acceptodds
Under review as a conference paper at ICLR 2027

FoRIS: Progressive Foreground Refinement for Training-Free In-Context Segmentation

Abstract

In-Context Segmentation (ICS) aims to precisely segment arbitrary semantic concepts, such as objects or parts, given one or a few annotated visual exemplars. In this paper, we revisit ICS from a more classical segmentation perspective, viewing it as a coarse-to-fine progressive refinement process. Rather than directly predicting the final mask through reference-query matching, we progressively refine the segmentation from coarse and ambiguous foreground responses to precise and complete foreground structures. Building upon this perspective, we propose a training-free in-context segmentation framework, termed FoRIS. Specifically, FoRIS consists of three key stages: Foreground Purification, Foreground Localization, and Foreground Consolidation, which progressively suppress background distractions, localize discriminative target regions, and recover complete foreground structures through semantic aggregation. Experimental results demonstrate that FoRIS achieves SOTA performance across semantic and part segmentation tasks, with average improvements of 4.5 and 4.8 mIoU points over existing approaches in the 1-shot and 5-shot settings, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Training-Free, In-Context Segmentation, DINOv3

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NDJgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FoRIS, a training-free framework for ICS that reformulates reference-query matching as a progressive foreground refinement process. The method consists of three components that foreground purification suppresses background interference and positional bias, foreground localization identifies reliable foreground regions through feature correspondence and multi-cue clustering, and foreground consolidation refines the segmentation results through semantic disagreement penalization and cluster-level reweighting. Experiments across multiple benchmarks demonstrate consistent improvements over existing training-free approaches.

Critical strengths
  1. The paper is well organized, and the motivation is clear.
  2. FoRIS achieves consistent improvements over existing methods across diverse datasets.
  3. The ablation studies are relatively comprehensive.
Critical weaknesses
  1. Although the method reformulates feature matching as a three-stage progressive refinement process, the unique contributions of the individual components could be further clarified, as several design choices appear closely related to those in INSID3.
  2. The response maps in Eqs. (13) and (17) are simply added together. Could the authors explain why this works well and whether different weights would affect the results?
  3. FoRIS achieves promising improvements but introduces considerable computational overhead compared with INSID3. Mybe providing a more detailed discussion of the trade-off between mIoU and inference efficiency would be better.
Questions for authors

Please see the questions raised in the Weaknesses section.

5 comments

We sincerely thank you for your professional and insightful review. Our responses are provided below:

Q1:

Core DimensionINSID3 Limitations (Old Paradigm)FoRIS Innovations (Our Advances)Key Problems Addressed / Advantages
Overall ParadigmOne-shot hard decision: Relies on hard-threshold intersection and cluster merging, resulting in limited error tolerance.Three-stage progressive soft refinement: S1→S2→S3S^1\rightarrow S^2\rightarrow S^3Prevents early mistakes from causing complete failure (e.g., empty predictions) and enables progressive coarse-to-fine refinement.
Debiasing and PrototypesUniform global debiasing distorts feature geometry; a single mean prototype fails to capture appearance diversity.Adaptive Positional Debiasing (APD) + Multiple Foreground Prototypes + Hard-Negative Background ModelingPreserves the original feature geometry when semantic evidence is strong, captures multimodal foreground appearances, and suppresses confusing background regions using Top-20% hard negatives.
Feature Gating and ScoringLacks explicit background suppression and dense response maps.Contrastive Feature Gating (gg) + Background-Orthogonalized Dense Response (S1S^1)Suppresses background interference at both feature and response levels, enhancing discrimination between foreground and confusing background regions.
Spatial Localization CuesHard intersection may discard thin structures; clustering based solely on DINO features produces coarse boundaries.Soft Voting (VV) and Soft Prior (PP) + Joint DINO/RGB/XY ClusteringIntegrates appearance and spatial information, converting localization cues into continuous response updates (S2S^2) for more accurate object localization and boundary alignment.
Conflict Resolution and ConsolidationLacks explicit evidence-conflict handling, allowing local noise to contaminate the final mask.Semantic Disagreement Penalty (SDP, Π\Pi) + Cluster-wise Semantic Refinement (SR, Δ\Delta)Explicitly penalizes foreground-background inconsistencies and conflicts across cues, while consolidating multi-source evidence (S3S^3) to suppress artifacts and improve prediction quality.

As shown in the table above, the key distinction lies not in the shared encoder or positional basis, but in how visual evidence is processed and integrated. Specifically, FoRIS introduces three complementary improvements:

  • Foreground Purification (FP): APD adaptively removes positional bias, while explicit foreground-background contrast and feature gating improve the reliability of foreground representations.

  • Foreground Localization (FL): Instead of relying on the hard intersection of forward and backward correspondences, FoRIS integrates soft voting and clustering-based spatial evidence to construct a more robust localization response.

  • Foreground Consolidation (FC): FoRIS explicitly models disagreements among different sources of evidence and performs cluster-wise semantic refinement, progressively correcting the intermediate response before the final prediction.

Therefore, FoRIS differs from INSID3 not merely by adding individual operations, but by introducing a unified progressive refinement mechanism that changes both the role of intermediate evidence and the way the final segmentation mask is generated.

Q 2.

We thank the reviewer for pointing out the potential scale mismatch among the intermediate response maps and the importance of validating the additive fusion coefficients. To address this, we clarify that our framework employs weighted additive fusion rather than simple unweighted addition. For simplicity, the original manuscript omitted the explicit weighting coefficients and their default values. Following the reviewer's suggestion, we have now included them in Eqs. (13) and (17) of the revised manuscript:

S(2)=S(1)+λV(V−12)+λP(P−12),S^{(2)} = S^{(1)} + \lambda_V\left(V-\frac{1}{2}\right) + \lambda_P\left(P-\frac{1}{2}\right),

S(3)=S(2)−λΠΠ+λΔΔ.S^{(3)} = S^{(2)} - \lambda_{\Pi}\Pi + \lambda_{\Delta}\Delta.

