acceptodds
Under review as a conference paper at ICLR 2027

UniRefDet: Universal Referring Detection of Salient and Camouflaged Objects in Unconstrained Scenes

Abstract

Salient object detection (SOD) and camouflaged object detection (COD) with language-driven understanding have attracted substantial attention in recent years. Existing studies predominantly focus on referential foreground object localization under individual task settings, while language-guided distinction between salient and camouflaged objects, particularly when these two visual attributes coexist in the same scene, remains underexplored. In this paper, we construct UniRefDet, a novel referring segmentation benchmark that unifies salient and camouflaged object detection tasks in unconstrained real-world scenes. The UniRefDet presents fine-grained referring expressions that explicitly describe object-environment relations. Built on this benchmark, we propose COREFNET, a SAM-based multimodal segmentation framework. We first propose a Language-Aware Semantic Adaptation (LASA), which inserts lightweight Text Adapters into the text encoder to adapt concept-level representations to fine-grained referring expressions. An Attribute Discrimination Module (ADM) is further designed, aiming to calibrate multiple proposals within each attribute and suppresses conflicting salient and camouflaged responses in coexistence scenes. Extensive experiments show that COREFNET achieves state-of-the-art performance across all scene configurations on UNIREFDET.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Referring image segmentation, salient object detection, camouflaged object detection, datasets and benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tyWFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces UniRefDet, a unified referring segmentation benchmark for salient and camouflaged objects in unconstrained scenes, and CoRefNet, a SAM 3-based framework tailored to fine-grained referring expressions. Experiments demonstrate that the proposed method achieves strong performance.

Critical strengths
  1. The proposed benchmark is practical and could stimulate further research in unconstrained scenes.
  2. The proposed framework is well-motivated and achieves competitive performance.
Critical weaknesses
  1. The primary contribution of this paper is the construction of the UniRefDet benchmark. However, the images are drawn from an existing benchmark, and the generation of referring annotations relies heavily on LMMs. I therefore consider the contribution to be incremental.
  2. When constructing the UniRefDet benchmark, how did the authors determine whether an object was camouflaged or salient in scenes where both types of objects coexist? Camouflage and saliency are continuous rather than discrete properties, meaning that an object may exhibit both to varying degrees (e.g., a camouflage score of 0.55 and a saliency score of 0.45). If the dataset simply assigns binary labels, could this lead to confusion for the model?
  3. The core objective of referring image segmentation is to segment the target object specified by a referring expression, with the output containing only one class. In contrast, camouflaged and salient object detection in unconstrained scenes aims to detect both classes simultaneously, producing outputs that contain two classes. I am therefore quite puzzled by the rationale for unifying these two tasks. In practice, if I want to segment a target using a referring expression, whether the target is salient or camouflaged is either irrelevant to my goal or already reflected in the expression. Alternatively, if I want to detect camouflaged and salient objects simultaneously, I would expect the model to make predictions directly, without requiring me to provide complex referring expressions.
Questions for authors

Please refer to Weaknesses.

2 comments

We sincerely thank you for the valuable and constructive feedback. Below, we address your concerns point by point. We would be grateful if you could let us know whether our responses adequately address your questions and concerns.

Q1. Is UniRefDet merely an incremental extension of an existing benchmark?

We acknowledge that UniRefDet reuses images from USC12K. However, our contribution goes beyond simply adding captions to existing images. UniRefDet establishes a fine-grained language-guided benchmark for attribute-aware referring segmentation, with expressions covering eight dimensions, including object appearance and object–environment relationships. LMMs are used only to generate initial descriptions, which may contain factual errors, hallucinations, or ambiguities regarding object categories, shapes, quantities, and visual attributes. To ensure annotation quality, an independent VLM first checks image–mask–expression consistency, followed by two rounds of manual review and refinement conducted by three trained annotators under standardized guidelines. This process involves correcting inaccuracies and removing ambiguous or unsupported descriptions, requiring approximately 2–4 minutes per image. Thus, UniRefDet contributes carefully verified fine-grained referring annotations and a unified evaluation protocol for target localization and visual-attribute assignment, while CoRefNet provides an effective solution to this new task.

Q2. Why are saliency and camouflage represented by binary labels?

We agree that saliency and camouflage are context-dependent perceptual properties that may vary continuously. However, existing SOD/COD benchmarks typically use binary segmentation masks for standardized training and evaluation. UniRefDet follows USC12K's three-class segmentation setting (background, salient, and camouflaged), directly inheriting its annotations rather than introducing new thresholds for attribute classification. These categories serve as operational labels for evaluation rather than quantitative measurements of continuous visual attributes.

In Scenes A and B, salient and camouflaged objects are relatively distinguishable. However, as you correctly point out, Scene C presents a more challenging situation where both attributes coexist, and relying solely on binary labels may lead to attribute confusion. This motivates our fine-grained referring annotations. Beyond object categories, our expressions describe target appearance and object–environment relationships, including how a target visually contrasts with or blends into its surroundings. These descriptions provide complementary contextual cues for distinguishing targets with different visual attributes, rather than relying solely on categorical labels.

