Pith. sign in

REVIEW 4 major objections 7 minor 73 references

Topology-Aware Modeling for Unsupervised Simulation-to-Reality Point Cloud Recognition

T0 review · 4 major / 7 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 Topology-aware modeling closes the sim-to-real gap for point cloud recognition without target labels.

desk verdict A solid empirical recipe for point cloud Sim2Real UDA with consistent gains, but some equation and validation issues to fix before I'd trust the exact numbers. read the letter →

arxiv 2506.21165 v1 pith:CO2Y5HVB submitted 2025-06-26 cs.CV

Longkun Zou , Kangjun Liu , Ke Chen , Kailing Guo , Kui Jia , Yaowei Wang This is my paper · ORCID
classification cs.CV
keywords unsuperviseddomainadaptationpointcloudclassificationsimulation-to-realitytopology-awaremodelingimplicitfieldsself-trainingcontrastivelearningFourierpositionalencoding
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

Simulation-to-reality (Sim2Real) adaptation for point cloud recognition is difficult because synthetic point clouds are clean and complete while real sensor scans are noisy and partial. This paper proposes Topology-Aware Modeling (TAM), an unsupervised domain adaptation framework that tries to bridge this gap by learning representations of global spatial topology and of the topological relations among local geometric features, without using any target labels. The framework couples a Fourier positional encoding branch that captures low-level high-frequency 3D structure with a self-supervised task that predicts projection directions and distances of query points to an approximate object surface, aggregating these local implicit features through a Part-based Cloud Graph. A self-training stage with cross-domain contrastive learning is added to resist noisy pseudo-labels. The paper reports consistent accuracy gains over prior methods on PointDA-10, Sim-to-Real, and GraspNetPC-10, suggesting that topology-aware supervision is a viable way to make point cloud classifiers generalize across domains.

What carries the argument

The framework's load-bearing pieces are four. Fourier Positional Encoding is a trigonometric map of point coordinates that exposes low-level high-frequency spatial structure and is used by the global branch. Cross-Domain Mixup (CDMix) forms convex combinations of a source and a target point cloud and penalizes the difference between the prediction on the mix and the blended virtual labels, regularizing the global representation. The local implicit field is a self-supervised decoder that takes a query point and the features of its k nearest points and predicts the projection direction and distance to a triangle approximation of the object surface. The Part-based Cloud Graph (PCG) treats the local implicit features as nodes, connects each node to its K nearest neighbors, applies graph convolution, and max-pools the result into a global descriptor that is aligned with the global feature by a cosine-similarity loss. Self-training with cross-domain contrastive learning then refines the representation with pseudo-labels while reducing the influence of noisy labels.

What would settle it

On a dataset that provides ground-truth object meshes, replace the triangle-approximated distances in Eq. 12 with exact mesh distances; if the accuracy of TAM changes materially, or if a variant trained under heavy occlusion and thin structures loses its advantage over source-only training, the surface-proxy assumption is shown to be load-bearing.

Watch

Extended reading notes

Core claim

The paper's central claim is that the simulation-to-reality gap in point cloud classification is best reduced by exposing and modeling topology, not by aligning global semantics alone. It argues that high-frequency 3D spatial structure, obtained through Fourier positional encoding of raw coordinates, is domain-insensitive and complements the local geometric information that deep networks tend to overfit. To capture local geometry in a domain-agnostic way, the paper learns implicit fields that encode, for each query point, the direction and distance of its projection onto an approximate surface formed by the nearest input points; these per-part features are then organized into a Part-based Cloud Graph whose node relations, pooled into a global descriptor, regularize the global feature. Finally, a self-training loop with category-level cross-domain contrastive learning assigns pseudo-labels to confident target samples while softening the alignment to reduce sensitivity to label noise. On PointDA-10 the method reaches 62.1% and 59.3% on the two Sim2Real tasks, on Sim-to-Real 76.19% and 63.85%, and on GraspNetPC-10 an average of 90.9% over the four adaptation scenarios.

Load-bearing premise

The framework assumes that the distance from a query point to the nearest triangle formed by its M nearest input points is a faithful, reliable approximation to the distance to the true object surface, and that the resulting projection direction is reliable supervision.

