REVIEW 5 major objections 5 minor 31 references
SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds
T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read SMART-PC claims that predicting skeletal representations during pretraining lets online test-time adaptation for 3D point clouds update only BatchNorm statistics, without backpropagation, and outperform backprop-based MATE on ModelNet40-C…
desk verdict A genuinely efficient BN-only TTT recipe for point clouds, but the skeleton-specific mechanism is asserted more than demonstrated; deserves a serious referee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the predicted skeletal sphere: each tokenized patch outputs a center $\mathbf{c}_s \in \mathbb{R}^3$ and a radius $r(\mathbf{c}_s)$, and the three loss terms — point-to-sphere, sampling (Chamfer on sphere-surface samples), and negative-radius regularization — jointly enforce that input points lie on the spheres and that spheres broadly cover the shape. This skeletal mesh abstracts the shape as a discrete approximation of the medial axis transform, which is the mechanism that filters corruption noise so that a BatchNorm-statistics-only update can adapt the model online. The skeleton branch shares its encoder with the classification head, and the encoder and decoder features are summed before the classifier.
What would settle it
Replace the skeletal point and radius targets in the skeleton branch with random 4D vectors sampled uniformly in the object's bounding box, keeping every loss and the online BatchNorm-statistics adaptation identical. If ModelNet40-C accuracy remains near the 70.8% of the backpropagation-free variant, the skeletons are not the active ingredient and the gains come from the training setup or backbone; a large drop would confirm the skeletal abstraction carries the robustness.
Extended reading notes
Core claim
The central claim is that predicting skeletal representations during pre-training forces the encoder to capture the intrinsic geometric structure of a point cloud, filtering out high-frequency noise and local distortions that plague surface-level reconstructions. Consequently, the model's features remain meaningful under distribution shift, and test-time adaptation in the online setting can be achieved simply by updating BatchNorm running mean and variance, with no gradient updates at all. The paper demonstrates this by reporting state-of-the-art results on ModelNet40-C and ScanObjectNN-C, including a backpropagation-free variant that nearly matches the full online model and beats MATE-Online by a wide margin on ScanObjectNN-C. A companion experiment swapping several pretrained encoders into BFTT3D's backpropagation-free adaptation shows that the skeleton-pretrained encoder is the best base across all three corrupted datasets, which the authors take as evidence that the skeleton branch, not the adaptation rule, is what supplies robustness.
Load-bearing premise
The load-bearing premise is that the skeleton branch, trained only on clean data, actually learns an abstraction that filters corruption noise, so the extracted features remain meaningful for classification under shifted test distributions; this premise is asserted from the known abstraction property and supported only by indirect accuracy gains, not by a direct measurement of feature robustness.
Editorial extensions
If this is right
- Test-time adaptation for 3D point clouds no longer needs backpropagation in online settings: one forward pass plus a BatchNorm-statistics refresh can adapt to corruption, making real-time deployment at high frame rates feasible.
- Pretext-task choice matters more than adaptation rule: a self-supervised objective that targets abstraction (skeleton) rather than surface reconstruction (MAE) yields features that transfer under corruption, suggesting future TTT methods should be designed around robustness-inducing pretexts.
- Backpropagation-free online adaptation becomes a viable route on ModelNet40-C and ScanObjectNN-C, with the skeleton-pretrained backbone also improving other backpropagation-free adaptation methods such as BFTT3D.
- The same pretrained features are better even in source-only evaluation (no adaptation): SMART-PC-SO beats MATE-SO and the vanilla backbone across all three benchmarks, implying skeleton pretraining alone is a strong robustness intervention.
- Because the online mode resets per corruption type, the gains depend on within-corruption continuity; standard mode with full parameter updates gives only marginal gains over MATE, so the method's advantage is concentrated in the online/streaming setting.
Reading between the lines
- If the abstraction-through-skeleton explanation is correct, then the clarity of the skeleton (e.g., reconstruction fidelity or sparsity) should correlate with downstream robustness; a testable extension would be to train the skeleton branch with different capacities or regularization strengths and measure feature drift under corruption.