We acknowledge that such annotations cannot fully eliminate the inherent ambiguity of continuous perceptual properties, particularly in borderline cases. Nevertheless, they provide richer supervision for studying attribute-aware referring segmentation, especially in coexistence scenes. We will clarify this motivation and the limitations of categorical annotations in the revised manuscript.

Q3. Why is it necessary to unify referring segmentation with SOD/COD?

We appreciate this important question. Our key motivation is that objects of the same semantic category may exhibit different visual attributes within the same scene. For example, two animals of the same category may coexist, with one visually salient and the other blending into its surroundings. A user may wish to locate each animal through a different referring expression. However, applying SOD and COD independently may produce overlapping predictions or ambiguous attribute assignments, particularly when objects exhibit competing visual characteristics. UniRefDet provides a unified framework for investigating how fine-grained language descriptions can help distinguish such targets and alleviate attribute ambiguity.

This setting also has potential practical applications. For instance, in wildlife monitoring, researchers may need to distinguish animals clearly visible in open areas from those concealed by surrounding vegetation, rather than detecting all animals indiscriminately. Similarly, in UAV-assisted search and rescue, operators may use language to specify particular individuals who are either clearly visible or partially concealed by debris or vegetation. The value of UniRefDet lies in providing a complementary benchmark for studying language-guided target disambiguation under competing visual attributes, rather than simply combining existing tasks. We will further clarify this motivation and its distinction from conventional referring segmentation and SOD/COD in the revised manuscript.

I appreciate the authors’ responses, which have addressed some of my concerns. However, my main concern regarding the motivation for unifying reasoning segmentation with COD/SOD remains unresolved. I am therefore inclined to maintain my current score. Thank you!

Reviewer_vCsPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces UniRefDet, a language-guided segmentation benchmark for jointly handling salient and camouflaged objects in unconstrained scenes. Built upon USC12K, the benchmark provides fine-grained referring expressions across salient-only, camouflaged-only, coexistence, and background-only scenes. The authors also propose CoRefNet, a SAM 3-based framework incorporating Language-Aware Semantic Adaptation (LASA) and an Attribute Discrimination Module (ADM) for proposal aggregation and cross-attribute suppression. Experiments demonstrate strong performance on UniRefDet, particularly in coexistence scenes, along with promising cross-dataset transfer results.

Critical strengths
  1. A potentially useful benchmark for language-guided salient and camouflaged object segmentation. UniRefDet extends existing visual annotations with fine-grained referring expressions and jointly evaluates salient and camouflaged objects, including their coexistence. This provides a useful setting for investigating language-conditioned segmentation under competing visual attributes.
  2. Strong empirical performance with meaningful diagnostic analyses. CoRefNet achieves 82.53% overall mIoU and 82.54% overall oIoU. In particular, its Scene-C camouflaged-object mIoU improves from the strongest baseline's 60.48% to 65.97%. The ablations and linguistic perturbation experiments in Tables 3–7 further provide evidence that language adaptation, proposal calibration, and contextual descriptions contribute to performance.
Critical weaknesses
  1. The distinct learning challenge introduced by UniRefDet is not sufficiently established. USC12K already addresses unified salient and camouflaged object segmentation across four scene configurations, while existing referring segmentation methods, such as GRES, handle relational grounding, multiple targets, and absent referents. Three concerns remain: (1) Unclear task novelty: It is unclear what new learning challenges arise from combining these established settings. (2) Attribute information provided in the input: Although the model is expected to distinguish salient from camouflaged objects, the referring expressions and their attribute-specific assignments already convey this information, calling into question the need for additional visual attribute reasoning. (3) Simplified no-target setting: Background-only scenes use the explicit prompt "background" rather than a referring expression with no matching target, limiting the evaluation of genuine no-target rejection. Consequently, the benchmark primarily evaluates attribute-conditioned referring segmentation, rather than demonstrating a distinct challenge in visual attribute reasoning or no-target detection. This weakens its claimed novelty and significance.

  2. The independent methodological contribution of CoRefNet remains insufficiently demonstrated. (1) LASA relies on conventional lightweight text adapters, while ADM employs proposal-score calibration, weighted aggregation, and cross-attribute suppression. Closely related work such as SAM3-I (ACL 2026) has already investigated adapting SAM 3 to complex natural-language instructions. Although the paper includes useful component ablations and comparisons with established methods, it lacks a sufficiently controlled comparison against simpler SAM 3-based adaptation and dual-prompt fusion strategies. (2) Furthermore, Table 4 shows that ADM's improvement in Scene-C camouflaged-object oIoU is not accompanied by an improvement in mIoU, leaving its claimed attribute-discrimination mechanism incompletely characterized. These gaps make it difficult to determine how much of the reported gain comes from the proposed mechanisms rather than the pretrained backbone, fine-grained supervision, and conventional adaptation.

