Pith. sign in

REVIEW 5 major objections 6 minor 38 references

Whitening before attention and recoloring after it makes few-shot 3D prototype generation deterministic and state of the art.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 16:27 UTC pith:FLVKZINX

load-bearing objection WARM is a sensible incremental contribution with a useful FPS-instability diagnostic, but the SOTA claim is overreaching given only two baselines, no variance reporting, and an undefined Q-K distance metric. the 5 major comments →

arxiv 2509.13907 v3 pith:FLVKZINX submitted 2025-09-17 cs.CV

White Aggregation and Restoration for Few-shot 3D Point Cloud Semantic Segmentation

Jiyun Im , SuBeen Lee , Miso Lee , Jae-Pil Heo This is my paper
classification cs.CV
keywords few-shot learning3D point cloudsemantic segmentationprototype generationcross-attentionwhiteningZCAFarthest Point Sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the standard method for building prototypes in few-shot 3D point cloud segmentation—Farthest Point Sampling—is unstable because its output depends on a random seed, and this seed variation can swing performance dramatically. It proposes replacing FPS with cross-attention, but shows that vanilla cross-attention fails because learnable prototypical tokens and support point features live in very different distributions. The proposed WARM module applies ZCA whitening to support features before attention to align them with the tokens, then applies the inverse coloring after attention to restore the original statistics. The result is a deterministic, trainable prototype generator that achieves state-of-the-art results on S3DIS and competitive results on ScanNet, while also making attention maps smoother and more semantically focused.

Core claim

The central claim is that a large distributional gap between learnable prototypical tokens and support point features—quantified by intra-class, inter-class, and instance dispersion metrics—prevents vanilla cross-attention from producing semantically meaningful prototypes. By temporarily whitening support features (zero mean, decorrelated channels) before cross-attention and restoring the original mean and covariance afterward, WARM aligns queries and keys in a shared space. This allows tokens to attend to semantically related regions rather than to a few nearby points, yielding deterministic prototypes that are fully differentiable and seed-independent. The paper reports state-of-the-art pe

What carries the argument

The White Aggregation and Restoration Module (WARM): a cross-attention layer sandwiched between a ZCA whitening step and its inverse coloring. Whitening (Eq. 10) removes per-class mean and covariance from support features so they match the distribution of learnable prototypical tokens; the cross-attention (Eq. 11) aggregates the whitened features into prototypes; coloring (Eq. 12) re-injects the original mean and covariance to preserve instance-specific statistics. This transforms prototype generation from a fixed, seed-dependent sampling algorithm into a trainable, deterministic process.

Load-bearing premise

The method assumes that ZCA whitening computed from a single support cloud's per-class covariance makes support features align with the learnable prototypical tokens after the learned query/key projections, and that this covariance is a stable, invertible stand-in for the true distribution.

What would settle it

If one measured the distance between whitened support features and prototypical tokens after the learned query/key projections in a well-trained model and found it still as large as without whitening (Table 2 scale), the alignment story would be falsified. Alternatively, if replacing ZCA whitening with diagonal per-channel scaling (preserving correlations) matched WARM's accuracy, then decorrelation is not the key ingredient.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, prototype construction in few-shot 3D point cloud segmentation no longer depends on random seeds, making results reproducible across initializations.
  • Attention-based prototypes are end-to-end trainable, so the prototype generator can be optimized jointly with the segmentation head rather than being a fixed preprocessing step.
  • WARM's whitening-alignment pattern could generalize to other few-shot tasks where learnable queries and support features are distributionally mismatched.
  • The paper's demonstration that a simple distance-based classifier with strong prototypes beats complex decoders suggests future FS-PCS work should focus on prototype quality rather than decoder complexity.