- The failure on ShapeNet-C (67.1% vs MATE-Online's 69.1%) suggests the advantage is not universal; a natural extension is to identify corruption types where surface-level reconstruction is actually more informative than skeletal abstraction — for instance, density-decay corruptions that thin the point cloud may destroy the skeletal geometry itself.
- The paper's backpropagation-free variant is essentially DUA with a better-pretrained backbone; the implication is that previously proposed BN-statistics-only methods may have been underperforming largely because their backbones were pretrained with surface-sensitive objectives, and re-evaluating DUA or TENT with skeleton-based pretraining could change their ranking.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes SMART-PC, a test-time training framework for 3D point cloud classification. During pretraining, the model is jointly trained with a classification head and a skeletal branch that predicts skeletal points and radii using point-to-sphere, sampling, and radius regularization losses. At test time, SMART-PC adapts in two modes: a standard mode with backpropagation through the skeletal loss, and an online mode that either updates all parameters or, in a backpropagation-free variant, updates only BatchNorm statistics. The paper reports accuracy gains over MATE on ModelNet40-C and ScanObjectNN-C, a small loss on ShapeNet-C, and a large FPS advantage for the BatchNorm-only online variant. The authors argue that the abstraction property of skeletons makes the learned features robust to corruption, so that lightweight BatchNorm statistics updates suffice for adaptation.
Significance. If the central claim is substantiated, this would be a worthwhile contribution to test-time adaptation for point clouds: a pretext task that improves source-only robustness and supports backpropagation-free online adaptation, with code released and experiments on three standard corruption benchmarks. The Source-only and online improvements on ModelNet40-C and ScanObjectNN-C are suggestive, and the BFTT3D comparison in Table 3 is a reasonable first step toward isolating the effect of pretraining. However, the paper currently does not demonstrate that the skeleton-specific losses, rather than the presence of any auxiliary decoder or the different adaptation augmentation, cause the gains. The loss coefficients are selected on the corrupted test benchmark, and there are no error bars. The backpropagation-free variant, which is the paper's stated headline contribution, does not outperform MATE-Online on ShapeNet-C. These issues leave the main mechanistic claim under-supported, though the approach remains plausible and testable.
major comments (5)
- [Sec. 4.1 and Table 1] The main comparison between SMART-PC-Online and MATE-Online is confounded by the adaptation augmentation: the paper states that MATE employs random masking while SMART-PC employs random rotation. Since online adaptation uses the augmentation to generate test views, the accuracy gap could be driven by the augmentation choice rather than by the skeletal representation. The ablation in Fig. 4 shows that removing augmentation lowers SMART-PC from 47.36% to 46.42% on ScanObjectNN-C, but there is no analogous experiment applying random rotation to MATE or random masking to SMART-PC. Without such a control, the claimed superiority of skeleton-based pretraining is not established.
- [Sec. 4.3, Table 2] The skeletal loss coefficients (λ1, λ2) and the overall configuration are selected using the corrupted test accuracy on ModelNet40-C. Table 2 explicitly reports ``Corrupted Acc(%)'' and the best set (0.3, 1.0, 0.4) is chosen from that column. This is test-set tuning, which can inflate the reported robustness numbers and makes the method's advantage partly a result of selection on the benchmark. The authors should select hyperparameters on a clean validation set or report the full sensitivity surface without using the corrupted test split for model selection.
- [Sec. 3.4, Sec. 4.3, and Table 3] The load-bearing premise, that skeletal abstraction filters corruption noise, is never directly tested. The only evidence is indirect: SMART-PC-SO outperforms MATE-SO, and the BFTT3D comparison in Table 3 shows SMART-PC-SO is the best pretrained model for a fixed backpropagation-free adaptation algorithm. However, these comparisons vary both the auxiliary task and the decoder architecture simultaneously; they do not isolate the skeletal loss. A control experiment with the same architecture but a non-skeletal auxiliary loss, such as point reconstruction via Chamfer distance or a dummy regression task, under the same BN-only adaptation protocol, is needed to determine whether the skeleton-specific losses are responsible for the robustness gains. As written, the paper cannot rule out the possibility that any auxiliary decoder would produce similar results.
