, pages 3788–3796, 2025. 3

    2025

  12. [19]

    Robust outlier rejection for 3d reg- istration with variational bayes

    Haobo Jiang, Zheng Dang, Zhen Wei, Jin Xie, Jian Yang, and Mathieu Salzmann. Robust outlier rejection for 3d reg- istration with variational bayes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1148– 1157, 2023. 3, 6, 7

    2023

  13. [21]

    Difnet: Semantic segmentation by diffusion networks

    Peng Jiang, Fanglin Gu, Yunhai Wang, Changhe Tu, and Baoquan Chen. Difnet: Semantic segmentation by diffusion networks. In Advances in Neural Information Processing Systems, pages 1637–1646, 2018. 3

    2018

  14. [22]

    Johnson and Martial Hebert

    Andrew E. Johnson and Martial Hebert. Using spin images for efficient object recognition in cluttered 3d scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(5):433–449, 1999. 2

    1999

  15. [23]

    Le, Thanh-Toan Do, Tuan Hoang, and Ngai-Man Cheung

    Huu M. Le, Thanh-Toan Do, Tuan Hoang, and Ngai-Man Cheung. SDRSAC: semidefinite-based randomized ap- proach for robust point cloud registration without correspon- dences. In IEEE Conference on Computer Vision and Pattern Recognition, pages 124–133, 2019. 3

    2019

  16. [24]

    Gesac: Robust graph enhanced sample consensus for point cloud registra- tion

    Jiayuan Li, Qingwu Hu, and Mingyao Ai. Gesac: Robust graph enhanced sample consensus for point cloud registra- tion. ISPRS Journal of Photogrammetry and Remote Sens- ing, 167:363–374, 2020. 3

    2020

  17. [25]

    Density-invariant features for distant point cloud registration

    Quan Liu, Hongzi Zhu, Yunsong Zhou, Hongyang Li, Shan Chang, and Minyi Guo. Density-invariant features for distant point cloud registration. In IEEE/CVF International Confer- ence on Computer Vision, pages 18169–18179, 2023. 3

    2023

  18. [26]

    Extend your own correspondences: Unsupervised distant point cloud registration by progressive distance extension

    Quan Liu, Hongzi Zhu, Zhenxi Wang, Yunsong Zhou, Shan Chang, and Minyi Guo. Extend your own correspondences: Unsupervised distant point cloud registration by progressive distance extension. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20816–20826, 2024. 3

    2024

  19. [27]

    Overlap-guided gaussian mix- ture models for point cloud registration

    Guofeng Mei, Fabio Poiesi, Cristiano Saltori, Jian Zhang, Elisa Ricci, and Nicu Sebe. Overlap-guided gaussian mix- ture models for point cloud registration. In IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 4500–4509, 2023. 3

    2023

  20. [28]

    Unsu- pervised deep probabilistic approach for partial point cloud registration

    Guofeng Mei, Hao Tang, Xiaoshui Huang, Weijie Wang, Juan Liu, Jian Zhang, Luc Van Gool, and Qiang Wu. Unsu- pervised deep probabilistic approach for partial point cloud registration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13611–13620, 2023. 3

    2023

  21. [29]

    Mian, Mohammed Bennamoun, and Robyn A

    Ajmal S. Mian, Mohammed Bennamoun, and Robyn A. Owens. Automatic correspondence for 3d modeling: an ex- tensive review. Int. J. Shape Model. , 11(2):253–291, 2005. 1

    2005

  22. [30]

    Kim, and Niloy J

    Luca Morreale, Noam Aigerman, Vladimir G. Kim, and Niloy J. Mitra. Neural semantic surface maps. Computer Graphics Forum, 43(2):i–iii, 2024. 3

    2024

  23. [31]

    Geometric transformer for fast and robust point cloud registration

    Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 11133–11142,

  24. [32]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674– 10685, 2022. 2, 3

    2022

  25. [33]

    Aligning point cloud views using persistent feature histograms

    Radu Bogdan Rusu, Nico Blodow, Zoltan Csaba Marton, and Michael Beetz. Aligning point cloud views using persistent feature histograms. In International Conference on Intelli- gent Robots and Systems, pages 3384–3391, 2008. 3