The additive formulation is motivated by the fact that all response maps share the same target patch grid and provide spatially aligned complementary evidence. Specifically, S(1)S^{(1)} provides the primary semantic response, while the auxiliary maps progressively refine foreground predictions through geometric localization and regional semantic consistency. Additive fusion integrates these complementary signals while preserving the original semantic response, without introducing additional learnable fusion modules. Moreover, the explicit weighting coefficients control the contributions of different auxiliary signals and prevent them from overwhelming the primary response. The default settings are (λV,λP)=(0.20,0.25)(\lambda_V,\lambda_P)=(0.20,0.25) and (λΠ,λΔ)=(1,1)(\lambda_{\Pi},\lambda_{\Delta})=(1,1) To further investigate the impact of different weights, we conduct systematic weight sensitivity experiments on COCO, ISIC, and Lung. Specifically, we independently vary each coefficient in Eqs. (13) and (17), while keeping the remaining coefficients fixed at their default values. The corresponding experimental results and analyses have been added to the revised manuscript.

Table 1. Sensitivity to the localization fusion weights

λV\lambda_VλP\lambda_PCOCOISICLung
0.000.2560.2063.8085.40
0.200.2560.9062.9087.60
0.500.2556.8558.6088.50
1.000.2549.4053.1088.00
0.200.0059.1061.3088.20
0.200.5060.8561.9085.60
0.201.0059.6054.1080.90

Table 2. Sensitivity to the consolidation fusion weights

λΠ\lambda_{\Pi}λΔ\lambda_{\Delta}COCOISICLung
0.00.058.4560.2086.20
0.01.060.1762.6087.10
0.51.060.6562.9087.40
1.01.060.9062.9087.60
2.01.061.1262.6087.80
1.00.059.9061.1087.50
1.00.560.6062.3087.80
1.02.060.8363.6086.50

Q3: To provide a more comprehensive assessment, we conduct additional experiments on COCO-20$^i$ using a single RTX 3090, including a component-wise runtime analysis and a resolution-wise evaluation of the accuracy-efficiency trade-off. These experiments aim to identify the main sources of computational overhead and explore more efficient inference configurations.

(1) Accuracy-efficiency comparison with INSID3.

At the default input resolution of 1024 px, following the original setting of INSID3, FoRIS achieves 60.9% mIoU, outperforming INSID3 (57.6%) by 3.3 percentage points. Meanwhile, FoRIS requires 1,620.77 ms per in-context example, approximately 2.0$\times$ the 812 ms required by INSID3. To identify the sources of this additional overhead, we provide a component-wise runtime breakdown in Table below. Among the three refinement stages, Foreground Localization (346.44 ms) and Foreground Consolidation (342.50 ms) account for the largest computational costs, whereas Foreground Purification introduces only 38.74 ms of additional computation. The optional CRF refinement further contributes 190.82 ms. This analysis quantifies the computational cost of each component and identifies the main sources of inference overhead.

(2) Resolution-dependent accuracy-efficiency trade-off.

We further investigate the effect of input resolution on segmentation performance and inference efficiency to explore more efficient inference configurations. As shown in Table below, FoRIS achieves 59.5% mIoU at 512 px with an inference time of only 258 ms. Increasing the resolution to 720 px, 896 px, and 1024 px improves mIoU to 60.3%, 60.7%, and 60.9%, with corresponding runtimes of 593 ms, 1,108 ms, and 1,620.77 ms, respectively.

Notably, further increasing the resolution does not yield commensurate performance gains. For example, at 1440 px and 1600 px, the runtime increases to 5.47 s and 9.02 s, respectively, without improving mIoU. Although the 1760 px setting achieves a marginally higher mIoU of 61.0%, it requires 11.24 s per example, approximately 6.9$\times$ the runtime at 1024 px, for only a 0.1-point improvement. These results indicate that simply increasing the input resolution is computationally inefficient and provides limited benefits to segmentation performance.

Overall, FoRIS provides a flexible accuracy-efficiency trade-off through input resolution selection. We retain 1024 px as the default input resolution in our main experiments, following the original setting of INSID3 to ensure a consistent evaluation protocol and fair comparison. Meanwhile, our resolution-wise experiments demonstrate that lower resolutions can substantially reduce inference time while maintaining competitive segmentation performance, offering practical alternatives for efficiency-constrained applications.

Table: Detailed computational analysis of FoRIS on COCO-20$^i$.

We report the inference time of each component in milliseconds (ms, ↓\downarrow) for a single in-context example. Runtime is measured at a resolution of 1024 px on a single RTX 3090.

ComponentRuntime
Encoder forward702.27 ms
Foreground purification38.74 ms
Foreground localization346.44 ms
Foreground consolidation342.50 ms
CRF refinement190.82 ms
Total inference time1620.77 ms
INSID3812 ms

Table: Resolution-wise computational analysis of FoRIS on COCO-20$^i$.

We report the inference time and mIoU of FoRIS under different input resolutions. Runtime is measured for a single in-context example on a single RTX 3090. The 1024 px setting follows the default resolution of INSID3 and is used in the component-wise analysis.

Metric512 px720 px896 px1024 px1440 px1600 px1760 px
Runtime258 ms593 ms1108 ms1621 ms5468 ms9018 ms11235 ms
mIoU (%)59.560.360.760.960.860.161.0
Reviewer_cbBmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes FoRIS, a training-Free In-Context Segmentation method built on frozen DINOv3 features. The method refine the segmentation as progressive foreground refinement through foreground purification, foreground localization, and foreground consolidation.

  1. Foreground purification removes positional layout bias and background clutter via adaptive positional debiasing (APD)and two-stage foreground refinement (FR).
  2. Foreground localization refines score using debiased reference and target features.
  3. Foreground consolidation refines score again via semantic disagreement penalization (SDP) and semantic reweighting (SR).

This paper evaluates the method on one-shot semantic and part segmentation benchmarks, and the proposed method reaches SOTA performance.

Critical strengths
  1. The paper addresses a relevant problem: extracting reliable segmentation from frozen visual representations without task-specific optimization or a mask decoder. The proposed method target recognizable weaknesses of reference-query matching, namely background leakage, localization ambiguity, and fragmented response.

  2. The paper presents rich experimental evaluations, including baseline comparisons and detailed ablation studies to validate each component. Table 1 reports results compare with task-specific fine-tuning and training-free approaches over six semantic datasets and two part-segmentation datasets.

  3. The manuscript exhibits good readability. Its core insight and key technical approach are articulated very clearly. Fig 1 represents the high-level motivation well, and Fig 3 provides a helpful overview of a fairly elaborate training-free pipeline. The separation between purification, localization, and consolidation gives readers a useful conceptual organization.