- [Sec. 3.5 and Table 1] The abstract and introduction highlight the elimination of backpropagation, but the best online accuracy numbers (SMART-PC-Online, 72.9 on ModelNet40-C, 47.4 on ScanObjectNN-C, 67.1 on ShapeNet-C) are obtained with the backpropagation-based variant, not with the BatchNorm-only variant. The backpropagation-free variant (dagger) achieves 70.8, 46.7, and 65.9 respectively, and it is worse than MATE-Online on ShapeNet-C (65.9 vs. 69.1). The claim of ``superior classification performance'' while ``eliminating backpropagation'' is therefore not supported across all evaluated datasets. The paper should either present the BatchNorm-only variant as the primary result, with the backpropagation-based variant clearly labeled as an upper-bound analysis, or temper the abstract's claims accordingly.
- [All experiments, Tables 1, 5, 6, and 7] No results include error bars or multiple seeds. Several headline comparisons are small, such as SMART-PC-Standard vs. MATE-Standard on ModelNet40-C (63.1 vs. 63.0) and SMART-PC-Online† vs. MATE-Online on the same dataset (70.8 vs. 70.6 from Table 5). Without variance estimates or significance tests, the claimed improvements may not be statistically meaningful. At minimum, the authors should report the mean and standard deviation over at least three independent pretraining runs and adaptation runs.
minor comments (5)
- [Table 1 and Sec. 4.2] The ModelNet40-C numbers for MATE are inconsistent: Table 1 lists MATE-Online* as 70.6, while the text in Sec. 4.2 says 69.6, and Table 5 reports MATE-Online* as 69.6. Similarly, the text says MATE-SO is 53.7% on ModelNet40-C, while Table 1 lists 54.4%. These discrepancies need to be reconciled.
- [Sec. 4.1 and Fig. 3] The FPS comparison in Fig. 3 is not fully apples-to-apples because MATE uses random masking and SMART-PC uses random rotation during adaptation; the different augmentations may themselves have different computational costs. The paper should state whether the reported FPS includes the augmentation cost and ideally compare the methods under the same augmentation protocol.
- [Eq. (14) and Table 2] The total skeletal loss in Eq. (14) is written as ``Lskel = L p2s + λ1 Lsampling + λ2 Lradius'', but the notation is not fully typeset and the equation is hard to read. Also, Table 2's column headers are abbreviated (``Pt2Sphere'', ``Sampling'', ``RadiusReg''); the table would benefit from a legend defining these abbreviations.
- [Sec. 3.5] The description of the online adaptation mode says ``the model is reset at the beginning of adaptation for each new corruption type,'' but it is not specified whether the BatchNorm statistics are reset as well. This detail matters for interpreting the online results across corruption types.
- [Sec. 2 and References] Several references are cited with inconsistent numbering or incomplete metadata (e.g., reference [2] appears as an arXiv preprint without a clear venue, and reference [4] is listed with page number ``1''). A careful reference cleanup is needed.
Circularity Check
No definitional circularity; mild test-set selection of skeleton-loss coefficients inflates headline accuracy.
-
fitted input called prediction
[Section 4.3, Skeleton Loss Coefficients (Table 2) vs. Table 1 headline result.]
"As shown in Tab. 2, the best performance is obtained with the coefficient set (0.3,1.0,0.4), which corresponds to the original settings in the Point2Skeleton paper [13]. This configuration achieves the highest corrupted accuracy of 72.95%, confirming that each skeletal loss term contributes meaningfully to the learning of robust features under corruption."
