acceptodds
Under review as a conference paper at ICLR 2027

CGPRNet: Cross-Scale Graph Prototype Refinement for Robust Few-Shot Semantic Segmentation

Abstract

Few-shot semantic segmentation (FSS) learns to segment novel classes with only a few annotated support images. Despite the promising results of recent methods, there remain three critical limitations: (1) current FSS methods do not explicitly model the cross-scale semantic dependencies between support and query; (2) they generate prototypes by simple embedding statistics while disregarding the progressive refinement ability of graph; (3) they fuse the multi-scale information in a principled manner. To address these challenges, We propose CGPR, a framework that jointly integrates query-conditioned graph prototype refinement, adaptive multi-scale matching, and dual-branch feature encoding to improve prototype quality, feature representation, and robustness to cross-scale variation. Concretely, we design a Cross-Attention Graph Prototype Generator to iteratively refine the prototypes by aggregating the cross-scale support-query relationships, and a Multi-Scale Prototype Matching module to compute the similarity at different receptive fields in an adaptive way. Additionally, we introduce a dual-branch feature encoder with frozen generic features and trainable task-specific features to prevent overfitting while maintaining discriminative power. We conduct extensive experiments on PASCAL-5i, COCO-20i, and FSS-1000 datasets to show the superiority of CGPRNet over the previous state-of-the-art methods. Under both 1-shot and 5-shot settings, our CGPRNet obtains better mIoU, FB-IoU and Dice scores on these three benchmarks, demonstrating the effectiveness of cross-scale and query-conditioned prototype refinement learning for robust few-shot segmentation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-Shot Semantic Segmentation, Prototype Refinement, Cross-Scale Feature Learning, Graph Neural Networks, Computer Vision

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dgY7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CGPRNet for the traditional Few-Shot Semantic Segmentation task. The network mainly consists of three modules: the Dual-Branch Feature Encoder uses a frozen pretrained CNN backbone to extract generic features, while learning task-specific features through a trainable adapter and then fusing them using a gating mechanism; the Cross-Attention Graph Prototype Generator initializes foreground and background prototypes based on support masks and iteratively updates the prototypes through query attention and message passing between graph nodes; the Multi-Scale Prototype Matching module applies 3×3, 5×5, and 7×7 convolutions to query features and adaptively fuses matching results from different receptive fields. The authors conduct 1-shot and 5-shot experiments on PASCAL-5i, COCO-20i, and FSS-1000, and report some competitive results.

Critical strengths

The author’s methodological framework introduces some novel ideas, and the experimental results show a certain level of competitiveness.

Critical weaknesses
  1. There are several issues in the experimental results and configurations that require further clarification. For example, the mIoU, Dice, and FB results under the ResNet-50 configuration on FSS-1000 are inconsistent across Tables 3, 4, and 5. Additionally, the standard evaluation protocol for FSS-1000 divides the 1,000 categories into 520 training classes, 240 validation classes, and 240 testing classes, with evaluation performed on the test set. However, the authors report their results based on four folds. It is unclear how F0/F1/F2/F3 are constructed and whether the evaluation follows the standard protocol.

  2. The authors' claim of State-of-the-art Performance does not seem sufficiently convincing. First, among the reported experimental results, the proposed method fails to achieve the best performance in several settings on PASCAL-5i. On COCO-20i, the mIoU under the R101 (5-shot) setting is nearly 5 points lower than HSNet. However, the authors still highlight their own results in bold, and there are also several incorrect bold/underlined markings in the tables. Second, most of the baseline methods listed in the tables were published between 2020 and 2022, with only a few more recent methods included. However, several stronger few-shot segmentation methods proposed in recent years, such as HDMNet (CVPR 2023), AMFormer (NeurIPS 2023), and HMNet (NeurIPS 2024), are not included in the comparisons. Among them, AMFormer uses ResNet-101 as the backbone and achieves 70.7% and 73.6% mIoU on PASCAL-5i under the 1-shot and 5-shot settings, respectively, while the corresponding results of CGPRNet are 64.1% and 69.6%.

  3. The authors name their module the Cross-Attention Graph Prototype Generator and emphasize that it can learn cross-scale support-query relationships through iterative graph reasoning. However, in Table 4, removing Graph Refinement only leads to average mIoU drops of 0.4 and 0.7 in the 1-shot and 5-shot settings, respectively. In contrast, removing the Dual-Branch Encoder or Multi-Scale Matching leads to performance drops of 6.9 and 6.6 in the 5-shot setting. This suggests that the most significant improvements in the experiments seem to come from the dual-branch feature encoding and multi-scale matching, rather than the Graph Prototype Refinement emphasized in the paper's title. This does not mean that the graph module has no value, but the authors need to more accurately clarify the contributions of each module.