Editorial extensions

If this is right

  • On PointDA-10, TAM with self-training reaches 62.1% and 59.3% on the ModelNet-10 to ScanNet-10 and ShapeNet-10 to ScanNet-10 tasks, exceeding the numbers reported for prior methods.
  • On Sim-to-Real, TAM reaches 76.19% on ModelNet-11 to ScanObjectNN-11 and 63.85% on ShapeNet-9 to ScanObjectNN-9.
  • On GraspNetPC-10, TAM averages 90.9% across the four sensor-transfer scenarios, with 97.8% on synthetic-to-Kinect and 82.6% on Kinect-to-Realsense.
  • Ablations show that CDMix and local implicit fields are complementary: each alone gives modest gains, together they provide most of the improvement, and the contrastive self-training adds the remainder.
  • Replacing the Fourier-encoded Point-PN global backbone with DGCNN, PointNet, or Point-NN lowers Sim2Real accuracy, indicating that high-frequency positional encoding is the main carrier of the global topology signal.

Reading between the lines

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

  • Because the local implicit supervision only needs an approximate surface proxy, the same triangle-based distance could be used to adapt other point cloud tasks, such as part segmentation or single-view completion, wherever local geometry is shared across domains.
  • The paper's evidence that high-frequency global structure transfers across domains suggests that Fourier positional encoding could be inserted into existing UDA methods as a plug-in module, potentially improving them without changing their alignment objectives.
  • A direct test of the framework's core premise would be to vary the number of query points and neighbors in the PCG module while holding everything else fixed; if the accuracy gain tracks the quality of the surface proxy rather than the graph depth, the implicit-field supervision is the active ingredient.
  • The reported noise-robustness table suggests TAM may also serve as a label-free calibration step for classifiers deployed on new depth sensors, before any target labels become available.
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 / 7 minor

Summary. The paper proposes Topology-Aware Modeling (TAM), an unsupervised domain adaptation method for Sim2Real point cloud classification. The framework combines Fourier positional encoding to capture global high-frequency spatial structure, a cross-domain mixup regularizer (CDMix), a self-supervised branch that learns local geometric implicit fields with Part-based Cloud Graph (PCG) aggregation, and self-training with cross-domain contrastive learning (CLST). Experiments on PointDA-10, Sim-to-Real, and GraspNetPC-10 report consistent improvements over prior methods, supported by ablations, error bars over three seeds, A-distance analysis, noise-robustness tests, and computational cost comparisons.

Significance. If the reported results hold, the paper makes a strong empirical contribution: it achieves state-of-the-art accuracy on three public Sim2Real benchmarks, and the ablation study in Table IV gives evidence that CDMix, the local implicit SSL branch, and CLST each contribute to the final accuracy. The A-distance analysis in Fig. 6 and the cost comparison in Table VIII are useful additions. However, the manuscript as written contains mathematical errors in the loss definitions and an underspecified hyperparameter selection protocol for the self-training threshold. These issues prevent me from endorsing the reported numbers as stated and require correction before the central claim can be accepted.