The coefficients λ1, λ2 of the skeleton loss (Eq. 14) are selected by ranking corrupted accuracy on ModelNet40-C: Table 2 lists accuracies from 67.82 to 72.95 across five coefficient sets, and the best set is retained. The same benchmark then yields the paper's headline SMART-PC-Online result of 72.9% in Table 1. The reported robustness figure is therefore the maximum over the tested configurations on the very test set used for selection, rather than an independent prediction from a pre-specified model. This is a mild form of fitting the evaluation outcome on the target benchmark: the headline number is partially selected rather than predicted.
full rationale
The central derivation is not circular: the skeletal branch is trained with point-to-sphere, sampling, and radius losses (Eqs. 11-14) that depend only on the input point cloud, not on class labels or corrupted-test accuracy, and the classification branch is supervised independently (Eq. 15). The test-time adaptation by BatchNorm-statistics updates is a known external mechanism (AdaBN/DUA, cited as [10,17]) applied to held-out corruptions; it is not defined in terms of the classification accuracy. The self-citations ([1],[2],[3]) appear only as related-work context or as a baseline (SVWA), and they are not used to justify the skeleton-abstraction premise, which is attributed to the external Point2Skeleton work. The main caveat is evaluation-side: λ1 and λ2 were selected using corrupted accuracy on ModelNet40-C (Table 2), so the reported 72.9% is partially tuned on that benchmark. In addition, the comparison with MATE changes both the auxiliary task and the adaptation augmentation (random rotation vs. random masking), and no control auxiliary task isolates the skeleton-specific losses. These are experimental confounds and benchmark-selection issues, not definitional reductions, and they do not collapse the method's derivation into its inputs.
Assumptions & free parameters
free parameters (4)
- lambda_1 (sampling loss weight) =
1.0
- lambda_2 (radius regularization weight) =
0.4
- M (number of FPS centers / patches) =
Not reported
- K (KNN neighborhood size) =
Not reported
assumptions (4)
- domain assumption Skeletal representations abstract shapes and filter high-frequency noise and local distortions.
- ad hoc to paper Optimizing point-to-sphere, sampling, and radius losses without skeleton ground truth yields a meaningful skeleton that supports classification.
- domain assumption BatchNorm statistics encode domain-specific information while network weights encode label knowledge.
- domain assumption ModelNet40-C, ShapeNet-C, and ScanObjectNN-C corruption benchmarks approximate the distribution shifts relevant to real point cloud applications.
Cite this review
Pith. "Pith review of SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds." pith.science (2026). https://pith.science/paper/ZP4Q5JOS
@misc{pith2026250519546,
author = {Pith},
title = {Pith review of: SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds},
year = {2026},
howpublished = {\url{https://pith.science/paper/ZP4Q5JOS}},
note = {Machine review of arXiv:2505.19546}
}
read the original abstract
Test-Time Training (TTT) has emerged as a promising solution to address distribution shifts in 3D point cloud classification. However, existing methods often rely on computationally expensive backpropagation during adaptation, limiting their applicability in real-world, time-sensitive scenarios. In this paper, we introduce SMART-PC, a skeleton-based framework that enhances resilience to corruptions by leveraging the geometric structure of 3D point clouds. During pre-training, our method predicts skeletal representations, enabling the model to extract robust and meaningful geometric features that are less sensitive to corruptions, thereby improving adaptability to test-time distribution shifts. Unlike prior approaches, SMART-PC achieves real-time adaptation by eliminating backpropagation and updating only BatchNorm statistics, resulting in a lightweight and efficient framework capable of achieving high frame-per-second rates while maintaining superior classification performance. Extensive experiments on benchmark datasets, including ModelNet40-C, ShapeNet-C, and ScanObjectNN-C, demonstrate that SMART-PC achieves state-of-the-art results, outperforming existing methods such as MATE in terms of both accuracy and computational efficiency. The implementation is available at: https://github.com/AliBahri94/SMART-PC.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[1]