Questions for authors
  1. Could the authors explain why the mIoU, Dice, and FB-IoU results of the full model with ResNet-50 on FSS-1000 are inconsistent across Tables 3, 4, and 5?
  2. Could the authors clarify how the four folds (F0/F1/F2/F3) on FSS-1000 are constructed and whether the evaluation follows the standard FSS-1000 protocol?
  3. Why are several recent and stronger few-shot segmentation methods, such as HDMNet, AMFormer, and HMNet, not included in the comparisons?
  4. The ablation study in Table 4 shows that removing Graph Refinement leads to relatively small performance drops compared with the other two modules. Could the authors further explain how the graph refinement module contributes to cross-scale support-query relationships and why its performance contribution is relatively limited?
3 comments

W1. Inconsistencies in FSS-1000 results across Tables 3, 4, and 5

We thank the reviewer for identifying these discrepancies. After carefully examining our training logs, we confirmed that the ResNet-50 results reported in Tables 3 and 5 were inadvertently taken from an intermediate 5-epoch checkpoint, whereas the full-model results in Table 4 correspond to the completed 50-epoch training schedule.

Importantly, all configurations in Table 4 were trained for 50 epochs under the same training settings. Therefore, the ablation comparisons are internally consistent. The discrepancy arose from incorrect checkpoint selection during preparation of the main results tables, rather than from different training budgets within the ablation study.

Table. FSS-1000 ResNet-50 results from the reported checkpoints (%).

Metric5 epochs (Tables 3/5)50 epochs (Table 4)Improvement
1-shot mIoU81.1283.60+2.48
5-shot mIoU82.1784.00+1.83
1-shot FB-IoU78.0980.90+2.81
5-shot FB-IoU78.6181.00+2.39
1-shot Dice88.7390.50+1.77
5-shot Dice89.3990.80+1.41

These results explain the systematic differences across all three metrics. We will correct Tables 3 and 5 using the verified final-checkpoint evaluation results, explicitly specify the training duration and checkpoint-selection protocol, and ensure consistency throughout the manuscript.

We acknowledge this reporting oversight and appreciate the reviewer for bringing it to our attention.

W2. FSS-1000 evaluation protocol and four-fold reporting

We thank the reviewer for highlighting the need for a clearer description of the FSS-1000 evaluation protocol.

The standard FSS-1000 benchmark contains 1,000 categories, partitioned into 520 training, 240 validation, and 240 testing categories. Our experiments use episodic KK-shot evaluation, where each episode consists of KK annotated support image-mask pairs and a query image from the target category.

The F0–F3 columns in our manuscript present fold-wise evaluation results, and the reported mean is calculated across these four folds. However, we acknowledge that the manuscript does not sufficiently specify the category assignments underlying these folds or their precise relationship to the official FSS-1000 test partition.

To address this concern, we will explicitly document the fold construction, category membership, training/validation/test partitions, episode-sampling procedure, and checkpoint-selection settings. We will also clearly distinguish any additional fold-based evaluation from the official FSS-1000 test-set protocol.

Furthermore, we will ensure that direct comparisons with existing methods are restricted to compatible evaluation settings, avoiding potentially misleading conclusions arising from different dataset partitions.

We appreciate this observation, as these clarifications are important for reproducibility and fair benchmarking.

W3. Missing recent baselines and overstated state-of-the-art claims

We appreciate the reviewer for identifying relevant recent methods, including HDMNet (CVPR 2023), AMFormer (NeurIPS 2023), and HMNet (NeurIPS 2024).

We agree that these methods should be discussed to provide a more complete assessment of recent progress. We also acknowledge that the statement claiming superior performance across all backbones and settings is too broad.

For example, AMFormer achieves stronger results than CGPRNet on PASCAL-5ⁱ with ResNet-101.

Table. PASCAL-5ⁱ comparison with AMFormer (ResNet-101, mean mIoU %).

Method1-shot5-shot
AMFormer70.7073.60
CGPRNet64.1069.60

Likewise, on COCO-20ⁱ with ResNet-101, HSNet achieves 49.5% 5-shot mean mIoU, compared with 44.6% for CGPRNet. We therefore do not claim that CGPRNet universally outperforms existing approaches.

Nevertheless, CGPRNet remains competitive in several evaluated configurations. For instance, on PASCAL-5ⁱ using ResNet-50, our method achieves 68.6% and 70.8% mean mIoU in the 1-shot and 5-shot settings, respectively, compared with 67.3% and 69.2% for RiFeNet.

The primary contribution of CGPRNet is its integration of three complementary mechanisms:

  1. Dual-branch feature adaptation. Preserves generic pretrained representations while incorporating task-specific information.
  2. Cross-attention graph prototype generation. Refines semantic prototypes using query-conditioned feedback and inter-prototype interactions.
  3. Multi-scale prototype matching. Combines information from different receptive fields through spatially adaptive weighting.

These mechanisms address complementary challenges in few-shot segmentation, rather than guaranteeing the best performance under every backbone and benchmark.

We will expand the related-work discussion to include HDMNet, AMFormer, and HMNet, incorporate additional quantitative comparisons where compatible results are available, correct the highlighted best-result markings, and revise the performance claims to accurately reflect the evidence.

W4. Contribution of graph prototype refinement

We thank the reviewer for questioning the relative contribution of graph refinement. We agree that the Dual-Branch Encoder and Multi-Scale Matching module produce larger improvements in the reported ablations, particularly in the 5-shot setting.