major comments (4)
  1. [III-B1, Eq. (14)] The self-supervised local implicit loss is written as L_imp = -1/N * sum_c ||ConCat(ê_n, ê_d) - ConCat(n, d)||_2. Minimizing a negative L2 term drives the predicted projection direction and distance away from the supervision targets; a regression loss of this form should have a positive sign. Because this loss supervises the entire local-implicit branch, and the ablation in Table IV attributes part of the gain to that branch, the derivation as printed does not support the method. Please correct the sign and confirm that the reported results use the positive L2 loss.
  2. [III-C, Eq. (26)] The category-based cross-domain contrastive loss is ill-posed as written: the denominator is sum_{i=1}^{n_t} φ(z_i^t, z^{s+}) + φ(z_i^t, z^{s-}), where the second term uses an undefined index i outside the summation and there is no summation over negatives. A valid per-sample contrastive loss would be -log[φ(z_i^t, z^{s+}) / (φ(z_i^t, z^{s+}) + sum_{z^-} φ(z_i^t, z^-))] with multiple negatives. Since CLST is one of the three components ablated in Table IV, please reformulate Eq. (26) precisely and specify how the positive and negative source features are defined for each target anchor.
  3. [IV-B and Fig. 7] Algorithm 1 initializes the self-training threshold to θ0 = 0.8 and the text says this is the default, yet Fig. 7 shows that M10→S*10 and M11→SO*11 perform best near θ0 = 0.8 while S10→S*10 and S9→SO*9 perform best near θ0 = 0.7. The paper does not describe a validation procedure for selecting these per-task thresholds. If the values were chosen using the target test accuracy reported in Tables I–III, the comparison is not made under a standard UDA protocol. Please either fix a single threshold, describe a validation-based selection, or report the sensitivity analysis without claiming the best per-task value.
  4. [III-B1, Eq. (12)] The local implicit branch assumes that the distance from a query point to the nearest triangle formed by its M nearest input points is a faithful approximation of the distance to the true object surface. On the target domains, which are described in Sec. IV-A as partial, noisy, and incomplete, the M nearest points can span holes or connect outlier points, so the resulting projection directions and distances may not correspond to the physical surface. The noise-robustness experiment in Table VII evaluates accuracy after adding test-time noise and does not assess the fidelity of this triangle proxy during training. Given that the SSL branch contributes the accuracy gains shown in the Table IV ablations, please provide evidence, either analytical or experimental, that the proxy remains reliable on noisy and partial target clouds.
minor comments (7)
  1. [IV-C heading] The section heading “Comparsion to the State-of-the-art” contains a typo and should read “Comparison.”
  2. [I, Contributions] One contribution bullet states that experiments were conducted on “two public benchmarks”, but the paper evaluates on PointDA-10, Sim-to-Real, and GraspNetPC-10; please make the count consistent.
  3. [III-C, Eqs. (23)-(24)] The prediction p^s_{i,c} is defined as the product p^{1s}_{i,c} p^{2s}_{i,c} of two probability vectors. The elementwise product of two probability distributions is not itself a normalized probability distribution; please clarify whether an ensemble average or another combination rule is intended.
  4. [III-B1] The query point sampling hyperparameters are not reported: the voxel resolution l, the distance range [D_l, D_u], the number of nearest points M, and the number of query points N are all left unspecified. Without these values the self-supervised branch is not reproducible from the paper alone.
  5. [III-A2, Eq. (10) and Algorithm 1] In Algorithm 1, line 5 writes “min_{Φ_REG} L_mix with (P^s_i, y^s_i) and P^t_i”, but CDMix in Eq. (10) uses only the point clouds and virtual labels; please clarify whether the source label y^s_i is actually used and how the virtual labels are formed in the algorithm listing.
  6. [III-B2, Eq. (22)] The notation around Φ_reg is confusing: Eq. (22) writes Φ_REG = Φ_cls(z_reg) ∘ Φ_reg(z_g), but the text then says Φ_reg denotes the global feature encoder Φ_g regularized by CDMix and PCG. Please define Φ_reg and z_reg consistently with its use in the equation.
  7. [Abstract] The source code URL in the abstract ends with “TAG.git”, which does not match the name TAM; please update the link if the repository name is different.

Circularity Check

1 steps flagged · score 4.0 of 10

Per-benchmark threshold selection on the target test sets makes some reported accuracies partly fitted; the core method derivation is otherwise self-contained.

  1. fitted input called prediction [Section IV-D, 'Influence of confidence threshold θ in self-training' (Fig. 7)]
    "Therefore, to obtain a suitable threshold, we conducted multiple experiments with different thresholds on four Sim2Real settings. As shown in Fig. 7, we found that when the threshold θ is close to 0.8, M10→S*10 and M11→SO*11 achieve the best performance, and when the threshold θ is close to 0.7, S10→S*10 and S9→SO*9 achieve the best performance."

    The reported PointDA-10 and Sim-to-Real accuracies are produced after choosing the self-training confidence threshold θ separately for each task, with the choice based on the best accuracy on the same target test benchmarks whose accuracies are later reported. For each task, the table entry is therefore a selected maximum over the evaluated thresholds rather than an independent evaluation at a fixed, pre-specified setting. Algorithm 1 initializes θ0=0.8, but the per-task 0.7 exception is introduced from test-set accuracy, so the fitted threshold is part of the evaluation loop and the reported 'consistent improvements' are partly fitted numbers.

full rationale

Apart from the threshold selection, the derivation chain is not circular. The self-supervised local-implicit branch (Eqs. 12-14) uses a triangle-proxy distance to the approximate surface computed from each input cloud; this proxy is a heuristic approximation that may be biased by noise or occlusion, but it is not equivalent to the classification output and does not use target labels. CDMix (Eq. 10) and SPST (Eqs. 24-25) use the model's own current predictions as virtual or pseudo labels; this is self-referential but standard consistency/self-training, and the final accuracy is not equal to those labels by construction. Fourier positional encoding, the PCG module, and the contrastive loss are grounded in cited prior work without a load-bearing self-citation chain. The one concrete fitted-input concern is the per-benchmark threshold tuning in Fig. 7, which inflates the specific reported numbers but does not by itself invalidate the qualitative comparison or the independent content of the proposed components.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four assumptions, none independently verified outside the benchmarks. Hyperparameters include two Fourier encoding scales, mixup beta, loss weights, and a per-task pseudo-label threshold. No new physical entities are introduced.