GeoMask3D: Geometrically Informed Mask Selection for Self-Supervised Point Cloud Learning in 3D
Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Milad Cheraghalikhani, Gustavo Adolfo Vargas Hakim, David Os- owiechi, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Geomask3d: Geometrically informed mask se- lection for self-supervised point cloud learning in 3d.arXiv preprint arXiv:2405.12419, 2024. 1
work page Pith review arXiv 2024
-
[2]
Test-Time Adaptation in Point Clouds: Leveraging Sampling Variation with Weight Averaging
Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sa- har Dastani Oghani, Milad Cheraghalikhani, David Os- owiech, Farzad Beizaee, Ismail Ben Ayed, Christian Desrosiers, et al. Test-time adaptation in point clouds: Leveraging sampling variation with weight averaging.arXiv preprint arXiv:2411.01116, 2024. 2, 7
work page Pith review arXiv 2024
-
[3]
Spectral Informed Mamba for Robust Point Cloud Processing
Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sahar Dastani, Milad Cheraghalikhani, David Osowiechi, Gustavo Adolfo Vargas Hakim, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Spectral informed mamba for robust point cloud processing.arXiv preprint arXiv:2503.04953,
-
[4]
Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015
Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository.arXiv preprint arXiv:1512.03012, 2015. 6, 1
arXiv 2015
-
[5]
Mathematical theory of medial axis transform.pacific jour- nal of mathematics, 181(1):57–88, 1997
Hyeong In Choi, Sung Woo Choi, and Hwan Pyo Moon. Mathematical theory of medial axis transform.pacific jour- nal of mathematics, 181(1):57–88, 1997. 2
work page 1997
-
[6]
Test-time training with masked autoencoders.Advances in Neural Information Processing Systems, 35:29374–29385,
Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders.Advances in Neural Information Processing Systems, 35:29374–29385,
-
[7]
Back to the source: Diffusion- driven adaptation to test-time corruption
Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shel- hamer, and Dequan Wang. Back to the source: Diffusion- driven adaptation to test-time corruption. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11786–11796, 2023. 7
work page 2023
-
[8]
Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generaliza- tion.Advances in Neural Information Processing Systems, 34:2427–2440, 2021. 2, 7
work page 2021
Show all 31 references
-
[9]
Q-mat: Computing medial axis trans- form by quadratic error minimization.ACM Transactions on Graphics (TOG), 35(1):1–16, 2015
Pan Li, Bin Wang, Feng Sun, Xiaohu Guo, Caiming Zhang, and Wenping Wang. Q-mat: Computing medial axis trans- form by quadratic error minimization.ACM Transactions on Graphics (TOG), 35(1):1–16, 2015. 2
2015
-
[10]
Adaptive batch normalization for practical do- main adaptation.Pattern Recognition, 80:109–117, 2018
Yanghao Li, Naiyan Wang, Jianping Shi, Xiaodi Hou, and Jiaying Liu. Adaptive batch normalization for practical do- main adaptation.Pattern Recognition, 80:109–117, 2018. 2
2018
-
[11]
Pointmamba: A simple state space model for point cloud analysis.arXiv preprint arXiv:2402.10739, 2024
Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis.arXiv preprint arXiv:2402.10739, 2024. 1
2024 arXiv
-
[12]
Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation
Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation. InInternational conference on machine learning, pages 6028–6039. PMLR, 2020. 1, 2, 7
2020
-
[13]
Point2skeleton: Learning skele- tal representations from point clouds
Cheng Lin, Changjian Li, Yuan Liu, Nenglun Chen, Yi-King Choi, and Wenping Wang. Point2skeleton: Learning skele- tal representations from point clouds. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4277–4286, 2021. 2, 3, 4, 5, 8
2021
-
[14]
Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808–21820, 2021
Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808–21820, 2021. 2
2021
-
[15]
The norm must go on: Dynamic unsuper- vised domain adaptation by normalization
M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsuper- vised domain adaptation by normalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14765–14775, 2022. 1, 2, 7
2022
-
[16]
Mate: Masked au- toencoders are online 3d test-time learners
M Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Pos- segger, In So Kweon, Kuk-Jin Yoon, et al. Mate: Masked au- toencoders are online 3d test-time learners. InProceedings of the IEEE/CVF International Conference on Co...