However, the graph component serves a distinct role: it enables query-conditioned prototype updates and interaction between foreground and background representations.

Table. Component-level ablation on FSS-1000 (ResNet-50, 50 epochs; mIoU %).

Configuration1-shot5-shotΔ 1-shotΔ 5-shot
Full CGPRNet83.6084.00——
w/o Dual-Branch82.5077.10−1.10−6.90
w/o Graph Refinement83.2083.30−0.40−0.70
w/o Multi-Scale Matching83.0077.40−0.60−6.60

These results show that graph refinement provides a consistent, although comparatively modest, contribution to the complete architecture.

To further examine the underlying design, we report the existing targeted prototype-graph ablations.

Table. Prototype-graph design ablation on FSS-1000 (ResNet-50, mIoU %).

Configuration1-shot5-shotΔ 1-shotΔ 5-shot
Full CGPRNet83.6084.00——
Softmax over K83.0083.50−0.60−0.50
w/o Background Node82.8083.20−0.80−0.80
w/o Inter-Prototype Interaction82.9083.30−0.70−0.70

Removing the background node reduces mIoU by 0.8 percentage points in both settings, suggesting that explicit background modeling is beneficial. Similarly, removing inter-prototype interaction decreases performance by 0.7 points in both settings, supporting the usefulness of information exchange among prototype nodes.

The attention-normalization ablation also shows that spatially normalized query feedback is preferable to prototype-axis normalization under the evaluated settings.

Importantly, we clarify that CGPRNet does not explicitly construct scale-specific graph nodes or cross-scale graph edges. Graph-based refinement operates on semantic prototypes, while multi-scale feature matching is performed subsequently by the matching module. These components are complementary but architecturally distinct.

We acknowledge that the overall graph contribution is smaller than those of the other major components. Accordingly, we will revise the manuscript to emphasize its role as a complementary refinement mechanism rather than attributing the majority of the performance improvement to graph reasoning.

We also recognize that the current ablations do not fully separate the effects of graph connectivity from additional learnable refinement capacity. We will clarify this limitation rather than making unsupported claims about the superiority of graph message passing.

Reviewer_AYL2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CGPRNet for few-shot semantic segmentation. The framework combines a frozen ImageNet-pretrained backbone with a trainable adapter branch, query-conditioned prototype refinement, and multi-scale prototype matching. Foreground prototypes from the support images and an additional background prototype form a fully connected graph. The prototypes are updated through attention to query features and inter-prototype message passing. Refined foreground prototypes are then matched against query features processed with different convolutional kernel sizes, and spatially adaptive weights combine the resulting similarity maps.

Experiments are reported on PASCAL-5i, COCO-20i, and FSS-1000 using VGG-16, ResNet-50, and ResNet-101. The strongest comparative results are obtained with ResNet-50, including 68.6%/70.8% mIoU on PASCAL-5i and 48.5%/58.8% on COCO-20i under the 1-shot/5-shot settings. Ablations examine the encoder, graph refinement, background node, scale weighting, and refinement depth.

Critical strengths

1. The architecture addresses a reasonable modeling problem through a coherent combination of components.

Conditioning support prototypes on the query can help accommodate support–query appearance differences. Combining this with an explicit background representation and spatially adaptive receptive-field weighting is a sensible design. The formulation also clarifies how multiple support examples participate as separate prototype nodes rather than being averaged before refinement. A single frozen backbone with trainable adaptation layers, as specified in Sections 2.3 and 3.2, is a practical alternative to training a second full backbone.

2. The ResNet-50 results are promising relative to the listed baselines, and the ablation coverage is reasonably broad.

On PASCAL-5i, the reported ResNet-50 results exceed RiFeNet by 1.5 and 1.2 percentage points in 1-shot and 5-shot mIoU. The reported COCO-20i improvements are larger. The paper also attempts to examine individual design choices, including the background node, inter-prototype messages, normalization axis, spatial scale gating, and number of refinement steps. If the inconsistencies in the experimental reporting can be resolved, these experiments could provide a useful starting point for understanding the method.

Critical weaknesses

1. Substantial inconsistencies in the experimental results and evaluation protocol prevent a reliable assessment of performance.

Several discrepancies affect central results rather than peripheral details:

  • Different full-model results on FSS-1000. Table 3 reports ResNet-50 mIoU of 81.12%/82.17%, whereas Table 4 reports 83.6%/84.0% for the full model, also described as ResNet-50 results averaged over four folds. The Dice results similarly differ between Tables 4 and 5. No change in configuration or protocol is stated.
  • Unclear FSS-1000 splits. The paper claims to follow standard splits but reports four-fold FSS-1000 evaluation without defining those folds or explaining their relationship to the benchmark’s standard class-disjoint training, validation, and test partitions.
  • Claims contradicted by the tables. Section 3.3 states that the ResNet-101 and VGG-16 models are “best in every column” on COCO-20i. However, ResNet-101 obtains 44.6% 5-shot mIoU, below HSNet’s 49.5%, and VGG-16 is also not best in every fold. On PASCAL-5i, the VGG-16 5-shot result is below HSNet, as are the ResNet-101 mean results in both shot settings. Thus, the claims of superiority “across all settings” are not supported even by the selected comparisons.
  • Contradictory refinement-depth interpretation. Table 6 shows a substantial decline from T=2T=2 to T=3T=3: 68.60% to 62.81% in 1-shot and 70.80% to 63.90% in 5-shot. The text instead states that additional iterations provide “marginal gains.”
  • Conflicting implementation and performance summaries. The limitations paragraph quotes COCO-20i results of 47.93% for VGG-16 and 49.62% for ResNet-101, inconsistent with Table 2. The manuscript states approximately 10M trainable parameters, while Table 6 reports 4.87M at the default T=2T=2, without explaining the accounting scope.

