Pith. sign in

REVIEW 2 minor 53 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

DHANet aggregates features at multiple scales spatially and along channels to reduce semantic and attribute over-alignment in cross-domain few-shot segmentation.

2026-06-26 01:01 UTC pith:XAYV42H7

load-bearing objection The paper introduces DHANet with HSA, HCA, and OPSB modules to target semantic and attribute over-alignment in CD-FSS beyond style shifts, but the abstract supplies no numbers or comparisons to support the SOTA claim.

arxiv 2606.24296 v1 pith:XAYV42H7 submitted 2026-06-23 cs.CV

Hierarchical Spatial and Channel Aggregation for Cross-domain Few-shot Segmentation

Sujun Sun , Mingwu Ren , Haofeng Zhang This is my paper
classification cs.CV
keywords cross-domain few-shot segmentationhierarchical spatial aggregationhierarchical channel aggregationsemantic over-alignmentattribute over-alignmentdual hierarchical aggregation networkonline probabilistic semantic bank
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 claims that existing methods for cross-domain few-shot segmentation overlook differences in class semantic granularity and discriminative attributes between domains, which produce semantic over-alignment and attribute over-alignment during support-query matching. It introduces the Dual Hierarchical Aggregation Network to generate multi-scale semantic-enhanced features through spatial aggregation and multi-scale attribute-enhanced features through channel aggregation. An online probabilistic semantic bank is added to build and sample from class probability distributions during inference, supplying extra pseudo-prototypes when support is limited. The method is evaluated on four target-domain datasets where it reports state-of-the-art results. A sympathetic reader would care because the approach targets the matching process directly rather than only style shifts.

Core claim

The Dual Hierarchical Aggregation Network comprises Hierarchical Spatial Aggregation that performs multi-scale region aggregation of pixel features to produce hierarchical semantic-enhanced features, Hierarchical Channel Aggregation that performs multi-scale attribute aggregation along the channel dimension to produce hierarchical attribute-enhanced features, and the Online Probabilistic Semantic Bank that progressively constructs class probability distributions from query predictions and samples multiple pseudo-prototypes as additional support.

What carries the argument

Dual Hierarchical Aggregation Network (DHANet) consisting of HSA for spatial multi-scale region aggregation, HCA for channel multi-scale attribute aggregation, and OPSB for online construction of class probability distributions and pseudo-prototype sampling.

Load-bearing premise

The main degradations in support-query matching arise from semantic over-alignment and attribute over-alignment caused by cross-domain differences in class semantic granularity and discriminative attributes.

What would settle it

An ablation study on a target dataset that measures the drop in segmentation accuracy when the hierarchical spatial and channel aggregation modules are removed while keeping all other components fixed.

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

If this is right

  • HSA produces hierarchical semantic-enhanced features that reduce semantic over-alignment.
  • HCA produces hierarchical attribute-enhanced features that reduce attribute over-alignment.
  • OPSB supplies sampled pseudo-prototypes that compensate for insufficient support information.
  • The combined modules yield state-of-the-art segmentation performance on four target-domain datasets.

Where Pith is reading between the lines

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

  • The same hierarchical aggregation pattern could be tested on other few-shot tasks that rely on support-query matching across domains.
  • If semantic granularity differences are the dominant factor, similar modules might improve performance even when style gaps are small.
  • The online bank mechanism suggests a general way to augment limited support sets using query-side predictions in inference-time adaptation.

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

0 major / 2 minor

Summary. The paper proposes the Dual Hierarchical Aggregation Network (DHANet) for Cross-domain Few-shot Segmentation (CD-FSS). It argues that existing methods overlook differences in class semantic granularity and discriminative attributes across domains, causing semantic over-alignment and attribute over-alignment in support-query matching. DHANet consists of a Hierarchical Spatial Aggregation (HSA) module for multi-scale spatial region aggregation to produce semantic-enhanced features, a Hierarchical Channel Aggregation (HCA) module for multi-scale channel-wise attribute aggregation, and an Online Probabilistic Semantic Bank (OPSB) that builds and samples from class probability distributions during inference to generate pseudo-prototypes. The central claim is that these components yield state-of-the-art results on four target-domain datasets.