Critical weaknesses
  1. Several hyper‑parameters in the paper are directly provided with their default values, without additional experiments to demonstrate their impacts on the proposed method. For example: beta, top 20%

  2. Certain statements in the manuscript lack sufficient precision. eg: FoRIS achieves the best performance on all semantic benchmarks. In fact, VIS-92i:SegIC 44.6 > FoRIS 42.8, COCO-20i:SegIC 76.1 > FoRIS 60.9

Questions for authors
  1. Please report the empirical range — and, if possible, the variance — of S^(1), (V − ½), (P − ½), Π, and Δ over actual episodes, and provide a sensitivity analysis for the fusion weights in Eqs. (13) and (17). Specifically: S^(2) contains the unnormalized S^(1) (Eq. 13). Please clarify whether S^(2) should be rescaled before Eq. (17) is applied.

  2. Please add controlled experiments on the key hyperparameters and show that the current values are reasonable design choices rather than values fitted to the evaluation benchmarks. Specifically: β = 0.07, the top-20% hard-negative ratio, the gate floor, and ε in Eq. (12).

4 comments

We sincerely thank you for your professional and insightful review. Our responses are provided below:

Q1:

We thank the reviewer for pointing out the potential scale mismatch among the intermediate response maps and the importance of validating the additive fusion coefficients.

(1) Sensitivity of additive fusion weights.

To investigate the relative contributions of different terms, we introduce explicit fusion coefficients into Eqs. (13) and (17):

S(2)=S(1)+λV(V−12)+λP(P−12),S^{(2)} = S^{(1)} + \lambda_V\left(V-\frac{1}{2}\right) + \lambda_P\left(P-\frac{1}{2}\right),

S(3)=S(2)−λΠΠ+λΔΔ.S^{(3)} = S^{(2)} - \lambda_{\Pi}\Pi + \lambda_{\Delta}\Delta.

We conduct controlled one-at-a-time sensitivity experiments on COCO (four-fold mean), ISIC, and Lung using an identical random seed. The default coefficients are (λV,λP)=(0.20,0.25)(\lambda_V,\lambda_P)=(0.20,0.25) and (λΠ,λΔ)=(1,1)(\lambda_{\Pi},\lambda_{\Delta})=(1,1).

For localization, the two priors provide complementary information. Removing VV decreases mIoU by 0.70 and 2.20 points on COCO and Lung, respectively, while removing PP decreases it by 1.80 and 1.60 points on COCO and ISIC. However, assigning unit weights to these bounded priors can overwhelm the original contrastive response. For example, increasing λV\lambda_V from 0.20 to 1.0 reduces COCO and ISIC performance by 11.50 and 9.80 points, respectively. Similarly, setting λP=1.0\lambda_P=1.0 decreases ISIC and Lung performance by 8.80 and 6.70 points.

These results demonstrate that localization priors should provide moderate corrections to the purified response rather than replace its underlying contrastive geometry. The default weights establish a balanced refinement regime across different domains.

For consolidation, removing both terms decreases mIoU by 2.45, 2.70, and 1.40 points on COCO, ISIC, and Lung, respectively. Removing either term individually also weakens performance, demonstrating the complementary roles of the disagreement penalty Π\Pi and cluster reweighting Δ\Delta.

Moreover, varying λΠ\lambda_{\Pi} between 0.5 and 2.0 results in only minor performance fluctuations, while increasing λΔ\lambda_{\Delta} to 2.0 benefits ISIC but slightly degrades Lung. Thus, the default unit coefficients provide a balanced cross-domain choice rather than an isolated benchmark-specific optimum.

(2) Clarification of score normalization.

Importantly, we clarify that S(2)S^{(2)} is not min--max normalized before Eq. (17). Instead, we directly apply the consolidation terms to the weighted localization response:

S(3)=S(2)−Π+Δ.S^{(3)}=S^{(2)}-\Pi+\Delta.

Only the final response S(3)S^{(3)} is normalized during the binarization stage.

This choice is supported by the fusion-weight sensitivity experiments. After applying the relatively small localization coefficients (λV=0.20\lambda_V=0.20 and λP=0.25\lambda_P=0.25), the resulting S(2)S^{(2)} retains the dominant contrastive structure of S(1)S^{(1)}, while the consolidation terms provide complementary corrections under unit coefficients. Therefore, an additional normalization step between localization and consolidation is unnecessary and is not used in our implementation.

We have added the corresponding fusion-weight sensitivity experiments and clarified the normalization procedure in the revised manuscript.

**Table 1. Sensitivity of localization fusion weights

Default: λV=0.20\lambda_V=0.20, λP=0.25\lambda_P=0.25.

λV\lambda_VλP\lambda_PCOCO ↑ISIC ↑Lung ↑
0.000.2560.2063.8085.40
0.200.2560.9062.9087.60
0.500.2556.8558.6088.50
1.000.2549.4053.1088.00
0.200.0059.1061.3088.20
0.200.5060.8561.9085.60
0.201.0059.6054.1080.90

**Table 2. Sensitivity of consolidation fusion weights

Default: λΠ=1\lambda_{\Pi}=1, λΔ=1\lambda_{\Delta}=1.

λΠ\lambda_{\Pi}λΔ\lambda_{\Delta}COCO ↑ISIC ↑Lung ↑
0.00.058.4560.2086.20
0.01.060.1762.6087.10
0.51.060.6562.9087.40
1.01.060.9062.9087.60
2.01.061.1262.6087.80
1.00.059.9061.1087.50
1.00.560.6062.3087.80
1.02.060.8363.6086.50

Step-wise gains of progressive refinement (folds separated).

Each entry is the class-wise mIoU gain (%) relative to the corresponding baseline: +V+V/+P+P/S(2)S^{(2)} vs. S(1)S^{(1)}; −Π-\Pi/+Δ+\Delta/S(3)S^{(3)} vs. S(2)S^{(2)}; and S(3)S^{(3)} vs. S(1)S^{(1)}.