No episode counts, repeated-run uncertainty, or sufficiently detailed model-selection protocol are provided to help interpret these discrepancies. Moreover, the FSS-1000 tables compare the authors’ own backbones rather than competing methods, so they do not establish state-of-the-art performance on that benchmark.

2. The central “cross-scale graph refinement” claim is not established by the formulation, and the incremental contribution is insufficiently isolated.

In Section 2.4, graph nodes are foreground support prototypes and a background prototype. Query feedback is computed from a single fused feature map. Explicit multi-scale processing appears only afterward, in Section 2.5, through different convolutional kernels applied to query features. As written, this is query-conditioned prototype refinement followed by multi-receptive-field matching; it does not explicitly model relationships between scale-indexed support and query representations within the graph. This distinction matters because cross-scale graph reasoning is the paper’s main motivation and title-level claim.

The graph’s behavior in the 1-shot setting also deserves closer treatment. With one foreground node and one background node, excluding self-loops leaves exactly one neighbor for each node. Therefore, Equation (5) assigns that neighbor weight one regardless of the learned affinity. Message passing remains possible, but learned inter-node attention cannot select among neighbors in this setting.

The paper does not adequately distinguish its contribution from the cited scale-aware graph network, iterative-refinement methods, adaptive prototype methods, and support–query attention approaches. Table 4 reports only 0.4/0.7 percentage-point gains from graph refinement, without uncertainty estimates or a parameter-matched non-graph refinement control. There is also no scale-stratified evaluation demonstrating that the proposed graph addresses support–query scale mismatch. Together, these omissions leave the central mechanism and its claimed advantage insufficiently supported.

Questions for authors

1. Which experimental results correspond to the final model and evaluation protocol?
Please reconcile the FSS-1000 full-model results in Tables 3–5, the COCO-20i numbers in the limitations paragraph, and the 10M versus 4.87M parameter counts. Please also correct the statements of superiority and the interpretation of Table 6. If different settings produced these numbers, specify those settings and identify the results used for the main claims.

2. What exactly is the evaluation and model-selection protocol?
Please define the four FSS-1000 folds and their relationship to the standard train/validation/test class split. How many episodes are evaluated per dataset and fold? Are support and query images distinct, and are identical evaluation episodes used across methods? What validation data select the checkpoint and T=2T=2? Are 1-shot and 5-shot models trained separately? Please report repeated-run or repeated-sampling variability for the main results and ablations.

3. Where does cross-scale interaction occur within the graph?
Equations (1)–(9) appear to operate on support prototypes and a single-resolution query representation, while multiple receptive fields are introduced only after refinement. Does the implementation contain scale-specific nodes or cross-scale edges that are absent from the description? If not, please clarify the claim and provide evidence distinguishing the benefit of graph refinement from that of the subsequent multi-scale matching module.

4. What is the specific benefit of learned graph message passing?
For 1-shot evaluation, Equation (5) has only one eligible neighbor per node, so its normalized edge weight is always one. How does the method’s interpretation account for this case? Can the authors compare the graph against a parameter-matched query-conditioned MLP or attention-only refinement baseline? In Table 4, what precisely is removed by “w/o Graph Refine,” and how are attended query features constructed in that variant?

5. How should the reported backbone and refinement-depth sensitivity be interpreted?
On COCO-20i, ResNet-50 exceeds ResNet-101 by 6.6 points in 1-shot and 14.2 points in 5-shot, despite the claimed modest backbone dependence. Table 6 also shows a sharp optimum at T=2T=2, rather than gradual improvement or saturation. Are these patterns reproducible under matched training and model-selection settings? Clarifying them would materially affect the claims of robustness and effective progressive refinement.

4 comments

We sincerely thank the reviewer for the detailed and constructive evaluation of CGPRNet. We appreciate the recognition of our motivation and the potential benefits of query-conditioned prototype refinement, explicit background modeling, and adaptive multi-scale matching. We carefully examined the concerns regarding experimental consistency, evaluation reproducibility, architectural novelty, graph message passing, and refinement sensitivity. We address each point below.

W1. We thank the reviewer for identifying the discrepancies between Tables 3–5.

After verifying our training records, we confirmed that the ResNet-50 results in Table 3 were inadvertently taken from an intermediate 5-epoch checkpoint, whereas the full-model results in Table 4 were obtained after the complete 50-epoch training schedule. Importantly, all configurations in the Table 4 ablation study were trained for 50 epochs.