Significance. If the experimental claims hold, the work offers a targeted architectural response to granularity and attribute mismatches in CD-FSS via explicit hierarchical aggregation along spatial and channel axes plus inference-time prototype augmentation. This could strengthen generalization in few-shot settings where support is limited and domain shifts involve semantic structure rather than only style.

minor comments (2)
  1. [Abstract] Abstract: the state-of-the-art claim is stated without any numerical results, dataset names, or baseline comparisons, which reduces immediate readability even though the full experimental section presumably supplies them.
  2. [Introduction] The motivation paragraph introduces 'semantic over-alignment' and 'attribute over-alignment' as key degradations; a brief illustrative figure or toy example early in the paper would clarify these terms for readers unfamiliar with the precise failure modes.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary of our DHANet approach to cross-domain few-shot segmentation and for recommending minor revision. No specific major comments were provided in the report, so we have no individual points to address point-by-point. We are pleased that the significance of the hierarchical spatial/channel aggregation and online probabilistic bank is recognized.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is an empirical architectural proposal for DHANet in cross-domain few-shot segmentation. It identifies two alignment issues from domain differences, maps them to three modules (HSA for spatial multi-scale aggregation, HCA for channel attribute aggregation, OPSB for pseudo-prototypes), and reports SOTA results on four target datasets. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim rests on external experimental outcomes rather than any internal reduction of outputs to inputs by construction. This is the standard non-circular case for method papers.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The abstract does not enumerate fitted parameters or new physical entities; contributions are network modules resting on standard assumptions of deep feature learning.

axioms (1)
  • domain assumption Existing methods overlook differences in class semantic granularity and discriminative attributes, producing semantic and attribute over-alignment.
    This premise is stated explicitly as the motivation for the new modules.

pith-pipeline@v0.9.1-grok · 5744 in / 1196 out tokens · 68502 ms · 2026-06-26T01:01:30.576404+00:00 · methodology

0 comments
read the original abstract

Cross-domain Few-shot Segmentation (CD-FSS) aims to learn generalizable segmentation capability from abundant annotated samples in the source domain, enabling accurate segmentation of novel classes in the target domain with only a few annotated samples. Existing CD-FSS methods mainly focus on mitigating feature distribution shifts caused by style gaps while ignoring significant differences in class semantic granularity and discriminative attributes across domains, leading to two key degradations in support-query matching: semantic over-alignment and attribute over-alignment. To this end, we propose the Dual Hierarchical Aggregation Network (DHANet), which comprises three key modules. First, the Hierarchical Spatial Aggregation (HSA) module performs multi-scale region aggregation of pixel features along the spatial dimension, generating hierarchical semantic-enhanced features to alleviate semantic over-alignment. Additionally, the HCA module conducts multi-scale attribute aggregation along the channel dimension, generating hierarchical attribute-enhanced features to mitigate attribute over-alignment. Finally, we propose the Online Probabilistic Semantic Bank (OPSB), which progressively constructs and updates class probability distributions from query predictions during inference, and samples multiple pseudo-prototypes as additional support information to mitigate insufficient support. Extensive experiments on four target-domain datasets demonstrate that our method achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2606.24296 by Haofeng Zhang, Mingwu Ren, Sujun Sun.