free parameters (6)
  • Fourier positional encoding scale alpha = 100
    Empirically set in Sec. IV-B, controls frequency and scale of the positional embedding.
  • Fourier positional encoding frequency beta = 500
    Empirically set in Sec. IV-B, paired with alpha to determine high-frequency sensitivity.
  • CDMix beta distribution parameter kappa = 2
    Controls mixup strength in Eq. (10); set empirically in Sec. IV-B.
  • Self-training confidence threshold theta_0 = 0.8 or 0.7 depending on task
    Fig. 7 shows multiple threshold experiments and selects 0.8 for M10->S*10 and M11->SO*11, and 0.7 for S10->S*10 and S9->SO*9.
  • Self-paced threshold increment epsilon = 0.005
    Set in Sec. IV-B; governs how fast the confidence threshold grows.
  • Loss weights lambda_t, lambda_cdc, lambda_imp, lambda_mix, lambda_sim = 1, 1, 1, 1, 0.1
    Trade-off coefficients in Eq. (27), empirically set in Sec. IV-B.
assumptions (4)
  • domain assumption Fourier positional encoding captures global spatial topology that is invariant across source and target domains
    Central motivation in Sec. I and Sec. III-A; not proved, only empirically supported by Table VI.
  • domain assumption Local geometric structures and their implicit fields are largely shared across domains
    Sec. III-B states local geometry is more likely shared; this assumption is borrowed from PDG [36].
  • ad hoc to paper A query point's distance to nearby triangles approximates distance to the underlying surface
    Sec. III-B1, Eq. (12); needed to generate self-supervised projection signals, but no error analysis is given.
  • domain assumption Mixing source and target clouds with virtual labels regularizes the latent space
    CDMix objective in Eq. (11) relies on the linearity assumption from interpolation consistency training [56].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topology-Aware Modeling for Unsupervised Simulation-to-Reality Point Cloud Recognition." pith.science (2026). https://pith.science/paper/CO2Y5HVB

@misc{pith2026250621165,
  author       = {Pith},
  title        = {Pith review of: Topology-Aware Modeling for Unsupervised Simulation-to-Reality Point Cloud Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CO2Y5HVB}},
  note         = {Machine review of arXiv:2506.21165}
}
read the original abstract

Learning semantic representations from point sets of 3D object shapes is often challenged by significant geometric variations, primarily due to differences in data acquisition methods. Typically, training data is generated using point simulators, while testing data is collected with distinct 3D sensors, leading to a simulation-to-reality (Sim2Real) domain gap that limits the generalization ability of point classifiers. Current unsupervised domain adaptation (UDA) techniques struggle with this gap, as they often lack robust, domain-insensitive descriptors capable of capturing global topological information, resulting in overfitting to the limited semantic patterns of the source domain. To address this issue, we introduce a novel Topology-Aware Modeling (TAM) framework for Sim2Real UDA on object point clouds. Our approach mitigates the domain gap by leveraging global spatial topology, characterized by low-level, high-frequency 3D structures, and by modeling the topological relations of local geometric features through a novel self-supervised learning task. Additionally, we propose an advanced self-training strategy that combines cross-domain contrastive learning with self-training, effectively reducing the impact of noisy pseudo-labels and enhancing the robustness of the adaptation process. Experimental results on three public Sim2Real benchmarks validate the effectiveness of our TAM framework, showing consistent improvements over state-of-the-art methods across all evaluated tasks. The source code of this work will be available at https://github.com/zou-longkun/TAG.git.

Figures

Figures reproduced from arXiv: 2506.21165 by the authors.

Figure 1
Figure 1. Illustration of alignment based on synergizing topological configura [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed TAM framework, which consists of two branches: the top branch captures low-level global high-frequency 3D spatial [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Illustration of query points sampling and surface projection. Black [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Illustration of graph construction of local implicit fields. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: (a) A-distance of different aggregated features between source and target domain. (b) Discrimination ability of global-level features regularized by different aggregated part-level feature measured by classification accuracy on test data of source and target domain. in…
Figure 8
Figure 8. Figure 8: Saliency map visualization of various comparison methods under the [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Confusion matrices of classifying testing samples on target domain. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: The t-SNE visualization of feature distribution on the target domain. Colors indicate different classes. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 60 canonical work pages

  1. [1]

    Pointgl: a simple global-local framework for efficient point cloud analysis,

    J. Li, J. Wang, and T. Xu, “Pointgl: a simple global-local framework for efficient point cloud analysis,”IEEE Transactions on Multimedia, 2024

    work page 2024

  2. [2]