Table. Reconciliation of FSS-1000 results (ResNet-50, %).

MetricTable 3 (5 epochs)Table 4 (50 epochs)Difference
1-shot mIoU81.1283.60+2.48
5-shot mIoU82.1784.00+1.83
1-shot FB-IoU78.0980.90+2.81
5-shot FB-IoU78.6181.00+2.39
1-shot Dice88.7390.50+1.77
5-shot Dice89.3990.80+1.41

The discrepancy therefore reflects an intermediate-checkpoint reporting error, rather than unequal training budgets among the ablation configurations. We will correct Table 3 using the final-checkpoint benchmark evaluation and reconcile the corresponding Dice results in Table 5. We will also make the checkpoint-selection procedure explicit.

We additionally acknowledge the following reporting inconsistencies:

  • COCO-20ⁱ. The values quoted in the limitations paragraph do not match Table 2. We will correct the text using the actual benchmark results.
  • Parameter count. The approximately 10M trainable-parameter statement is inconsistent with the 4.87M reported for T=2 in Table 6. We will clarify the counting scope and correct the inconsistent statement.
  • State-of-the-art claims. We agree that the claim of superiority in every COCO-20ⁱ column is overstated. For example, the ResNet-101 5-shot mIoU of 44.6% is below HSNet's 49.5%. We will restrict our claims to comparisons supported by the reported results.

We acknowledge these reporting errors and appreciate the opportunity to correct them.

W2.

We agree that reproducibility requires precise documentation of dataset partitions, evaluation episodes, model selection, and statistical variability.

Our framework follows episodic few-shot segmentation, where each episode contains KK annotated support image-mask pairs and one query image from the target class. Training and evaluation use disjoint base and novel classes, as described in Section 2.1.

The implementation settings are summarized below.

Table. Experimental configuration.

ComponentConfiguration
DatasetsFSS-1000, PASCAL-5ⁱ, COCO-20ⁱ
BackbonesResNet-50, ResNet-101, VGG-16
PretrainingImageNet
Backbone weightsFrozen
Feature dimension256
Graph MLP hidden dimension128
Refinement iterationsT=2
Training epochs50
OptimizerAdamW
Learning rate1 × 10⁻⁴
Weight decay1 × 10⁻⁴
Batch size8
Gradient clippingMaximum norm 1.0
Input resolution473 × 473
Output stride8
Evaluation metricsmIoU, FB-IoU, Dice

We recognize that the original manuscript does not sufficiently specify the four FSS-1000 folds or their relationship to the standard benchmark partition. We will explicitly document their class membership and distinguish standard benchmark evaluation from any additional fold-based analysis.

We will also clarify the evaluation-episode count, support-query sampling constraints, random seeds, shared evaluation episodes across methods, checkpoint-selection criterion, and training arrangements for the 1-shot and 5-shot settings.

Regarding statistical reliability, we agree that point estimates alone cannot establish whether small improvements are statistically significant. The existing results do not include repeated-run confidence intervals. Accordingly, we do not claim statistical significance for the modest graph-refinement improvements without additional evidence.

These clarifications will make the evaluation protocol more transparent and allow readers to assess the comparability and reproducibility of the reported results.

W3.

We appreciate the reviewer for identifying the distinction between explicit cross-scale graph interactions and multi-scale matching.

We clarify that CGPRNet does not construct scale-specific graph nodes or explicitly learn graph edges between different feature scales. Instead, the architecture combines two complementary mechanisms.

First, the Cross-Attention Graph Prototype Generator (CAGPG) refines foreground and background prototypes through query-conditioned spatial attention and inter-prototype message passing.

Second, the Multi-Scale Prototype Matching module (MSPM) evaluates the refined prototypes using query representations with 3×3, 5×5, and 7×7 receptive fields. The resulting similarity maps are combined through spatially adaptive weighting.

Thus, graph refinement operates on semantic prototypes, whereas scale adaptation occurs during multi-scale matching. We agree that the original terminology could imply explicit cross-scale graph edges and will revise the description accordingly.

Table. Component ablation on FSS-1000 (ResNet-50, 50 epochs).

Configuration1-shot mIoU5-shot mIoUΔ 1-shotΔ 5-shot
Full CGPRNet83.6084.00——
w/o Dual-Branch82.5077.10−1.10−6.90
w/o Graph Refinement83.2083.30−0.40−0.70
w/o Multi-Scale Matching83.0077.40−0.60−6.60

The ablation results show that removing Multi-Scale Matching reduces 5-shot mIoU by 6.60 percentage points, while removing Graph Refinement reduces it by 0.70 points. These results indicate that multi-scale matching contributes more strongly to aggregate performance, while graph refinement provides a smaller complementary improvement.

We therefore do not interpret the existing results as evidence that graph refinement independently resolves scale mismatch. Rather, they support the combined architecture.

We also acknowledge that the manuscript does not provide a scale-stratified evaluation separating objects by size or support-query scale discrepancy. Consequently, we will avoid claiming that the graph mechanism itself has been demonstrated to improve scale robustness.