Where Pith is reading between the lines

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

  • Because the whitening step uses a covariance estimated from a single support cloud, its benefit may partly come from the regularizing effect of whitening rather than from true alignment; an experiment that decorrelates these two effects would clarify the mechanism.
  • WARM's deterministic prototypes could serve as a drop-in replacement for FPS in other 3D few-shot pipelines, such as object detection or part segmentation, where prototype stability matters.
  • The coloring step restores support-set statistics, but if the backbone features shift between support and query (domain gap), the restored statistics may be less useful; a testable extension is to color with a weighted blend of support and query statistics.
  • The reported entropy and diversity changes suggest WARM flattens attention maps; these metrics could be reused as a training-time regularizer or as an early-stopping signal for attention-based prototype methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper addresses prototype generation in few-shot 3D point cloud semantic segmentation. It observes that FPS-based prototype construction is seed-sensitive and proposes WARM, which replaces FPS with learnable prototypical tokens aggregated through cross-attention. WARM whitens per-class support features (ZCA) before attention and applies the inverse coloring after attention to restore original statistics. Experiments on S3DIS and ScanNet compare with AttMPTI, COSeg, and an FPS+min-dist baseline, reporting mIoU across two class folds. The authors claim state-of-the-art on S3DIS and competitive performance on ScanNet, with ablations and attention analyses supporting the alignment narrative.

Significance. If validated, WARM is a simple module that makes prototype construction deterministic and trainable, addressing a real instability of FPS in FS-PCS. The paper ships public code, clearly motivates the problem, and includes component ablations and attention visualizations. The observation that a simple distance-based baseline can outperform a sophisticated decoder is also valuable. However, the empirical validation and the alignment evidence are currently too thin for the stated conclusions: only two prior methods are compared, no training variance is reported, Dist(Q,K) is undefined and scale-sensitive, and the whitening transform is not conditioned for low-support episodes.

major comments (5)
  1. [§6.2, Tables 3–4] The 'state-of-the-art' claim rests on only two existing methods (AttMPTI and COSeg) plus a hand-built FPS baseline, while §2.1 cites several recent FS-PCS methods (Refs. [3], [20], [24], [38]) that are not evaluated. On ScanNet WARM is not the best method in the 1-way 1-shot row (COSeg 41.78 vs 39.76), so the abstract's 'state-of-the-art on multiple benchmarks' is not supported. Please add the missing recent baselines under the same protocol, or restrict the claim to 'outperforms AttMPTI and COSeg.'
  2. [§6.3.1, Table 5; §5.2, Eq. (10)] Dist(Q,K) is never formally defined. If it is a Euclidean distance, the reported drop from 288.81 to 13.13 is expected from scale alone: ZCA whitening in Eq. (10) forces the sample covariance of the key features toward a scaled identity, compressing key-vector norms. This does not establish that queries and keys are angularly or semantically aligned after the learned projections W_q/W_k. Please define the metric, report a scale-invariant version (e.g., cosine distance or normalized Euclidean distance), and include a control where whitening uses a random covariance matrix to separate scale effects from alignment.
  3. [§5.2, Eqs. (9)–(10)] ZCA whitening requires Σ_c to be invertible. In the low-shot episodes the paper targets, L_c (number of support points for a class) can be below the feature dimension D, making Σ_c singular. No regularization, pseudo-inverse, or rank analysis is reported. Please report the range of L_c across classes and episodes relative to D, use a damped inverse (Σ_c + εI)^{-1/2} with a sensitivity study, and confirm that results in the singular regime are not numerical artifacts.
  4. [§6.2, Tables 3–4; §6.1.2] The paper's motivation is seed-induced instability (Table 1), but the main results report no variance or repeated training runs. The S-0/S-1 columns are class folds, not random seeds. Please report mean ± std over at least three training runs for the main comparisons, or explicitly state that a single run was used and assess sensitivity. Without this, the 'deterministic and stable prototype generation' claim is only about inference-time FPS, not about the whole pipeline.
  5. [§6.3.2, Table 6 and Fig. 3] The semantic-aggregation claim is supported by attention entropy and diversity. Higher entropy alone can indicate over-smoothed or uniform attention rather than semantic grouping, and lower diversity can indicate collapse. Please add a direct semantic metric—e.g., the proportion of attention mass on support points of the same ground-truth class as the token's assigned class, or correlation with a class-prototype similarity map—and report the distribution of entropy values, not only the mean.
minor comments (6)
  1. [§5.2, Eq. (10)] With Σ_c defined in Eq. (9) with denominator L_c−1, Z_c^T Z_c = (L_c−1)I, not I. Either change the covariance normalization or include a 1/sqrt(L_c−1) factor, and ensure the implementation matches the stated identity.
  2. [Abstract / Conclusion] The abstract and conclusion claim 'state-of-the-art performance with a significant margin on multiple benchmarks,' but ScanNet results are competitive rather than state-of-the-art. Please temper the claim.
  3. [§5.2, Eq. (12)] Coloring is applied to P0 + CA(P0, Z_c), so the support statistics are also injected into the learnable token P0, not only into the attention output. Please clarify whether this is intentional and discuss its effect on the learned token distribution.
  4. [§4.2, Eqs. (6)–(7)] The definitions of D_intra and D_inter use FG=FG' and FG≠FG' but do not explain how the two episodes are sampled. Please clarify the protocol for computing these metrics.
  5. [§6.3.1, Table 5] For coloring variants (e)–(g), the table reports S-0 but not Dist(Q,K). Reporting Dist(Q,K) for the final prototypes would help separate the alignment contribution from the restoration contribution.
  6. [Fig. 5] The number-of-tokens ablation has no error bars or repetition information. Please state how many runs were averaged.