    2008

  26. [34]

    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 IEEE International Conference on Robotics and Automation, pages 3212–3217, 2009. 2, 3

    2009

  27. [35]

    Lee, Jonathan Ho, Tim Salimans, David J

    Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mo- hammad Norouzi. Palette: Image-to-image diffusion mod- els. In SIGGRAPH, pages 15:1–15:10, 2022. 3

    2022

  28. [36]

    Buffer-x: Towards zero-shot point cloud registration in diverse scenes

    Minkyun Seo, Hyungtae Lim, Kanghee Lee, Luca Car- lone, and Jaesik Park. Buffer-x: Towards zero-shot point cloud registration in diverse scenes. arXiv preprint arXiv:2503.07940, 2025. 3

    2025

  29. [37]

    Semantic diffusion network for semantic segmentation

    Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Advances in Neural Information Processing Systems, pages 8702–8716, 2022. 3

    2022

  30. [38]

    Emergent correspondence from image diffusion

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems, 2023. 2, 3, 5

    2023

  31. [39]

    You only hypothesize once: Point cloud registration with rotation-equivariant descriptors

    Haiping Wang, Yuan Liu, Zhen Dong, and Wenping Wang. You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In ACM International Con- ference on Multimedia, pages 1630–1641, 2022. 2

    2022

  32. [40]

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

    Haiping Wang, Yuan Liu, Bing Wang, Yujing Sun, Zhen Dong, Wenping Wang, and Bisheng Yang. Freereg: Image- to-point cloud registration leveraging pretrained diffusion models and monocular depth estimators. In International Conference on Learning Representations, 2024. 2, 3, 5

    2024

  33. [41]

    Zero-shot point cloud registration

    Weijie Wang, Guofeng Mei, Bin Ren, Xiaoshui Huang, Fabio Poiesi, Luc Van Gool, Nicu Sebe, and Bruno Lepri. Zero-shot point cloud registration. arXiv preprint arXiv:2312.03032, 2023. 3

    2023 arXiv

  34. [42]

    Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features

    Thomas Wimmer, Peter Wonka, and Maks Ovsjanikov. Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4154–4164, 2024. 3

    2024

  35. [43]

    MAC: max- imal cliques for 3d registration

    Jiaqi Yang, Xiyu Zhang, Peng Wang, Yulan Guo, Kun Sun, Qiao Wu, Shikun Zhang, and Yanning Zhang. MAC: max- imal cliques for 3d registration. IEEE Transactions on Pattern Analysis and Machine Intelligence , 46(12):10645– 10662, 2024. 3, 6, 7

    2024

  36. [44]

    Rpm-net: Robust point matching using learned features

    Zi Jian Yew and Gim Hee Lee. Rpm-net: Robust point matching using learned features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11821– 11830, 2020. 6

    2020

  37. [45]

    REGTR: end-to-end point cloud correspondences with transformers

    Zi Jian Yew and Gim Hee Lee. REGTR: end-to-end point cloud correspondences with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6667–6676, 2022. 2, 3

    2022

  38. [46]

    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. In Advances in Neural In- formation Processing Systems, pages 23872–23884, 2021. 2, 6, 7

    2021

  39. [47]

    Exploring self-supervised learning for 3d point cloud registration

    Mingzhi Yuan, Qiao Huang, Ao Shen, Xiaoshui Huang, and Manning Wang. Exploring self-supervised learning for 3d point cloud registration. IEEE Robotics and Automation Let- ters, 10(1):25–31, 2025. 3

    2025

  40. [48]

    Funkhouser

    Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas A. Funkhouser. 3dmatch: Learning local geometric descriptors from RGB- D reconstructions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 199–208, 2017. 2, 6

    2017

  41. [49]

    A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Pola- nia Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements DINO for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems, 2023. 2, 3, 5

    2023

  42. [50]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision , pages 3813–3824, 2023. 2

    2023

  43. [51]

    Fastmac: Stochastic spectral sampling of correspondence graph

    Yifei Zhang, Hao Zhao, Hongyang Li, and Siheng Chen. Fastmac: Stochastic spectral sampling of correspondence graph. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17857–17867, 2024. 3, 6, 7

    2024

Pith tools

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