W4. We thank the reviewer for the precise mathematical observation regarding Equation (5).

For K=1K=1, the graph contains two nodes: one foreground prototype and one background prototype. Because self-loops are excluded, each node has exactly one eligible neighbor. Therefore, the normalized edge weight becomes

Gk,l=1,l≠k.G_{k,l} = 1, \qquad l \neq k .

We agree that the edge-attention mechanism cannot select among multiple neighbors in this configuration.

However, the message-passing operation does not become an identity mapping. The neighboring prototype is still processed through a learned transformation, and its message contributes to the residual prototype update. Thus, in the 1-shot setting, the graph component performs learned foreground-background message transformation rather than adaptive neighbor selection.

In the 5-shot setting, the graph contains six nodes, allowing each node to aggregate information from five eligible neighbors through nontrivial learned affinity weights.

Importantly, Table 4 includes targeted ablations examining the prototype-graph design.

Table. Prototype-graph ablations on FSS-1000 (ResNet-50).

Configuration1-shot mIoU5-shot mIoUΔ 1-shotΔ 5-shot
Full CGPRNet83.6084.00——
w/o Graph Refinement83.2083.30−0.40−0.70
Softmax over K83.0083.50−0.60−0.50
w/o Background Node82.8083.20−0.80−0.80
w/o Inter-Prototype Interaction82.9083.30−0.70−0.70

Removing the background node decreases both 1-shot and 5-shot mIoU by 0.80 points, supporting the contribution of explicit background modeling.

Removing inter-prototype interaction reduces performance by 0.70 points in both settings. This provides more targeted evidence for the usefulness of prototype interaction than the complete graph-removal ablation alone.

Furthermore, replacing spatial attention normalization with prototype-axis softmax decreases performance by 0.60 and 0.50 points, supporting our choice of spatially normalized prototype-query attention.

Nevertheless, we agree that these ablations do not completely separate graph structure from additional learnable transformation capacity. A parameter-matched query-conditioned MLP or attention-only refinement baseline would provide a more decisive comparison. We therefore characterize the existing results as evidence for the usefulness of prototype interaction within CGPRNet, rather than proof of superiority over all alternative refinement mechanisms.

We will also clarify the exact operations removed in the "w/o Graph Refine" configuration.

W5.

Adaptive multi-scale weighting. Beyond removing the entire Multi-Scale Matching module, our existing ablation compares spatially adaptive scale weighting with global weighting.

Table. Multi-scale weighting ablation on FSS-1000.

Configuration1-shot mIoU5-shot mIoUΔ 1-shotΔ 5-shot
Full adaptive weighting83.6084.00——
Global scale weights83.1083.70−0.50−0.30
w/o Multi-Scale Matching83.0077.40−0.60−6.60

The spatially adaptive weighting mechanism improves mIoU by 0.50 and 0.30 points over global weighting in the 1-shot and 5-shot settings, respectively. These findings support a modest additional benefit from spatial adaptation beyond incorporating multiple receptive fields.

Backbone sensitivity. We agree that the reported COCO-20ⁱ performance differences across ResNet-50 and ResNet-101 are substantial and cannot be characterized as uniformly modest. The existing experiments demonstrate backbone-dependent variation, but do not isolate whether it arises from feature quality, optimization, or interactions with the adaptation modules. We will revise the limitations discussion and avoid unsupported causal explanations.

Refinement-depth sensitivity. We also acknowledge that the original interpretation of Table 6 was incorrect.

Table. Refinement-depth ablation on PASCAL-5ⁱ (ResNet-50).

T1-shot mIoU1-shot FB-IoU5-shot mIoU5-shot FB-IoUGFLOPsParams (M)Latency (ms)
160.2169.8360.8470.115.84.4517
268.6078.7070.8079.306.34.8719
362.8171.7463.9072.416.85.2822
462.7871.7063.8572.367.35.7025

Increasing T from 1 to 2 improves 1-shot and 5-shot mIoU by 8.39 and 9.96 percentage points. However, increasing T from 2 to 3 reduces performance by 5.79 and 6.90 points, respectively. Therefore, the results demonstrate a non-monotonic relationship between refinement depth and performance. T=2 achieves the best observed accuracy among the evaluated configurations.

At T=2, the model requires 6.3 GFLOPs and 19 ms latency, compared with 7.3 GFLOPs and 25 ms at T=4.

We also acknowledge that the number of parameters increases with T, meaning that refinement depth and model capacity are not independently controlled. Accordingly, the current results establish an empirical configuration preference rather than a causal explanation for the performance degradation at larger T.

Finally, we will correct the Table 6 caption, as the T=2 results correspond to the four-fold means in Table 1 rather than the fold-specific values indicated in the original caption.

Reviewer_xSwqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes CGPRNet, which combines a frozen shared CNN backbone with a trainable adapter, query-conditioned prototype refinement, and spatially adaptive matching across three receptive fields. Foreground and background prototypes exchange information with query features and with each other before segmentation. The paper evaluates the method on PASCAL-5, COCO-20, and FSS-1000.