Figure 1
Figure 1. Figure 1: Motivation of DHANet. (a) Semantic over-alignment prevents the model from distinguishing semantic categories at other segmentation granularities (e.g., a cat’s head and body). Hierarchical spatial aggregation aligns features at multiple granu￾larities, adapting to target domains with varying granularities. (b) Attribute over￾alignment causes the degradation of source-insensitive attributes. Hierarchical ch… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our method. The support and query images are first fed into GSR for style perturbation, and features are extracted by the image encoder. Subsequently, the HSA and HCA modules sequentially perform hierarchical semantic and attribute enhancement on the features. The enhanced features are then fed into the main branch and the auxiliary branch to compute the query foreground confidence maps, respec… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Qualitative results of our model for 1-way 1-shot setting. (b) The heatmaps of the foreground similarity maps show that our method can extract hierarchical se￾mantic features to alleviate semantic over-alignment [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of spatial and channel slots demonstrates that multi-stage slots facilitate constructing features with hierarchical semantic and attribute enhancement [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: Parameter studies. (a) The mIoU curves over different loss weight λ. (b) The mIoU curves over different sampling numbers P in OPSB. sible reason is that excessive low-level features weaken the semantic information of high-level features [PITH_FULL_IMAGE:figures/full_fig_p022_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: More qualitative results of our method for 1-way 1-shot segmentation. The support labels are highlighted in blue, while the predictions and ground truth of query images are presented in red. Source Domain Target Domain Image Mask Baseline Ours [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The heatmaps of the foreground similarity maps for source and target domains demonstrate that our method can extract hierarchical semantic features to alleviate semantic over-alignment. 4 More Visualizations Qualitative results. In [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The channel-wise correlation maps show that our method mitigates attribute over-alignment and reduces channel redundancy. Brighter colors represent higher cor￾relation values. tic discriminative ability, we visualize more corresponding heatmaps in [PITH_FULL_IMAGE:figures/full_fig_p024_4.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

53 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    Advances in neural information processing systems , volume=

    Segnext: Rethinking convolutional attention design for semantic segmentation , author=. Advances in neural information processing systems , volume=

  2. [2]

    arXiv preprint arXiv:2404.16573 , year=

    Multi-scale representations by varying window attention for semantic segmentation , author=. arXiv preprint arXiv:2404.16573 , year=

    work page arXiv

  3. [3]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    SegMAN: Omni-scale context modeling with state space models and local attention for semantic segmentation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  4. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2022 , publisher=

    2022

  5. [5]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Scene parsing through ade20k dataset , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  6. [6]

    European conference on computer vision , pages=

    Self-support few-shot semantic segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    VRP-SAM: SAM with visual reference prompt , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [8]

    arXiv preprint arXiv:2505.14100 , year=

    Unlocking the power of sam 2 for few-shot segmentation , author=. arXiv preprint arXiv:2505.14100 , year=

    work page arXiv

  9. [9]

    European conference on computer vision , pages=

    Cross-domain few-shot semantic segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Domain-rectifying adapter for cross-domain few-shot segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [11]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    The devil is in low-level features for cross-domain few-shot segmentation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  12. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Apseg: auto-prompt network for cross-domain few-shot semantic segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [13]

    arXiv preprint arXiv:2506.07376 , year=

    Adapter naturally serves as decoupler for cross-domain few-shot semantic segmentation , author=. arXiv preprint arXiv:2506.07376 , year=

    work page arXiv

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    SAM-Aware Graph Prompt Reasoning Network for Cross-Domain Few-Shot Segmentation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Cross-domain few-shot segmentation via iterative support-query correspondence mining , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  16. [16]

    arXiv preprint arXiv:2506.02677 , year=

    Self-disentanglement and re-composition for cross-domain few-shot segmentation , author=. arXiv preprint arXiv:2506.02677 , year=

    work page arXiv

  17. [17]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Bridging Granularity Gaps: Hierarchical Semantic Learning for Cross-domain Few-shot Segmentation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  18. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Learning invariant inter-pixel correlations for superpixel generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  19. [19]

    proceedings of the IEEE/CVF international conference on computer vision , pages=

    Panet: Few-shot image semantic segmentation with prototype alignment , author=. proceedings of the IEEE/CVF international conference on computer vision , pages=

  20. [20]

    Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VIII 16 , pages=

    Prototype mixture models for few-shot semantic segmentation , author=. Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VIII 16 , pages=. 2020 , organization=

    2020

  21. [21]

    Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part IX 16 , pages=

    Part-aware prototype network for few-shot semantic segmentation , author=. Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part IX 16 , pages=. 2020 , organization=