    Domain adaptive lidar point cloud segmentation with 3d spatial consistency,

    A. Xiao, D. Guan, X. Zhang, and S. Lu, “Domain adaptive lidar point cloud segmentation with 3d spatial consistency,”IEEE Transactions on Multimedia, 2023

    work page 2023

  3. [3]

    Cmnet: Component-aware matching network for few-shot point cloud classification,

    L. Lai, J. Chen, G. Lin, and Q. Wu, “Cmnet: Component-aware matching network for few-shot point cloud classification,”IEEE Transactions on Multimedia, 2024

    work page 2024

  4. [4]

    Geometric back-projection net- work for point cloud classification,

    S. Qiu, S. Anwar, and N. Barnes, “Geometric back-projection net- work for point cloud classification,”IEEE Transactions on Multimedia, vol. 24, pp. 1943–1955, 2021

    work page 1943

  5. [5]

    Cattrack: Single-stage category-level 6d object pose tracking via convolution and vision trans- former,

    S. Yu, D. Zhai, Y . Xia, D. Li, and S. Zhao, “Cattrack: Single-stage category-level 6d object pose tracking via convolution and vision trans- former,”IEEE Transactions on Multimedia, vol. 26, pp. 1665–1680, 2023

    work page 2023

  6. [6]

    Real-time 3d single object tracking with transformer,

    J. Shan, S. Zhou, Y . Cui, and Z. Fang, “Real-time 3d single object tracking with transformer,”IEEE Transactions on Multimedia, vol. 25, pp. 2339–2353, 2022

    work page 2022

  7. [7]

    Vpfnet: Improving 3d object detection with virtual point based lidar and stereo data fusion,

    H. Zhu, J. Deng, Y . Zhang, J. Ji, Q. Mao, H. Li, and Y . Zhang, “Vpfnet: Improving 3d object detection with virtual point based lidar and stereo data fusion,”IEEE Transactions on Multimedia, vol. 25, pp. 5291–5304, 2022

    work page 2022

  8. [8]

    Centertube: Tracking multiple 3d objects with 4d tubelets in dynamic point clouds,

    H. Liu, Y . Ma, Q. Hu, and Y . Guo, “Centertube: Tracking multiple 3d objects with 4d tubelets in dynamic point clouds,”IEEE Transactions on Multimedia, vol. 25, pp. 8793–8804, 2023

    work page 2023

Show all 73 references
  1. [9]

    Shapenet: An information- rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Suet al., “Shapenet: An information- rich 3d model repository,”ArXiv, vol. 1512.03012, 2015

    2015 arXiv

  2. [10]

    Modelnet: Towards a datacenter emulation environment,

    K. V . Vishwanath, D. Gupta, A. Vahdat, and K. Yocum, “Modelnet: Towards a datacenter emulation environment,” in2009 IEEE Ninth International Conference on Peer-to-Peer Computing, 2009, pp. 81–82

    2009

  3. [11]

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

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inIEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 652–660

    2017

  4. [12]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” inAdv. Neural Inform. Process. Syst., 2017, pp. 5099–5108

    2017

  5. [13]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,”Acm Transactions On Graphics (tog), vol. 38, no. 5, pp. 1–12, 2019

    2019

  6. [14]

    Kpconv: Flexible and deformable convolution for point clouds,

    H. Thomas, C. R. Qi, J. Deschaud, B. Marcotegui, F. Goulette, and L. J. Guibas, “Kpconv: Flexible and deformable convolution for point clouds,” inInt. Conf. Comput. Vis., 2019, pp. 6411–6420

    2019

  7. [15]

    Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling,

    X. Yan, C. Zheng, Z. Li, S. Wang, and S. Cui, “Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling,”IEEE Conf. Comput. Vis. Pattern Recog., 2020

    2020

  8. [16]

    Pct: Point cloud transformer,

    M. Guo, J. Cai, Z. Liu, T. Mu, R. R. Martin, and S. Hu, “Pct: Point cloud transformer,”Computational Visual Media, vol. 7, no. 2, pp. 187–199, 2021

    2021

  9. [17]

    V oxelnet: End-to-end learning for point cloud based 3d object detection,

    Y . Zhou and O. Tuzel, “V oxelnet: End-to-end learning for point cloud based 3d object detection,” inIEEE Conf. Comput. Vis. Pattern Recog., 2018, pp. 4490–4499

    2018

  10. [18]

    Modeling point clouds with self-attention and gumbel subset sampling,

    J. Yang, Q. Zhang, B. Ni, L. Li, J. Liu, M. Zhou, and Q. Tian, “Modeling point clouds with self-attention and gumbel subset sampling,”IEEE Conf. Comput. Vis. Pattern Recog., 2019