Critical strengths

Encouraging ResNet-50 results. The reported PASCAL-5 results improve over RiFeNet from 67.1/69.6 to 68.6/70.8 mIoU for one/five shots. The reported COCO-20 improvements are larger, particularly in five-shot segmentation.

Critical weaknesses
  1. There are several central experimental results contradict one another. Table 3 reports FSS-1000 ResNet-50 mIoU of 81.12/82.17, whereas Table 4 gives 83.6/84.0 for the full model on the same dataset and backbone.

  2. The iteration analysis is also misinterpreted: Table 6 shows performance falling from 68.60/70.80 at T=2T=2 to 62.81/63.90 at T=3T=3, while the text says additional iterations provide marginal gains. Its caption specifies PASCAL-5, but the T=2T=2 values reproduce the four-fold means in Table 1 rather than its fold-0 results.

  3. The claim that the ResNet-101 model is best in every COCO column is contradicted by Table 2: its five-shot mIoU is 44.6, below HSNet’s 49.5. Table 3 contains no comparisons with other methods on FSS-1000. Moreover, the original HSNet paper reports ResNet-50 results of 85.5/87.8 on that benchmark, above the submitted model’s 81.12/82.17.

  4. Several citations in the introduction do not appear to support the claims they accompany. Li et al. (2025), cited when discussing FSS feature fusion, concerns 3D point-cloud detection. Xie et al. (2022) addresses saliency detection, and Zhang et al. (2024) addresses human skeleton estimation, yet both are used to characterize existing FSS methods. There are also reporting errors: lines 139–142 refer to Sections 3.3–3.6 instead of 2.3–2.6, and line 240 contains an unresolved “Eq. (??)” reference. Precision is introduced as an evaluation metric in Section 3.1 but is never reported.

Questions for authors

See weakness section.

4 comments

We sincerely thank the reviewer for the thorough evaluation of our manuscript and for identifying several important issues concerning experimental consistency, performance comparisons, and presentation. We appreciate the opportunity to clarify these concerns and strengthen the technical interpretation of our work. Due to the comment length limit, our response is split across several comments.

*W1. We sincerely thank the reviewer for identifying the discrepancies between Tables 3 and 4. Following this observation, we carefully re-examined our experimental protocols, training records, and checkpoint selection to identify the source of the inconsistency.

Our verification confirmed that the ablation results reported in Table 4 were obtained using the complete 50-epoch training protocol, consistent with Section 3.2. However, the ResNet-50 results reported in Table 3 were inadvertently taken from an intermediate 5-epoch evaluation rather than the final 50-epoch evaluation.

Table. Comparison of the originally reported FSS-1000 results (ResNet-50, %).

MetricMain evaluation (5 epochs)Full-model ablation (50 epochs)
1-shot mIoU81.1283.60
5-shot mIoU82.1784.00
1-shot FB-IoU78.0980.90
5-shot FB-IoU78.6181.00
1-shot Dice88.7390.50
5-shot Dice89.3990.80

The discrepancy therefore originated from an incorrect selection of intermediate-checkpoint results for the main benchmark table, rather than from differences in the ablation training protocol.

Importantly, all configurations in the ablation study were trained using the same 50-epoch schedule, ensuring a consistent training budget for evaluating the contributions of the Dual-Branch Feature Encoder, Graph Prototype Refinement, and Multi-Scale Prototype Matching modules.

Table. Component-wise ablation results on FSS-1000 (50 epochs, ResNet-50).

Configuration1-shot mIoU5-shot mIoU5-shot Δ
Full CGPRNet83.6084.00—
w/o Dual-Branch82.5077.10−6.90
w/o Graph Refinement83.2083.30−0.70
w/o Multi-Scale Matching83.0077.40−6.60

Removing the Dual-Branch Encoder causes the largest 5-shot degradation (6.9 percentage points), followed by Multi-Scale Matching (6.6 points), while removing Graph Refinement produces a smaller but measurable decrease (0.7 points). These comparisons support the complementary contributions of the proposed components under a consistent training budget.

We acknowledge this reporting oversight and apologize for the confusion. We will correct Table 3 using the verified final 50-epoch evaluation results and ensure that the corresponding mIoU, FB-IoU, and Dice scores are consistently reflected across the manuscript. We have additionally audited the remaining result tables against their training records and confirmed that the checkpoint-selection error was confined to the ResNet-50 entries of Table 3.

We will also clarify the checkpoint-selection procedure and explicitly distinguish intermediate evaluations from final benchmark reporting in the revised experimental section.

We appreciate the reviewer for bringing this issue to our attention. This correction will improve the consistency, transparency, and reproducibility of our experimental evaluation.

*W2. We appreciate the reviewer for highlighting the inconsistency between the refinement iteration results and their interpretation. We re-examined Table 6 and agree that the original statement suggesting marginal improvements with additional iterations was inaccurate.

Table. Effect of refinement iterations on segmentation performance.