Split+V+V+P+PS(2)S^{(2)}−Π-\Pi+Δ+\DeltaS(3)S^{(3)}S(3)−S(1)S^{(3)}{-}S^{(1)}
ISIC-0.13-0.26+1.47+0.85+2.43+2.72+4.19
X-Ray+8.04+2.18+7.24+1.24+0.82+1.34+8.58
SUIM+4.19+4.69+6.02-0.07+0.44+0.13+6.15
iSAID-0+3.70+1.71+4.46+1.35+1.01+1.89+6.35
iSAID-1+2.01+2.91+3.78+0.26+0.45+0.19+3.97
iSAID-2+2.98+3.73+4.67+0.45+0.04+0.21+4.88
COCO-0+3.34+6.99+7.71+1.62+1.17+2.27+9.98
COCO-1+1.48+6.27+6.38+1.00+1.36+1.54+7.92
COCO-2+1.39+5.28+4.64+1.41+2.35+3.54+8.18
COCO-3+2.17+6.47+6.11+1.85+2.24+2.64+8.75
PASCAL-Part-0+8.05+1.78+6.97+2.37+1.34+2.89+9.86
PASCAL-Part-1+2.54+3.17+6.00-1.50-0.72-1.79+4.21
PASCAL-Part-2+6.46+5.51+8.05+1.26+0.46+1.30+9.35
PASCAL-Part-3+6.42+2.04+6.22+1.10+0.13+0.75+6.97
PACO-Part-0+4.43+0.19+3.60+1.73+0.85+2.15+5.75
PACO-Part-1+6.38+1.28+5.30+2.19+1.24+3.00+8.30
PACO-Part-2+5.04-0.42+3.50+1.81+1.15+2.60+6.10
PACO-Part-3+3.99+0.43+2.43+1.86+1.20+2.71+5.14
LVIS-0+2.98+4.24+4.83+2.94+2.34+4.22+9.05
LVIS-1+4.97+4.22+6.50+2.91+1.73+3.75+10.25
LVIS-2+4.42+4.92+6.06+3.49+2.16+4.89+10.95
LVIS-3+5.14+5.27+7.03+3.64+2.59+4.76+11.79
LVIS-4+4.12+4.55+5.84+3.08+2.42+4.51+10.35
LVIS-5+4.86+6.90+7.99+2.78+2.38+4.24+12.23
LVIS-6+2.53+4.94+4.76+2.63+2.18+4.08+8.84
LVIS-7+4.37+4.84+6.52+3.07+1.56+3.71+10.23
LVIS-8+4.73+4.41+6.46+3.24+2.33+4.83+11.29
LVIS-9+3.59+4.12+4.70+3.46+2.31+4.62+9.32

Stage-wise progressive results (folds separated). For each fold we report class-wise mIoU (%) ±\pm standard deviation over episodes. S(1)S^{(1)}: purification only; +V+V/+P+P: enable one localization term; S(2)S^{(2)}: enable both; −Π-\Pi/+Δ+\Delta: enable one consolidation term on S(2)S^{(2)}; S(3)S^{(3)}: full model.

SplitS(1)S^{(1)}+V+V+P+PS(2)S^{(2)}−Π-\Pi+Δ+\DeltaS(3)S^{(3)}
ISIC58.74$\pm$26.4058.61$\pm$24.7858.48$\pm$26.9460.21$\pm$25.9861.06$\pm$24.9862.64$\pm$24.7262.93$\pm$24.07
X-Ray79.00$\pm$11.3987.04$\pm$7.0281.18$\pm$10.6186.24$\pm$8.2487.48$\pm$7.3687.06$\pm$7.5587.58$\pm$7.00
SUIM52.93$\pm$30.7457.12$\pm$30.6457.62$\pm$31.4458.95$\pm$31.0158.88$\pm$30.7959.39$\pm$31.1459.08$\pm$30.96
iSAID-057.53$\pm$27.1861.23$\pm$27.6259.24$\pm$27.6061.99$\pm$27.5363.34$\pm$27.6163.00$\pm$27.8163.88$\pm$27.90
iSAID-136.86$\pm$28.4738.87$\pm$28.8639.77$\pm$30.3540.64$\pm$30.2440.90$\pm$30.3341.09$\pm$30.4640.83$\pm$30.52
iSAID-251.26$\pm$27.0154.24$\pm$27.2954.99$\pm$26.8255.93$\pm$27.2956.38$\pm$27.3755.97$\pm$27.3356.14$\pm$27.46
COCO-048.86$\pm$28.5852.20$\pm$28.4855.85$\pm$28.8356.57$\pm$28.5958.19$\pm$28.0857.74$\pm$28.5558.84$\pm$28.38
COCO-155.60$\pm$29.2557.08$\pm$28.6061.87$\pm$28.3961.98$\pm$28.4762.98$\pm$27.8463.34$\pm$28.2663.52$\pm$27.81
COCO-252.93$\pm$28.6054.32$\pm$27.7758.21$\pm$28.0257.57$\pm$27.9158.98$\pm$27.3359.92$\pm$27.5361.11$\pm$27.00
COCO-351.45$\pm$29.3553.62$\pm$29.0357.92$\pm$29.3957.56$\pm$29.4459.41$\pm$28.9659.80$\pm$29.1260.20$\pm$28.71
PASCAL-Part-041.58$\pm$25.2749.63$\pm$24.3043.36$\pm$27.9148.55$\pm$26.5250.92$\pm$25.7349.89$\pm$26.4451.44$\pm$25.81
PASCAL-Part-160.54$\pm$24.7863.08$\pm$25.7963.71$\pm$26.6266.54$\pm$25.7765.04$\pm$26.3965.82$\pm$26.2464.75$\pm$26.67
PASCAL-Part-241.31$\pm$24.2547.77$\pm$23.5046.82$\pm$27.0249.36$\pm$25.8650.62$\pm$25.2549.82$\pm$25.8050.66$\pm$25.28
PASCAL-Part-349.61$\pm$25.9056.03$\pm$25.5951.65$\pm$26.6055.83$\pm$26.2756.93$\pm$25.6155.96$\pm$25.7256.58$\pm$25.19
PACO-Part-032.80$\pm$25.5037.23$\pm$27.2332.99$\pm$28.1736.40$\pm$28.5338.13$\pm$28.7437.25$\pm$28.8338.55$\pm$28.97
PACO-Part-136.63$\pm$25.4343.01$\pm$25.5337.91$\pm$28.2941.93$\pm$27.4844.12$\pm$27.4143.17$\pm$28.1144.93$\pm$27.79
PACO-Part-239.95$\pm$25.3344.99$\pm$26.6739.53$\pm$27.9043.45$\pm$27.8645.26$\pm$27.8744.60$\pm$28.0346.05$\pm$28.01
PACO-Part-334.99$\pm$25.9138.98$\pm$26.6935.42$\pm$28.3037.42$\pm$28.4339.28$\pm$28.5638.62$\pm$29.1040.13$\pm$29.22
LVIS-031.01$\pm$31.4233.99$\pm$31.3235.25$\pm$33.2435.84$\pm$33.2038.78$\pm$33.4738.18$\pm$33.7440.06$\pm$33.96
LVIS-131.98$\pm$31.5936.95$\pm$31.1836.20$\pm$33.4038.48$\pm$32.8741.39$\pm$32.9940.21$\pm$33.5142.23$\pm$33.39
LVIS-232.59$\pm$31.6737.01$\pm$31.6037.51$\pm$33.7138.65$\pm$33.4342.14$\pm$33.3840.81$\pm$33.6143.54$\pm$33.42
LVIS-334.26$\pm$31.5739.40$\pm$30.4539.53$\pm$32.5741.29$\pm$31.6744.93$\pm$31.7943.88$\pm$32.5046.05$\pm$32.27
LVIS-429.00$\pm$31.4933.12$\pm$31.2533.55$\pm$33.7034.84$\pm$33.1637.92$\pm$33.6037.26$\pm$34.0239.35$\pm$34.01
LVIS-534.38$\pm$32.0639.24$\pm$31.7741.28$\pm$34.1342.37$\pm$33.3245.15$\pm$33.1744.75$\pm$33.7046.61$\pm$33.63
LVIS-637.28$\pm$32.4739.81$\pm$31.4142.22$\pm$34.5642.04$\pm$33.5244.67$\pm$33.5644.22$\pm$34.2446.12$\pm$33.97
LVIS-733.97$\pm$31.7838.34$\pm$31.0438.81$\pm$33.7240.49$\pm$32.9643.56$\pm$33.1242.05$\pm$33.6044.20$\pm$33.47
LVIS-831.72$\pm$32.2436.45$\pm$32.0936.13$\pm$34.1638.18$\pm$33.6141.42$\pm$33.8440.51$\pm$34.2243.01$\pm$34.15
LVIS-930.12$\pm$30.7133.71$\pm$31.0834.24$\pm$33.2034.82$\pm$33.0338.28$\pm$33.4337.13$\pm$33.6939.44$\pm$33.77