    2019

  11. [19]

    So-net: Self-organizing network for point cloud analysis,

    J. Li, B. M. Chen, and G. H. Lee, “So-net: Self-organizing network for point cloud analysis,” inIEEE Conf. Comput. Vis. Pattern Recog., 2018

    2018

  12. [20]

    Spherical cnns,

    T. Cohen, M. Geiger, J. K ¨ohler, and M. Welling, “Spherical cnns,”ArXiv, vol. 1801.10130, 2018

    2018 arXiv

  13. [21]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inIEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 5828– 5839

    2017

  14. [22]

    Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data,

    M. A. Uy, Q. Pham, B. Hua, T. Nguyen, and S. Yeung, “Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data,” inIEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 1588–1597

    2019

  15. [23]

    Classification of object point clouds: A practical perspective,

    Z. Xu, K. Chen, C. Ding, Y . Wang, and K. Jia, “Classification of object point clouds: A practical perspective,”arXiv preprint arXiv:2012.10042, 2020

    2012 arXiv

  16. [24]

    Transferable representation learning with deep adaptation networks,

    M. Long, Y . Cao, Z. Cao, J. Wang, and M. I. Jordan, “Transferable representation learning with deep adaptation networks,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, pp. 3071–3085, 2019

    2019

  17. [25]

    Contrastive adaptation network for unsupervised domain adaptation,

    G. Kang, L. Jiang, Y . Yang, and A. G. Hauptmann, “Contrastive adaptation network for unsupervised domain adaptation,” inIEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 4888–4897

    2019

  18. [26]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavio- lette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,”J. Mach. Learn. Res., vol. 17, pp. 2096–2030, 2016

    2016

  19. [27]

    Maximum classi- fier discrepancy for unsupervised domain adaptation,

    K. Saito, K. Watanabe, Y . Ushiku, and T. Harada, “Maximum classi- fier discrepancy for unsupervised domain adaptation,” inIEEE Conf. Comput. Vis. Pattern Recog., 2018, pp. 3723–3732. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, MARCH 2024 14

    2018

  20. [28]

    Learning semantic represen- tations for unsupervised domain adaptation,

    S. Xie, Z. Zheng, L. Chen, and C. Chen, “Learning semantic represen- tations for unsupervised domain adaptation,” inInt. Conf. Mach. Learn., 2018, pp. 5423–5432

    2018

  21. [29]

    Transferrable prototypical networks for unsupervised domain adaptation,

    Y . Pan, T. Yao, Y . Li, Y . Wang, C. Ngo, and T. Mei, “Transferrable prototypical networks for unsupervised domain adaptation,” inIEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 2234–2242

    2019

  22. [30]

    Pointdan: A multi- scale 3d domain adaption network for point cloud representation,

    C. Qin, H. You, L. Wang, C.-C. J. Kuo, and Y . Fu, “Pointdan: A multi- scale 3d domain adaption network for point cloud representation,” in Adv. Neural Inform. Process. Syst., 2019, pp. 7192–7203

    2019

  23. [31]

    Self-supervised learning for domain adaptation on point clouds,

    I. Achituve, H. Maron, and G. Chechik, “Self-supervised learning for domain adaptation on point clouds,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 123– 133

    2021

  24. [32]

    Geometry-aware self-training for unsupervised domain adaptation on object point clouds,

    L. Zou, H. Tang, K. Chen, and K. Jia, “Geometry-aware self-training for unsupervised domain adaptation on object point clouds,” inInt. Conf. Comput. Vis., 2021, pp. 6403–6412

    2021

  25. [33]

    Domain adaptation on point clouds via geometry-aware implicits,

    Y . Shen, Y . Yang, M. Yan, H. Wang, Y . Zheng, and L. J. Guibas, “Domain adaptation on point clouds via geometry-aware implicits,” in IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 7223–7232

    2022

  26. [34]

    Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels,

    H. Fan, X. Chang, W. Zhang, Y . Cheng, Y . Sun, and M. Kankanhalli, “Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels,” inIEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 6377–6386

    2022

  27. [35]

    Quasi-balanced self- training on noise-aware synthesis of object point clouds for closing domain gap,

    Y . Chen, Z. Wang, L. Zou, K. Chen, and K. Jia, “Quasi-balanced self- training on noise-aware synthesis of object point clouds for closing domain gap,” inEur. Conf. Comput. Vis., 2022, pp. 728–745

    2022

  28. [36]

    Learning generalizable part-based feature representation for 3d point clouds,

    X. Wei, X. Gu, and J. Sun, “Learning generalizable part-based feature representation for 3d point clouds,” inAdvances in Neural Information Processing Systems, 2022