Iterations1-shot mIoU (%)5-shot mIoU (%)GFLOPsParameters (M)
T=160.2160.845.84.45
T=268.6070.806.34.87
T=362.8163.906.85.28
T=462.7863.857.35.70

The results demonstrate that increasing refinement depth from T=1 to T=2 improves 1-shot mIoU by 8.39 percentage points and 5-shot mIoU by 9.96 percentage points. However, further increasing the refinement depth to T=3 reduces performance by 5.79 and 6.90 percentage points, respectively, compared with T=2. This indicates that T=2 provides the most favorable balance between segmentation performance and computational complexity among the evaluated configurations.

From an architectural perspective, query-conditioned cross-attention enables the support prototypes to incorporate query-specific semantic information, while graph message passing facilitates interactions among foreground and background prototypes.

The observed performance degradation at higher iteration counts suggests that additional refinement is not necessarily beneficial. Repeated feature aggregation may introduce redundant information or reduce prototype discriminability, although establishing the precise cause would require further analysis.

Therefore, we will revise the discussion to emphasize that two refinement iterations achieve the best performance among the evaluated configurations, rather than claiming that increasing the iteration count consistently improves performance.

We also acknowledge the inconsistency in the Table 6 caption. The reported T=2 values (68.60 / 70.80) match the four-fold means in Table 1 rather than the PASCAL-5⁰ fold-specific values (69.30 / 71.50). We will verify the corresponding evaluation records and correct the caption or reported values accordingly.

W3.

We thank the reviewer for carefully examining our comparisons with existing few-shot segmentation methods.

We acknowledge that our original manuscript overstated the generality of its state-of-the-art claims. In particular, the statement that CGPRNet achieves the best performance across every COCO-20ⁱ configuration is not supported by all reported results.

Table. Comparison with HSNet based on reported benchmark results. CGPRNet FSS-1000 values are the corrected 50-epoch results described in our response to W1.

DatasetMethodBackbone1-shot mIoU5-shot mIoU
COCO-20ⁱHSNetResNet-101—49.50
COCO-20ⁱCGPRNetResNet-10141.9044.60
FSS-1000HSNetResNet-5085.5087.80
FSS-1000CGPRNetResNet-5083.6084.00

We recognize that HSNet outperforms the reported CGPRNet configurations in these specific comparisons. Accordingly, we will remove unsupported claims of universal superiority from the abstract, introduction, and conclusion.

Nevertheless, we would like to emphasize that the primary contribution of CGPRNet lies in its integrated architectural design, which addresses three complementary challenges in few-shot segmentation.

  1. Query-conditioned graph prototype refinement. Unlike conventional static masked-average-pooled prototypes, our Cross-Attention Graph Prototype Generator incorporates query-specific information through iterative cross-attention and inter-prototype message passing. The explicit background node additionally enables foreground-background interactions and prevents prototype-axis attention from becoming degenerate in the single-support setting.
  2. Adaptive multi-scale prototype matching. Our Multi-Scale Prototype Matching module computes similarity across 3×3, 5×5, and 7×7 receptive fields and learns spatially adaptive scale weights. This allows the matching process to accommodate variations in object size and local semantic structure.
  3. Shared-backbone dual-branch feature encoding. Our architecture combines frozen generic representations with lightweight trainable task-specific features through learned fusion gates, aiming to preserve pretrained knowledge while improving task adaptation without requiring a second fully trainable backbone.

The component-wise ablation results in Table R2 provide evidence for the contribution of these mechanisms under a consistent 50-epoch training protocol.

We agree that the novelty and effectiveness of CGPRNet should be assessed against existing graph-based, attention-based, and multi-scale segmentation methods rather than relying solely on aggregate benchmark rankings.

We will therefore revise our comparative discussion to accurately acknowledge stronger competing results, clarify the distinction between our proposed mechanisms and related methods, and restrict performance claims to the experimental evidence.

W4. We appreciate the reviewer for identifying inaccuracies in the references and experimental descriptions.

Citation corrections. We acknowledge that Li et al. (2025), Xie et al. (2022), and Zhang et al. (2024) were cited in contexts that do not accurately reflect their respective research contributions. We will replace these references with appropriate few-shot segmentation literature and revise the related-work discussion accordingly.

Section and equation references. We will correct the section references in lines 139–142 from Sections 3.3–3.6 to Sections 2.3–2.6 and resolve the missing equation reference at line 240.

Evaluation metrics. We agree that defining Precision without reporting corresponding results creates an inconsistency. We will revise the evaluation section to ensure that all discussed metrics are supported by reported experimental results.

Implementation and numerical consistency. We will also reconcile the parameter-count description and any numerical values in the discussion that do not match the corresponding experimental tables. In particular, the trainable-parameter count must be reported consistently with the model configuration and measurement convention used in the efficiency analysis.

These revisions will improve the technical accuracy, reproducibility, and overall clarity of the manuscript.

Comments on the paper

1 comment

Dear Authors,

We are approaching the end of the discussion period, and I note that no response has been provided to the reviewers' comments.

Please confirm if you intend to submit a rebuttal. If you have decided to withdraw this submission, please officially withdraw it through the system. If you are no longer able to respond, I will proceed with the evaluation based on the current reviews on file.

Best regards, Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.