Q2 To examine whether the selected hyperparameters are robust across different domains, we conduct controlled one-at-a-time experiments on COCO (four-fold mean), ISIC, and Lung, keeping the remaining settings and random seed fixed.

(1) Log-sum-exp temperature β\beta.

We evaluate β∈{0.01,0.07,0.20,0.50}\beta\in\{0.01,0.07,0.20,0.50\}. The results show that changing β\beta over nearly two orders of magnitude leads to negligible performance variations: COCO varies by at most 0.17 points, while ISIC and Lung remain within 0.20 points.

In particular, the default β=0.07\beta=0.07 achieves 60.90%, 62.90%, and 87.60% mIoU on COCO, ISIC, and Lung, respectively.

These results indicate that the proposed multi-prototype aggregation is insensitive to the precise temperature setting, and that β=0.07\beta=0.07 is a reasonable operating point rather than a narrowly optimized value.

(2) Hard-negative retention ratio.

We further evaluate the hard-negative retention ratio r∈{0.05,0.20,0.50,1.00}r\in\{0.05,0.20,0.50,1.00\}. An excessively small ratio (r=0.05r=0.05) under-samples confusing background regions, reducing mIoU by 1.98, 0.70, and 0.80 points on COCO, ISIC, and Lung, respectively.

Conversely, retaining the entire background (r=1.00r=1.00) introduces excessive easy-background information and substantially degrades Lung performance by 3.30 points.

Although r=0.50r=0.50 slightly outperforms the default on COCO and ISIC, it produces a marginal decrease on Lung. No single alternative consistently dominates across all three datasets. We therefore retain r=0.20r=0.20 as a balanced cross-domain setting that avoids both insufficient hard-negative sampling and excessive inclusion of easy background regions.

Table: Controlled sweeps of Foreground Purification hyperparameters (mIoU).

HyperparameterValueCOCO ↑\uparrowISIC ↑\uparrowLung ↑\uparrow
β\beta (LSE temperature)0.0160.8062.9087.60
0.0760.9062.9087.60
0.2060.9763.0087.50
0.5060.8563.1087.50
rr (hard-neg. ratio)0.0558.9262.2086.80
0.2060.9062.9087.60
0.5061.2562.3087.40
1.0060.8863.0084.30

(3) ϵ\epsilon.

ϵ\epsilon is a small constant introduced in min-max normalization to prevent division by zero. We have clarified its definition in the revised manuscript and explicitly specified its value as 1×10−61\times10^{-6}.

Q3: Overstated Performance Claims on Semantic Benchmarks

Thank you for pointing out this imprecise statement. We acknowledge that FoRIS does not achieve the best performance on every semantic benchmark. In particular, SegIC outperforms FoRIS on VIS-92i (44.6 vs. 42.8) and COCO-20i (76.1 vs. 60.9). We note that SegIC is trained on the LVIS and COCO datasets, which may provide it with an advantage on these benchmarks. To avoid an overly broad claim, we have revised the statement from "FoRIS achieves the best performance on all semantic benchmarks" to "FoRIS achieves the best performance on most semantic benchmarks."

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

This paper introduces a training-free method for in-context image segmentation. The method consists of three main steps: foreground purification, foreground localization, and foreground consolidation. The experiments were conducted on several benchmarks across different domains.

Critical strengths
  1. The authors conducted extensive experiments on different benchmarks.
  2. The appendix provides the missing information in the main text and includes a discussion on the limitations.
Critical weaknesses

There are several main concerns from my side. One is about the effectivenss of the claimed three steps. In addition, the proposed modules and steps are mostly built on the off-the-shelf methods, making the novelty limited. More importantly, the comparison to more recent related works is missing. Here follow the details.