Circularity Check

0 steps flagged

No significant circularity: WARM's effectiveness is an empirical architecture result, not a derivation forced by its inputs.

full rationale

The paper's central claim—that wrapping cross-attention with whitening/coloring (Eqs. 10–12) yields better prototypes and improved mIoU—is supported by direct benchmark comparisons (Tables 3–4) and ablations (Tables 5–6). The method is trained end-to-end on base classes and evaluated on held-out novel classes, so the reported performance is not a fitted input renamed as a prediction. Whitening and coloring are explicit transformations of support features with no dependence on the target metric; even if the undefined Dist(Q,K) metric in Table 5 is scale-sensitive and the covariance invertibility assumption is unverified, those are evidence-quality concerns, not cases where a conclusion is identical to an input by construction. The self-citations (Refs. [18] and [19]) occur in motivation and related-work contexts and are not load-bearing for the central result. No circular step is identifiable from the paper's own equations or citations.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

Central method depends on the learned parameters listed, on the frozen pretrained feature space, and on the unproven alignment claim. No formal derivation is offered; this is an empirical architecture paper. The assumptions above are the load-bearing premises not established by the paper's own equations.

free parameters (4)
  • Learnable prototypical tokens P_0 = 100 tokens per class, trained on base classes
    Learnable query tokens optimized during training; their initialization and optimization drive prototype quality, which is central to WARM.
  • Cross-attention projection layers W_q, W_k, W_v
    Learned projections in Eq. 2-3; the claimed alignment between queries and keys depends on these optimized parameters.
  • Simplification loss weight lambda = 0.5
    Hand-chosen coefficient in Eq. 17; no sensitivity analysis reported, but it shapes prototype behavior and final performance.
  • Number of prototypical tokens M = 100 per class
    Hand-chosen architecture hyperparameter; Fig. 5 shows robustness, but it remains a free choice that affects results.
axioms (5)
  • domain assumption Support and query features from the frozen Stratified Transformer share a metric space where L2 distance to prototypes is a valid classifier.
    Eq. 13-14 assign labels by nearest prototype distance in the frozen feature space; no proof that the backbone features are metric-appropriate.
  • domain assumption The per-class covariance matrix Sigma_c is full rank and its inverse square root exists for every class and episode.
    Eq. 9-10 require inverting Sigma_c; with limited support points or degenerate channels, ZCA whitening is undefined or unstable.
  • ad hoc to paper Whitening support features aligns them with learnable prototypical tokens in the attention projection space.
    Sec. 4.2 and 5.2 assert alignment; the projections W_q/W_k are learned and no mathematical guarantee is provided.
  • ad hoc to paper The distributional gap metrics and attention entropy analyses are causal evidence for the proposed fix.
    Tables 2 and 6 are observational diagnostics from the authors' trained models; they are interpreted as causes of attention failure.
  • domain assumption The S3DIS/ScanNet evaluation protocol and baselines used are sufficient to establish state-of-the-art.
    Claim of SOTA depends on fair comparison; recent methods cited in Sec. 2.1 are omitted, and no error bars are given.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of White Aggregation and Restoration for Few-shot 3D Point Cloud Semantic Segmentation." pith.science (2026). https://pith.science/paper/FLVKZINX