2023
-
[17]
Evaluating prediction-time batch normalization for robust- ness under covariate shift.arXiv preprint arXiv:2006.10963,
Zachary Nado, Shreyas Padhy, D Sculley, Alexander D’Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robust- ness under covariate shift.arXiv preprint arXiv:2006.10963,
2006 arXiv
-
[18]
V oronoi skeletons: the- ory and applications
Robert L Ogniewicz and Markus Ilg. V oronoi skeletons: the- ory and applications. InCVPR, pages 63–69, 1992. 2
1992
-
[19]
Masked autoencoders for point cloud self-supervised learning
Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. InEuropean conference on computer vision, pages 604–621. Springer, 2022. 1
2022
-
[20]
Benchmarking and analyzing point cloud classification under corruptions
Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. InIn- ternational Conference on Machine Learning, pages 18559– 18575. PMLR, 2022. 1
2022
-
[21]
Cloudfixer: Test-time adaptation for 3d point clouds via diffusion-guided geometric transformation
Hajin Shim, Changhun Kim, and Eunho Yang. Cloudfixer: Test-time adaptation for 3d point clouds via diffusion-guided geometric transformation. InEuropean Conference on Com- puter Vision, pages 454–471. Springer, 2024. 7
2024
-
[22]
Medial meshes–a compact and accurate representation of medial axis transform.IEEE transactions on visualization and computer graphics, 22(3):1278–1290, 2015
Feng Sun, Yi-King Choi, Yizhou Yu, and Wenping Wang. Medial meshes–a compact and accurate representation of medial axis transform.IEEE transactions on visualization and computer graphics, 22(3):1278–1290, 2015. 2
2015
-
[23]
Benchmarking ro- bustness of 3d point cloud recognition against common cor- ruptions.arXiv preprint arXiv:2201.12296, 2022
Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z Morley Mao. Benchmarking ro- bustness of 3d point cloud recognition against common cor- ruptions.arXiv preprint arXiv:2201.12296, 2022. 1, 6
2022 arXiv
-
[24]
Test-time training with self- supervision for generalization under distribution shifts
Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–
-
[25]
Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data
Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. InProceedings of the IEEE/CVF international conference on computer vision, pages 158...
2019
-
[26]
Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726,
Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726,
2006 arXiv
-
[27]
Backpropagation-free network for 3d test-time adapta- tion
Yanshuo Wang, Ali Cheraghian, Zeeshan Hayder, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt- Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Ha- randi. Backpropagation-free network for 3d test-time adapta- tion. InProceedings of the IEEE/CVF Conference on Com...
-
[28]
Learnable skeleton-aware 3d point cloud sampling
Cheng Wen, Baosheng Yu, and Dacheng Tao. Learnable skeleton-aware 3d point cloud sampling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17671–17681, 2023. 3, 4, 5
2023
-
[29]
V oxel cores: Effi- cient, robust, and provably good approximation of 3d medial axes.ACM Transactions on Graphics (TOG), 37(4):1–13,
Yajie Yan, David Letscher, and Tao Ju. V oxel cores: Effi- cient, robust, and provably good approximation of 3d medial axes.ACM Transactions on Graphics (TOG), 37(4):1–13,
-
[30]
Memo: Test time robustness via adaptation and augmentation.Ad- vances in neural information processing systems, 35:38629– 38642, 2022
Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation.Ad- vances in neural information processing systems, 35:38629– 38642, 2022. 2
2022
-
[31]
Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training.Advances in neural information processing sys- tems, 35:27061–27074, 2022
Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training.Advances in neural information processing sys- tems, 35:27061–27074, 2022. 1 SMART-PC Skeletal ...
2022
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.