  1. Figure 2 shows the major changes for the activation come from the last step, consolidation. However, the performance gains are hugely contributed from the first and second steps as shown from Figure 1.
  2. From my side, I do not agree with the incorrect similar white square-shaped wall. The reference is white square, as long with a refrigerator. The while square-shaped wall region also fits the request. It could be treated as an ambiguous problem for the reference, but not an incorrect prediction.
  3. The three coupled sources of uncertainty in Line 94 are not well clarified in the introduction, leading to unclear understanding of the motivation.
  4. The literature review of in-context segmentation is incomplete. The authors should include more recent works and incorporate discussions as well.
  5. The logical flows in progressive segmentation and mask refinement section is not smooth. One may be confused about the transition, which makes the readers difficult to follow. In addition, one may not know the exact contribution of this work compared to prior works.
  6. In the first step if FP, the proposed work directly follows the previous work INSID3. In the context of training-free learning, what is the key contribution added into this method section?
  7. The variables are quite mixed, e.g., u_j, u^fg. In addition, how did the authors determine the clusters? for instance, the threshold?
  8. I might not get the motivation of semantic matching. What is the step aiming for? And why is the debiased semantic geometry preserved in this manner?
  9. The candidate voting and prior scheme are not new in the domain of image segmentation. What is the difference from prior works?
  10. The experimental section does not include sufficient comparison to recent works for in-context image segmentation. While there are quite a few works in AAAI, ICLR, CVPR, ICML, ECCV venues, most compared methods were published before 2025, which is complete for a submission to ICLR 2027. Here I just listed some representative ones: PR-MaGIC: Prompt refinement via mask decoder gradient flow for in-context segmentation; REBASE: Reference-Background Subspace Elimination for Training-Free In-Context Segmentation; Foveate: Recursive Zoom for Training-Free, In-Context Instance Segmentation; Training-Free Open-Vocabulary Semantic Segmentation with Context Pyramid Refinement; Toward Robust In-Context Segmentation via Concept Guidance.
Questions for authors

See the weaknesses above.

9 comments

We sincerely thank you for your valuable comments. Our responses to the concerns raised are provided below:

Q1: We thank the reviewer for this insightful observation. We believe that the two figures evaluate different aspects of our method, and therefore, there is no contradiction between them. We have added more visualization examples to Figure~2 to highlight the specific challenges addressed by each stage of the progressive refinement process.

Q2: We agree with the reviewer that this example is more appropriately characterized as semantic ambiguity rather than an explicitly incorrect prediction. This is also one of the motivations of our work, highlighting a limitation of direct matching methods: when multiple regions are visually compatible with the reference target, a one-step correspondence map may struggle to distinguish the semantic concept that the user actually intends to query.

Specifically, the reference target (a white rectangular appliance/refrigerator) and the white wall region share similar low-level visual appearances. Consequently, when relying solely on visual similarity, both regions may become plausible matching candidates.

Q3:

The three sources of uncertainty correspond to the three fundamental challenges that motivate the design of FP, FL, and FC, respectively.

(1) Reference Feature Contamination / Background Clutter

Even when a reference image is provided with a manually annotated foreground mask, its features in the DINO space may still contain positional artifacts and tokens from neighboring background regions (Fig.~2(a)). Such impure reference features can contaminate foreground representations and subsequently lead to biased matching.

This issue is addressed by FP, which incorporates Adaptive Positional Debiasing and Two-stage Background Suppression.

(2) Ambiguous Correspondence / Localization Uncertainty

As illustrated in Fig.~1, under cosine-similarity matching, multiple query regions may simultaneously exhibit high similarity to the reference foreground, such as repeated object parts or visually similar regions. Therefore, local feature similarity alone is insufficient to accurately determine the actual spatial location of the target.

This issue is addressed by FL, which incorporates Backward Candidate Voting and Seed-cluster Prior. These mechanisms improve localization reliability by exploiting bidirectional correspondences and spatial coherence.

(3) Foreground--Background Semantic Ambiguity / Semantic Discrimination Uncertainty

As illustrated in Fig.~2(d), even after the target region has been preliminarily localized, its local features may remain semantically ambiguous. Some features may exhibit similarities to both foreground and background, making their semantic assignments uncertain. Moreover, semantic responses across different foreground feature clusters may be inconsistent, resulting in unclear foreground--background decision boundaries. Such ambiguity can further lead to foreground misclassification and incomplete predictions.

This issue is addressed by FC, which incorporates Semantic Disagreement Penalty and Cluster-level Reweighting. By explicitly modeling foreground--background semantic disagreement and adaptively adjusting the contributions of different feature clusters, FC enhances semantic separability between foreground and background.

Q4: We agree with the reviewer and expand the Related Work section to cover recent advances in ICS and training-free segmentation, including, but not limited to, the representative methods listed in Q10: PR-MaGIC, REBASE, Foveate, Context Pyramid Refinement, and Concept-Guided ICS (CG-ICS).

Meanwhile, we explicitly clarify the differences between FoRIS and these approaches. FoRIS relies exclusively on frozen DINOv3 features, without requiring a mask-supervised decoder or an MLLM, and progressively refines foreground representations through FP, FL, and FC.

Q5:

We have substantially revised the ``Progressive Segmentation and Mask Refinement'' section to improve its logical flow and clarify the methodological contribution of FoRIS.

Specifically, we reorganize the discussion into a more coherent progression. We first introduce the objective of training-free ICS and its central challenge of transforming pretrained visual representations into reliable foreground-background separation. We then categorize existing model-assisted approaches into two typical pipelines and highlight a key limitation of correspondence-based methods: they primarily rely on feature matching and similarity propagation, with limited systematic refinement of intermediate foreground evidence within the visual feature space.

To establish the motivation for our approach, we further introduce the progressive refinement principle in classical segmentation, illustrating how different methods iteratively resolve uncertainty in intermediate visual evidence to obtain coherent foreground structures. Building on this principle, we formulate training-free ICS as a progressive foreground refinement process and explicitly raise the question of how to achieve such refinement entirely within the feature space without updating model parameters or introducing auxiliary segmentation models.

More importantly, we clarify the distinction and contribution of FoRIS compared with prior correspondence-based approaches. Rather than directly projecting initial correspondences into segmentation predictions, FoRIS treats them as intermediate semantic evidence and progressively refines foreground representations through three stages: Foreground Purification (FP), Foreground Localization (FL), and Foreground Consolidation (FC). All three stages are integrated within the feature space of a single frozen DINOv3 encoder, without relying on external segmentation models, multimodal large language models, or auxiliary feature encoders. This establishes a unified progressive feature-space refinement paradigm that goes beyond conventional correspondence-based prediction.

Q6: FP is not a straightforward reproduction of INSID3. INSID3 provides a strong DINOv3-based matching baseline, whereas FoRIS reuses the same frozen encoder but introduces a different feature purification strategy and a different mask generation process.

Specifically, our contributions are twofold:

1. Adaptive Conditional Positional Debiasing (APD).

Conventional or INSID3-style approaches typically remove positional subspaces unconditionally. In contrast, we introduce a semantic-alignment-aware gating mechanism based on ssems_{\mathrm{sem}}, applying positional debiasing only when ssem<0.8s_{\mathrm{sem}} < 0.8 in our implementation.