@misc{pith2026250913907,
  author       = {Pith},
  title        = {Pith review of: White Aggregation and Restoration for Few-shot 3D Point Cloud Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLVKZINX}},
  note         = {Machine review of arXiv:2509.13907}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Few-shot 3D Point Cloud Semantic Segmentation (FS-PCS) aims to predict per-point labels for an unlabeled point cloud, given only a few labeled examples. To extract representations from the limited labeled set, existing methods have constructed prototypes with Farthest Point Sampling (FPS). However, we found that this convention results in performance instability due to its sensitivity to FPS-induced variations, while the prototype generation process remains underexplored in the field. This motivates us to investigate deterministic prototype generation method based on attention mechanism. Despite its potential, we found that vanilla attention module suffers from the distributional gap between prototypical tokens and support features. To overcome this, we provide a simple approach, White Aggregation and Restoration Module (WARM), which resolves the misalignment by wrapping cross-attention with whitening and coloring transformations. Specifically, whitening aligns the features to tokens before the attention process, and coloring subsequently restores the original distribution to the attended tokens. This design enables robust attention, thereby generating prototypes that capture the semantic relationships in support features. WARM achieves state-of-the-art performance with a significant margin on the S3DIS dataset, and competitive performance on the ScanNet dataset. Further experiments demonstrate its effectiveness in deterministic prototype generation. Code is publicly available at: https://github.com/JiyunIm00/WARM.git

Figures

Figures reproduced from arXiv: 2509.13907 by Jae-Pil Heo, Jiyun Im, Miso Lee, SuBeen Lee.

Figure 1
Figure 1. Figure 1: (a) Alignment between prototypical tokens and point cloud features dipicted with t-SNE. In na¨ıve cross-attention, queries (prototypical tokens) are often misaligned with keys (point cloud features), resulting in inaccurate matching that fails to cap￾ture relational structures between keys. In contrast, our method facilitates alignment between queries and keys, leading to more stable matching that reflects… view at source ↗
Figure 2
Figure 2. Figure 2: Overall pipeline in the 1-way 1-shot scenario. Initially, support and query features are extracted using a backbone network. Then, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of attention maps. For a fair com [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training stability and acceleration by whitening. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study of the number of prototypical tokens. The [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    Correlations are ruining your gradient de- scent.arXiv preprint arXiv:2407.10780, 2024

    Nasir Ahmad. Correlations are ruining your gradient de- scent.arXiv preprint arXiv:2407.10780, 2024. 7

    Pith/arXiv arXiv 2024

  2. [2]

    Re- thinking few-shot 3d point cloud semantic segmentation

    Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, and Serge Belongie. Re- thinking few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 3996–4006, 2024. 1, 2, 3, 6

    2024

  3. [3]

    Multimodality helps few-shot 3d point cloud semantic seg- mentation

    Zhaochong An, Guolei Sun, Yun Liu, Runjia Li, Min Wu, Ming-Ming Cheng, Ender Konukoglu, and Serge Belongie. Multimodality helps few-shot 3d point cloud semantic seg- mentation. InICLR, 2025. 2, 3

    2025

  4. [4]

    3d seman- tic parsing of large-scale indoor spaces

    Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese. 3d seman- tic parsing of large-scale indoor spaces. InProceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1534–1543, 2016. 1, 2, 6

    2016

  5. [5]

    independent components

    Anthony J Bell and Terrence J Sejnowski. The “independent components” of natural scenes are edge filters.Vision re- search, 37(23):3327–3338, 1997. 2, 5