    2022

  29. [37]

    Deep convolutional networks do not classify based on global object shape,

    N. Baker, H. Lu, G. Erlikhman, and P. J. Kellman, “Deep convolutional networks do not classify based on global object shape,”PLoS computa- tional biology, vol. 14, no. 12, p. e1006613, 2018

    2018

  30. [38]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel, “Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,”arXiv preprint arXiv:1811.12231, 2018

    2018 arXiv

  31. [39]

    Connecting the dots with landmarks: Discriminatively learning domain-invariant features for unsupervised domain adaptation,

    B. Gong, K. Grauman, and F. Sha, “Connecting the dots with landmarks: Discriminatively learning domain-invariant features for unsupervised domain adaptation,” inInt. Conf. Mach. Learn., 2013, pp. 222–230

    2013

  32. [40]

    Parts of recognition,

    D. D. Hoffman and W. A. Richards, “Parts of recognition,”Cognition, vol. 18, no. 1-3, pp. 65–96, 1984

    1984

  33. [41]

    Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis,

    R. Zhang, L. Wang, Y . Wang, P. Gao, H. Li, and J. Shi, “Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis,”arXiv preprint arXiv:2303.08134, 2023

    2023 arXiv

  34. [42]

    Reconstructing continuous distributions of 3d protein structure from cryo-em images,

    E. D. Zhong, T. Bepler, J. H. Davis, and B. Berger, “Reconstructing continuous distributions of 3d protein structure from cryo-em images,” arXiv preprint arXiv:1909.05215, 2019

    1909 arXiv

  35. [43]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  36. [44]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” Advances in Neural Information Processing Systems, vol. 33, pp....

    2020

  37. [45]

    Metasets: Meta- learning on point sets for generalizable representations,

    C. Huang, Z. Cao, Y . Wang, J. Wang, and M. Long, “Metasets: Meta- learning on point sets for generalizable representations,” inIEEE Conf. Comput. Vis. Pattern Recog., 2021, pp. 8863–8872

    2021

  38. [46]

    Improving semantic analysis on point clouds via auxiliary supervision of local geometric priors,

    L. Tang, K. Chen, C. Wu, Y . Hong, K. Jia, and Z. Yang, “Improving semantic analysis on point clouds via auxiliary supervision of local geometric priors,”ArXiv, vol. 2001.04803, 2020

    2001 arXiv

  39. [47]

    Pointcnn: Convolution on x-transformed points,

    Y . Li, R. Bu, M. Sun, W. Wu, X. Di, and B. Chen, “Pointcnn: Convolution on x-transformed points,” inAdv. Neural Inform. Process. Syst., 2018, pp. 820–830

    2018

  40. [48]

    Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis,

    C. Chen, G. Li, R. Xu, T. Chen, M. Wang, and L. Lin, “Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis,” inIEEE Conf. Comput. Vis. Pattern Recog., 2019

    2019

  41. [49]

    mixup: Beyond empirical risk minimization,

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,”ArXiv, vol. 1710.09412, 2017

    2017 arXiv

  42. [50]

    Manifold mixup: Better representations by interpolating hidden states,

    V . Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez-Paz, and Y . Bengio, “Manifold mixup: Better representations by interpolating hidden states,” inInt. Conf. Mach. Learn.PMLR, 2019, pp. 6438–6447

    2019

  43. [51]

    Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,” inInt. Conf. Comput. Vis., 2019, pp. 6023–6032

    2019

  44. [52]

    Puzzle mix: Exploiting saliency and local statistics for optimal mixup,

    J. Kim, W. Choo, and H. O. Song, “Puzzle mix: Exploiting saliency and local statistics for optimal mixup,” inInt. Conf. Mach. Learn.PMLR, 2020, pp. 5275–5285

    2020

  45. [53]

    Pointmixup: Augmentation for point clouds,

    Y . Chen, V . T. Hu, E. Gavves, T. Mensink, P. Mettes, P. Yang, and C. G. Snoek, “Pointmixup: Augmentation for point clouds,” inEur. Conf. Comput. Vis.Springer, 2020, pp. 330–345

    2020

  46. [54]

    Pointaugment: an auto-augmentation framework for point cloud classification,

    R. Li, X. Li, P. Heng, and C. Fu, “Pointaugment: an auto-augmentation framework for point cloud classification,” inIEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 6378–6387

    2020

  47. [55]

    Pointcutmix: Regularization strategy for point cloud classification,

    J. Zhang, L. Chen, B. Ouyang, B. Liu, J. Zhu, Y . Chen, Y . Meng, and D. Wu, “Pointcutmix: Regularization strategy for point cloud classification,”Neurocomputing, vol. 505, pp. 58–67, 2022