This design preserves useful geometric information when the reference and target images already exhibit strong semantic alignment.

Our ablation results show that adaptive debiasing improves performance over fixed debiasing on X-Ray, increasing the result from 84.9 to 87.6. On SPair-71k, adaptive debiasing also achieves PCK performance comparable to or better than fixed debiasing.

2. Background Suppression.

We introduce a multi-prototype Log-Sum-Exp foreground matching mechanism to score target tokens, together with an orthogonalized hard-negative background penalty. Specifically, we select the top 20% background tokens according to their similarity to the foreground prototype μfg\boldsymbol{\mu}_{\mathrm{fg}}, and orthogonalize them using the Gram--Schmidt process.

Q7: We unify the notation throughout the revised manuscript. The specific definitions used in our manuscript and implementation are as follows.

1. Notation

  • μfg\boldsymbol{\mu}_{\mathrm{fg}} and μbg\boldsymbol{\mu}_{\mathrm{bg}}: global foreground and background prototypes, respectively.
  • {μfg(k)}k=1Kfg\{\boldsymbol{\mu}_{\mathrm{fg}}^{(k)}\}_{k=1}^{K_{\mathrm{fg}}}: clustered reference foreground prototypes.
  • {πk}\{\boldsymbol{\pi}_k\}: target-side cluster prototypes used in the Foreground Localization (FL) stage.

For patch-level features, we use f\mathbf{f} to denote the original normalized features and f~\tilde{\mathbf{f}} to denote the features after APD.

In particular, we replace the previously mixed notations, such as μj\boldsymbol{\mu}_j and μfg\boldsymbol{\mu}^{\mathrm{fg}}, with the consistent subscript-based notation above. The only scalar function u(⋅)u(\cdot) used in the manuscript denotes the uncertainty weight in the Foreground Consolidation (FC) stage.

2. Shared Agglomerative Clustering

Whenever FoRIS partitions a set of tokens, we denote their corresponding L2-normalized descriptors as {zi}\{\mathbf{z}_i\}, with their specific definitions provided separately for each stage.

We perform average-linkage agglomerative clustering based on cosine similarity:

Dij=⟨zi,zj⟩.D_{ij} = \langle \mathbf{z}_i, \mathbf{z}_j \rangle.

We adopt a fixed threshold:

τ=0.6.\tau = 0.6.

Cluster merging terminates when no pair of clusters has an average similarity exceeding τ\tau. Therefore, the final number of clusters is adaptively determined by the input data. We provide a complete ablation study of τ\tau in Appendix L.

Q8:

As illustrated in Fig.2(a), the similarity maps between reference and target images may contain positional bias introduced by absolute spatial correlations. Although removing such bias helps suppress irrelevant positional interference, positional information is not always harmful. For example, the X-Ray dataset exhibits relatively strong spatial-layout consistency, where positional relationships can provide valuable spatial priors. The ablation results in Table~5 further demonstrate that preserving positional information is more beneficial for X-Ray, whereas removing positional bias is more beneficial for COCO.

To address this issue, we introduce Semantic Matching to evaluate the reliability of semantic correspondence between the reference foreground and the target image before applying positional debiasing:

s_{\mathrm{sem}} &= \frac{ \langle \boldsymbol{\mu}_{\mathrm{fg}}, \bar{\mathbf{f}}_t \rangle }{ \|\bar{\mathbf{f}}_t\|_2 } \end{aligned}$$ Specifically, our design has two aspects: - **Adaptive Debiasing.** When semantic matching is strong, positional information may provide useful spatial priors. Otherwise, positional debiasing helps suppress potential interference. For example, on X-Ray, removing positional bias reduces performance from 87.6 to below 85, indicating that unconditional debiasing may discard useful spatial information. - **Conditional Preservation of Spatial Structure.** When $$s_{\mathrm{sem}} \geq \tau_{\mathrm{sem}}, \qquad \tau_{\mathrm{sem}} = 0.8,$$ we preserve the original features; otherwise, we apply the positional debiasing projection matrix $\mathbf{P}_{\perp}$. Therefore, this mechanism adaptively determines whether positional debiasing is necessary, rather than always removing or always preserving positional information. Furthermore, experiments on SPair-71k show that adaptive debiasing achieves performance comparable to or better than fixed debiasing, further validating its effectiveness. Overall, Semantic Matching evaluates the reliability of semantic correspondence and adaptively controls positional debiasing, balancing the suppression of positional bias with the preservation of useful spatial structure.

Comments on the paper

5 comments
~mingjie_yang1OpenReview ↗

Questions on image provenance and implementation details

Dear authors,

While reading the current submission and comparing it with the anonymous code linked in the paper, we noticed the following points and would appreciate your clarification:

  1. Figure 9 on page 24 is labeled DeepGlobe-18, but its target image matches 11878675_15.tiff from the official Massachusetts Roads training split. We would like to know which dataset was used for the road results in Tables 11 and 13, and whether those results and the visualization use the same dataset.

  2. The paper's equations and the code's default weights differ: the background coefficient in Eq. 7 is 1, whereas the code uses 0.55; the two added terms in Eq. 13 both have coefficient 1, whereas the code uses 0.20 and 0.25. See models/foris.py, lines 33-37, 514, 739 and 753. We would like to understand which weights were used in the reported experiments, and whether these differences reflect a simplified presentation of the equations.

  3. Reference and query features are normalized along different axes: the reference tensor has shape [1,C,H,W], but line 582 uses dim=2, normalizing along height; the query features use dim=1 at line 578, normalizing along channels. We would like to confirm whether this is intentional and how it corresponds to Eq. 8.

  4. The multi-reference APD decision uses only the score of the last valid reference. _should_apply_positional_debias overwrites s_sem_last in the loop, and lines 336-338 directly compare that value with 0.8. We would like to understand whether the 5-shot experiments also use this rule and whether a fixed ordering of reference images is specified.

Thank you for clarifying these details.

Thank you very much for your interest in our work and for taking the time to carefully examine our paper and released implementation. We sincerely appreciate your thoughtful questions. Please find our responses below.

Q1. Dataset used for the road segmentation experiments.

The DeepGlobe-18 dataset contains 1,107 images in its training split, whereas its test split contains only 13 images. Given the extremely limited number of test images, we believe that the official test split is insufficient for a statistically meaningful evaluation. Therefore, we directly evaluated our method on all 1,107 images in the training split.