    2020

  22. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Hierarchical dense correlation distillation for few-shot segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  23. [23]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Self-calibrated cross attention network for few-shot segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  24. [24]

    European Conference on Computer Vision , pages=

    Eliminating feature ambiguity for few-shot segmentation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  25. [25]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Rethinking prior information generation with clip for few-shot segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  26. [26]

    IEEE Transactions on Circuits and Systems for Video Technology , year=

    Multi-modal Few-Shot Semantic Segmentation Based on Triple Attention Mechanism and Hierarchical Decoding Transformer , author=. IEEE Transactions on Circuits and Systems for Video Technology , year=

  27. [27]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Prompt-and-transfer: Dynamic class-aware enhancement for few-shot segmentation , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2024 , publisher=

    2024

  28. [28]

    Advances in Neural Information Processing Systems , volume=

    Bridge the points: Graph-based few-shot segment anything semantically , author=. Advances in Neural Information Processing Systems , volume=

  29. [29]

    DINOv2: Learning Robust Visual Features without Supervision

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  31. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Adapt before comparison: A new perspective on cross-domain few-shot segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Lightweight frequency masker for cross-domain few-shot semantic segmentation , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    Advances in neural information processing systems , volume=

    Object-centric learning with slot attention , author=. Advances in neural information processing systems , volume=

  34. [34]

    Advances in neural information processing systems , volume=

    Attrseg: open-vocabulary semantic segmentation via attribute decomposition-aggregation , author=. Advances in neural information processing systems , volume=

  35. [35]

    arXiv preprint arXiv:2405.15265 , year=

    Cross-domain few-shot semantic segmentation via doubly matching transformation , author=. arXiv preprint arXiv:2405.15265 , year=

    work page arXiv

  36. [36]

    International journal of computer vision , volume=

    The pascal visual object classes (voc) challenge , author=. International journal of computer vision , volume=. 2010 , publisher=

    2010

  37. [37]

    2011 international conference on computer vision , pages=

    Semantic contours from inverse detectors , author=. 2011 international conference on computer vision , pages=. 2011 , organization=

    2011

  38. [38]

    Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    Deepglobe 2018: A challenge to parse the earth through satellite images , author=. Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pages=

    2018

  39. [39]

    Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic) , author=. arXiv preprint arXiv:1902.03368 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  40. [40]

    Scientific data , volume=

    The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions , author=. Scientific data , volume=. 2018 , publisher=

    2018

  41. [41]

    IEEE transactions on medical imaging , volume=

    Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration , author=. IEEE transactions on medical imaging , volume=. 2013 , publisher=

    2013

  42. [42]

    IEEE transactions on medical imaging , volume=

    Automatic tuberculosis screening using chest radiographs , author=. IEEE transactions on medical imaging , volume=. 2013 , publisher=

    2013

  43. [43]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Fss-1000: A 1000-class dataset for few-shot segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  44. [44]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  45. [45]

    Advances in neural information processing systems , volume=

    Feature-proxy transformer for few-shot segmentation , author=. Advances in neural information processing systems , volume=

  46. [46]

    Personalize Segment Anything Model with One Shot

    Personalize segment anything model with one shot , author=. arXiv preprint arXiv:2305.03048 , year=

    work page Pith review arXiv

  47. [47]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Hypercorrelation squeeze for few-shot segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  48. [48]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Few-shot segmentation without meta-learning: A good transductive inference is all you need? , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  49. [49]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Adapting In-Domain Few-Shot Segmentation to New Domains without Source Domain Retraining , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  50. [50]

    COP: Customized Deep Model Compression via Regularized Correlation-Based Filter-Level Pruning

    Cop: Customized deep model compression via regularized correlation-based filter-level pruning , author=. arXiv preprint arXiv:1906.10337 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  51. [51]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    FORLA: Federated Object-Centric Representation Learning with Slot Attention , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  52. [52]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Dual-agent optimization framework for cross-domain few-shot segmentation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  53. [53]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=