    2022

  48. [56]

    Interpolation consistency training for semi-supervised learning,

    V . Verma, K. Kawaguchi, A. Lamb, J. Kannala, Y . Bengio, and D. Lopez-Paz, “Interpolation consistency training for semi-supervised learning,”Neural Networks, vol. 145, pp. 90–106, 2022

    2022

  49. [57]

    Vision gnn: An image is worth graph of nodes,

    K. Han, Y . Wang, J. Guo, Y . Tang, and E. Wu, “Vision gnn: An image is worth graph of nodes,”ArXiv, vol. 2206.00272, 2022

    2022 arXiv

  50. [58]

    Deepgcns: Can gcns go as deep as cnns?

    G. Li, M. Muller, A. Thabet, and B. Ghanem, “Deepgcns: Can gcns go as deep as cnns?” inInt. Conf. Comput. Vis., 2019, pp. 9267–9276

    2019

  51. [59]

    Learning cross-modal contrastive features for video domain adaptation,

    D. Kim, Y . Tsai, B. Zhuang, X. Yu, S. Sclaroff, K. Saenko, and M. Chandraker, “Learning cross-modal contrastive features for video domain adaptation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13 618–13 627

    2021

  52. [60]

    Learning a nonlinear embedding by preserving class neighbourhood structure,

    R. Salakhutdinov and G. Hinton, “Learning a nonlinear embedding by preserving class neighbourhood structure,” inArtificial intelligence and statistics. PMLR, 2007, pp. 412–419

    2007

  53. [61]

    Improving generalization via scalable neighborhood component analysis,

    Z. Wu, A. A. Efros, and S. X. Yu, “Improving generalization via scalable neighborhood component analysis,” inProceedings of the european conference on computer vision (ECCV), 2018, pp. 685–701

    2018

  54. [62]

    Supervised contrastive learn- ing,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,”Advances in neural information processing systems, vol. 33, pp. 18 661–18 673, 2020

    2020

  55. [63]

    Graspnet-1billion: A large- scale benchmark for general object grasping,

    H. Fang, C. Wang, M. Gou, and C. Lu, “Graspnet-1billion: A large- scale benchmark for general object grasping,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 444–11 453

    2020

  56. [64]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” ArXiv, vol. 1412.6980, 2014

    2014 arXiv

  57. [65]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational conference on machine learning. pmlr, 2015, pp. 448–456

    2015

  58. [66]

    Self-supervised deep learning on point clouds by reconstructing space,

    J. Sauder and B. Sievers, “Self-supervised deep learning on point clouds by reconstructing space,” inAdv. Neural Inform. Process. Syst., 2019, pp. 12 962–12 972

    2019

  59. [67]

    Self-distillation for unsupervised 3d domain adaptation,

    A. Cardace, R. Spezialetti, P. Z. Ramirez, S. Salti, and L. Di Stefano, “Self-distillation for unsupervised 3d domain adaptation,” inProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 4166–4177

    2023

  60. [68]

    Domain adaptive sampling for cross- domain point cloud recognition,

    Z. Wang, W. Li, and D. Xu, “Domain adaptive sampling for cross- domain point cloud recognition,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 12, pp. 7604–7615, 2023

    2023

  61. [69]

    Synergizing contrastive learning and optimal transport for 3d point cloud domain adaptation,

    S. Katageri, A. De, C. Devaguptapu, V . Prasad, C. Sharma, and M. Kaul, “Synergizing contrastive learning and optimal transport for 3d point cloud domain adaptation,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 2942–2951

    2024

  62. [70]

    A theory of learning from different domains,

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,”Mach. Learn., vol. 79, pp. 151–175, 2010

    2010

  63. [71]

    Domain adaptation: Learning bounds and algorithms,

    Y . Mansour, M. Mohri, and A. Rostamizadeh, “Domain adaptation: Learning bounds and algorithms,”ArXiv, vol. 0902.3430, 2009

    2009 arXiv

  64. [72]

    Pointcloud saliency maps,

    T. Zheng, C. Chen, J. Yuan, B. Li, and K. Ren, “Pointcloud saliency maps,” inInt. Conf. Comput. Vis., 2019, pp. 1598–1606

    2019

  65. [73]

    Visualizing data using t-sne,

    L. v. d. Maaten and G. Hinton, “Visualizing data using t-sne,”J. Mach. Learn. Res., vol. 9, pp. 2579–2605, 2008

    2008

Pith tools

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