    1997

  6. [6]

    Deep learning on 3d semantic segmentation: A detailed review.Remote Sensing, 17(2):298, 2025

    Thodoris Betsas, Andreas Georgopoulos, Anastasios Doulamis, and Pierre Grussenmeyer. Deep learning on 3d semantic segmentation: A detailed review.Remote Sensing, 17(2):298, 2025. 1

    2025

  7. [7]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. InEuropean confer- ence on computer vision, pages 213–229. Springer, 2020. 3

    2020

  8. [8]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3

    2022

  9. [9]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 1, 2, 6

    2017

  10. [10]

    Batch normalization prov- ably avoids ranks collapse for randomly initialised deep net- works.Advances in Neural Information Processing Systems, 33:18387–18398, 2020

    Hadi Daneshmand, Jonas Kohler, Francis Bach, Thomas Hofmann, and Aurelien Lucchi. Batch normalization prov- ably avoids ranks collapse for randomly initialised deep net- works.Advances in Neural Information Processing Systems, 33:18387–18398, 2020. 2

    2020

  11. [11]

    Attention is not all you need: Pure attention loses rank dou- bly exponentially with depth

    Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank dou- bly exponentially with depth. InInternational conference on machine learning, pages 2793–2803. PMLR, 2021. 2, 3

    2021

  12. [12]

    Self- support few-shot semantic segmentation

    Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self- support few-shot semantic segmentation. InEuropean con- ference on computer vision, pages 701–719. Springer, 2022. 2, 6

    2022

  13. [13]

    Prototype adaption and projection for few- and zero-shot 3d point cloud semantic segmentation.IEEE Transactions on Image Processing, 2023

    Shuting He, Xudong Jiang, Wei Jiang, and Henghui Ding. Prototype adaption and projection for few- and zero-shot 3d point cloud semantic segmentation.IEEE Transactions on Image Processing, 2023. 2

    2023

  14. [14]

    Decorre- lated batch normalization

    Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorre- lated batch normalization. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 791–800, 2018. 2, 5

    2018

  15. [15]

    Revealing the dark secrets of bert.arXiv preprint arXiv:1908.08593, 2019

    Olga Kovaleva, Alexey Romanov, Anna Rogers, and Anna Rumshisky. Revealing the dark secrets of bert.arXiv preprint arXiv:1908.08593, 2019. 2 8

    Pith/arXiv arXiv 1908

  16. [16]

    Stratified trans- former for 3d point cloud segmentation

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified trans- former for 3d point cloud segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8500–8509, 2022. 6

    2022

  17. [17]

    Samplenet: Differentiable point cloud sampling

    Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7578–7588, 2020. 5

    2020

  18. [18]

    Activating self- attention for multi-scene absolute pose regression.Advances in Neural Information Processing Systems, 37:38508–38529,

    Miso Lee, Jihwan Kim, and Jae-Pil Heo. Activating self- attention for multi-scene absolute pose regression.Advances in Neural Information Processing Systems, 37:38508–38529,

  19. [19]

    Temporal alignment-free video matching for few- shot action recognition

    SuBeen Lee, WonJun Moon, Hyun Seok Seong, and Jae- Pil Heo. Temporal alignment-free video matching for few- shot action recognition. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 5412–5421,

  20. [20]

    Localization and expansion: A decoupled frame- work for point cloud few-shot semantic segmentation

    Zhaoyang Li, Yuan Wang, Wangkai Li, Rui Sun, and Tianzhu Zhang. Localization and expansion: A decoupled frame- work for point cloud few-shot semantic segmentation. In European Conference on Computer Vision, pages 18–34. Springer, 2024. 2, 3

    2024

  21. [21]

    Masked dis- crimination for self-supervised learning on point clouds

    Haotian Liu, Mu Cai, and Yong Jae Lee. Masked dis- crimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657–675. Springer, 2022. 3

    2022

  22. [22]