We would also like to emphasize that FoRIS is a training-free method and does not require any training or fine-tuning on the target domain. Consequently, using the training split for evaluation does not involve any target-domain training or parameter optimization. From the perspective of our training-free cross-domain evaluation protocol, the training and test splits can both serve as target-domain evaluation sets.

Q2. Differences between the coefficients in the equations and the implementation.

Due to the page limitations of the ICLR main paper, we simplified the mathematical formulations of certain components based on the actual implementation in our released code, aiming to make the overall method easier to follow. The released code provides the complete experimental configuration, including the default coefficients actually used in our experiments.

To further clarify the roles and effects of these hyperparameters, we have also conducted corresponding sensitivity analyses, which are provided in Appendix K.

Q3. Different normalization axes for reference and query features.

Thank you for pointing out this inconsistency. To investigate its impact, we evaluated both normalization settings (dim=2 and dim=1) for the reference features. The results are highly comparable:

NormalizationCOCOISICX-ray
dim=261.062.787.6
dim=160.962.987.6

These results indicate that the choice of normalization axis has a negligible impact on performance. Nevertheless, we acknowledge that the implementation and mathematical description should be consistent. We have therefore revised the description of Eq. 8 to clarify the normalization operation and ensure that its mathematical formulation accurately reflects the intended implementation.

Q4. Multi-reference APD decision in the 5-shot setting.

Our primary experiments focus on the more challenging 1-shot setting, where adaptive positional debiasing is particularly important. The current implementation follows the design described in our method, and the 5-shot experiments directly adopt the same decision rule from the 1-shot setting without additional optimization. Specifically, the APD decision is based on the semantic similarity score of the last valid reference image in the reference loop.

We acknowledge that more effective APD strategies may be possible for the 5-shot setting, such as aggregating semantic similarity scores across multiple references, and leave this as future work. We also clarify that all reported results strictly follow the released implementation, ensuring consistency between the experimental results and the code.

Final clarification.

Finally, we would like to emphasize that the above questions concern certain implementation details, hyperparameter specifications, and presentation inconsistencies. They do not affect the central motivation of our proposed progressive framework, nor do they change the reported experimental results or the main conclusions of the paper. We appreciate your careful examination, which has helped us identify several aspects that can be further clarified to improve the transparency and reproducibility of our work.

Thank you again for your valuable feedback and interest in our research.

AuthorsOpenReview ↗

To All Reviewers and Scholars Interested in Our Work,

To All Reviewers and Scholars Interested in Our Work,

We sincerely thank all reviewers for their constructive and insightful comments. In response to the reviewers' concerns and suggestions, we have carefully revised the manuscript and conducted additional experiments to improve its technical clarity, methodological rigor, and experimental completeness. In the revised PDF, blue indicates modified content, while red indicates newly added content. Our major revisions and corresponding responses are summarized as follows:

  1. Clarification of the motivation behind the progressive framework.

    To further address the reviewers' concerns regarding the motivation and necessity of the progressive framework, we have revised Fig. 2 to explicitly clarify the roles of the three components: FP serves as a feature denoising stage, FL localizes the target based on the denoised features, and FC further resolves semantic ambiguity after target localization. These revisions highlight the complementary and progressive nature of the three stages. Furthermore, we have added extensive ablation studies in Appendix J to validate the individual contributions of each component and the progressive performance gains achieved by their combination.

  2. Expanded discussion of related ICS methods.

    In response to the suggestions regarding related work, we have expanded the Related Work section with additional comparisons and discussions of existing ICS methods. In particular, we further clarify that existing ICS approaches predominantly rely on multi-model collaborative segmentation paradigms, while relatively little attention has been paid to the classical coarse-to-fine segmentation paradigm and the potential of progressively refining features within a single model.

  3. Completion of methodological details and clarification of notation.

    To address the concerns regarding the completeness and clarity of the proposed method, we have added the previously omitted clustering procedure to the Method section and conducted additional ablation experiments on the clustering threshold in Appendix L. We have also carefully revised the definitions and usage of mathematical symbols throughout the manuscript to avoid ambiguity and potential confusion arising from similar notations.

  4. Revision of the response-map formulations and coefficient analysis.

    Following the reviewers' suggestions, we have revised the response-map formulations by introducing explicit weighting coefficients:

S(1)(p)=βlog⁡∑k=1Kfgexp⁡(⟨f~t(p),μfg(k)⟩β)−λbg⟨f~t(p),μbg⊥⟩.S^{(1)}(\mathbf{p}) = \beta \log \sum_{k=1}^{K_{\mathrm{fg}}} \exp\Big( \frac{ \langle \tilde{\mathbf{f}}_t(\mathbf{p}), \boldsymbol{\mu}_{\mathrm{fg}}^{(k)} \rangle }{\beta} \Big) - \lambda_{\mathrm{bg}} \langle \tilde{\mathbf{f}}_t(\mathbf{p}), \boldsymbol{\mu}_{\mathrm{bg}}^{\perp} \rangle.

S(2)=S(1)+λV(V−12)+λP(P−12),S^{(2)} = S^{(1)} + \lambda_V\left(V-\frac{1}{2}\right) + \lambda_P\left(P-\frac{1}{2}\right),

S(3)=S(2)−λΠΠ+λΔΔ.S^{(3)} = S^{(2)} - \lambda_{\Pi}\Pi + \lambda_{\Delta}\Delta.

In addition, we have conducted ablation studies on these coefficients in Appendix J to investigate their effects on the final performance.

  1. Additional comparisons and correction of performance claims.

    To provide a more comprehensive experimental evaluation, we have added comparisons with Foveate and PR-MaGIC (PerSAM) in the main experimental section. We have also revised the original statement, "FoRIS achieves the best performance on all semantic benchmarks," to the more accurate statement, "FoRIS achieves the best performance on most semantic benchmarks," to ensure that our conclusions are fully consistent with the experimental results.

  2. Additional ablation studies on β\beta and the hard-negative ratio.

    To further investigate the effects of important hyperparameters in FoRIS, we have added ablation experiments on β\beta and the hard-negative ratio in Appendix K.

  3. Comprehensive comparison between FoRIS and INSID3.

    To better clarify the methodological differences and innovations of FoRIS relative to INSID3, we have added a comprehensive comparison in Appendix I. This additional analysis highlights the key differences in their feature processing strategies and final mask generation mechanisms, thereby providing a clearer understanding of the contributions introduced by FoRIS.

We sincerely appreciate the reviewers' valuable comments and suggestions, which have helped us substantially improve the quality of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.