    Part-aware prototype network for few-shot semantic segmentation

    Yongfei Liu, Xiangyi Zhang, Songyang Zhang, and Xum- ing He. Part-aware prototype network for few-shot semantic segmentation. InEuropean conference on computer vision, pages 142–158. Springer, 2020. 2, 3

    2020

  23. [23]

    An end-to- end transformer model for 3d object detection

    Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to- end transformer model for 3d object detection. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 2906–2917, 2021. 3

    2021

  24. [24]

    Boosting few-shot 3d point cloud segmentation via query-guided enhancement

    Zhenhua Ning, Zhuotao Tian, Guangming Lu, and Wenjie Pei. Boosting few-shot 3d point cloud segmentation via query-guided enhancement. InProceedings of the 31st ACM international conference on multimedia, pages 1895–1904,

    1904

  25. [25]

    How does batch normalization help optimiza- tion?Advances in neural information processing systems, 31, 2018

    Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Alek- sander Madry. How does batch normalization help optimiza- tion?Advances in neural information processing systems, 31, 2018. 4

    2018

  26. [26]

    Mask3d: Mask trans- former for 3d semantic instance segmentation.arXiv preprint arXiv:2210.03105, 2022

    Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask trans- former for 3d semantic instance segmentation.arXiv preprint arXiv:2210.03105, 2022. 3

    Pith/arXiv arXiv 2022

  27. [27]

    Prototypical networks for few-shot learning.Advances in neural informa- tion processing systems, 30, 2017

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.Advances in neural informa- tion processing systems, 30, 2017. 2, 3, 6

    2017

  28. [28]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

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

    2017

  29. [29]

    Matching networks for one shot learning.Ad- vances in neural information processing systems, 29, 2016

    Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning.Ad- vances in neural information processing systems, 29, 2016. 3

    2016

  30. [30]

    Detr3d: 3d ob- ject detection from multi-view images via 3d-to-2d queries

    Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d ob- ject detection from multi-view images via 3d-to-2d queries. InConference on robot learning, pages 180–191. PMLR,

  31. [31]

    Unsupervised point cloud rep- resentation learning with deep neural networks: A survey

    Aoran Xiao, Jiaxing Huang, Dayan Guan, Xiaoqin Zhang, Shijian Lu, and Ling Shao. Unsupervised point cloud rep- resentation learning with deep neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 45(9):11321–11339, 2023. 1

    2023

  32. [32]

    Pixel-aligned recurrent queries for multi-view 3d object detection

    Yiming Xie, Huaizu Jiang, Georgia Gkioxari, and Julian Straub. Pixel-aligned recurrent queries for multi-view 3d object detection. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 18370–18380,

  33. [33]

    Stabilizing transformer training by pre- venting attention entropy collapse

    Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by pre- venting attention entropy collapse. InInternational Con- ference on Machine Learning, pages 40770–40803. PMLR,

  34. [34]

    Feature- proxy transformer for few-shot segmentation.Advances in neural information processing systems, 35:6575–6588,

    Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Feature- proxy transformer for few-shot segmentation.Advances in neural information processing systems, 35:6575–6588,

  35. [35]

    Threshold- consistent margin loss for open-world deep metric learning

    Qin Zhang, Linghan Xu, Qingming Tang, Jun Fang, Ying Nian Wu, Joe Tighe, and Yifan Xing. Threshold- consistent margin loss for open-world deep metric learning. arXiv preprint arXiv:2307.04047, 2023. 5

    Pith/arXiv arXiv 2023

  36. [36]

    Few-shot 3d point cloud semantic segmentation

    Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8873–8882, 2021. 1, 2, 3, 6

    2021

  37. [37]

    Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis

    Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707– 14717, 2024. 3

    2024

  38. [38]

    No time to train: Empowering non-parametric net- works for few-shot 3d scene segmentation

    Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ji- aming Liu, Han Xiao, Chaoyou Fu, Hao Dong, and Peng Gao. No time to train: Empowering non-parametric net- works for few-shot 3d scene segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3838–3847, 2024. 1, 2 9

    2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.