Questions for authors
  1. Task novelty and attribute information: Is UniRefDet intended to evaluate attribute-conditioned referring segmentation or the model's ability to infer salient/camouflaged attributes from visual context? Can the authors provide evidence of attribute discrimination without explicitly supplying the attribute identity through the query assignment or expression?
  2. No-target evaluation: Since Scene D uses the fixed expression "background," how does the method handle ordinary referring queries whose targets are absent? Could the authors clarify whether its reported background-only performance measures genuine no-target rejection?
  3. Methodological contribution: Can the authors provide controlled comparisons using the same SAM 3 backbone and training protocol, particularly against standard text adaptation and independent salient/camouflaged prompt segmentation followed by simple score fusion? A direct attribute-confusion analysis would also help establish the specific benefit of ADM.
2 comments

We sincerely thank you for the valuable and constructive feedback. Below, we address your concerns point by point. We would be grateful if you could let us know whether our responses adequately address your questions and concerns.

Q1. Task Novelty and Visual Attribute Discrimination

USC12K studies unified salient and camouflaged object segmentation, while GRES addresses relational grounding, multiple targets, and absent referents. UniRefDet focuses on jointly achieving language-guided instance selection and consistent visual-attribute assignment when salient and camouflaged targets coexist. For example, an image may contain two animals of the same category, one visually salient and the other blending into its surroundings. A user may wish to locate the camouflaged animal through one expression and the salient animal through another. However, independently applying SOD and COD may produce overlapping predictions or ambiguous attribute assignments. UniRefDet provides a unified setting for investigating how fine-grained language descriptions can distinguish these targets and alleviate attribute ambiguity.

This setting also has potential practical value. In wildlife monitoring, researchers may distinguish visible animals from those concealed in vegetation, rather than detect animals. Similarly, in UAV-assisted search and rescue, operators may use language to specify particular individuals who are clearly visible or partially concealed by debris or vegetation. UniRefDet therefore provides a complementary benchmark for studying language-guided target disambiguation under competing visual attributes, rather than simply combining existing tasks. USC12K does not provide fine-grained referring expressions for instance selection, while GRES does not specifically evaluate competition between salient and camouflaged attributes. UniRefDet jointly evaluates these aspects within a unified setting. We will clarify this motivation in the revised manuscript.

The primary task is attribute-conditioned referring segmentation. To further examine visual attribute discrimination beyond explicit attribute cues, we conducted end-to-end retraining and evaluation after removing Attribute IDs from Scenes A/B/D. We then additionally removed explicit attribute words from the referring expressions to examine whether the model could use the remaining appearance and contextual descriptions for segmentation.

Table R1. Performance with reduced explicit attribute information (mIoU/oIoU, %).

SettingA: SalientB: CamouflagedC: SalientC: Camouflaged
Without Attribute IDs87.82/88.6883.15/86.3880.67/80.3573.83/76.59
Further removing explicit attribute words87.82/88.5881.51/84.0280.65/80.3773.83/76.44

The model retains strong performance after removing explicit attribute identifiers and attribute words. In particular, Scene-C camouflaged-object segmentation still achieves 73.83% mIoU and 76.44% oIoU after removing explicit attribute words. These results indicate that the model does not rely exclusively on explicit attribute labels and can use the remaining appearance descriptions and contextual cues for attribute-conditioned target segmentation. We will explicitly position UniRefDet as an attribute-conditioned referring segmentation benchmark with joint attribute-aware evaluation, and clarify the capabilities and limitations revealed by the additional experiments.

Q2. Genuine No-Target Rejection

We appreciate this important question and agree that the original fixed “background” prompt should not be considered a genuine absent-referent query. First, replacing the fixed background token with a complete natural-language background-selection instruction yields 100% background mIoU/oIoU on all 400 Scene-D images. This demonstrates that the model can accurately follow an explicit background-selection instruction, but does not itself measure rejection when a referred target is absent.

We therefore additionally evaluated 400 ordinary natural-language candidate queries describing absent targets on the 400 background-only images. The model produced strictly empty foreground masks for 359/400 queries (89.75%), while the remaining 41 queries (10.25%) produced nonempty foreground predictions. Background mIoU/oIoU reached 99.49%/99.66%, and predicted foreground pixels accounted for 0.3449% of all evaluated pixels. Compared with the fixed background prompt, this evaluation more directly examines the model's behavior under potential absent-referent conditions. These results provide evidence of no-target rejection behavior on background-only images. We will clearly distinguish the two evaluation protocols and their respective evidence for background selection and no-target rejection in the revised manuscript.

Q3. Methodological Contributions of LASA and ADM

We acknowledge that lightweight adapters, score calibration, and candidate aggregation have precedents in existing research. Our methodological contribution lies in their task-specific design for fine-grained language grounding and competing salient/camouflaged predictions. UniRefDet expressions describe not only object categories but also shape, color, quantity, and object-environment relationships. LASA is designed to adapt SAM 3 to these fine-grained descriptions, while ADM coordinates attribute-specific candidate predictions rather than simply merging two masks. We do not claim that every underlying operation is individually novel; instead, the components are specifically designed for the UniRefDet setting. We provide additional experiments below and will clarify LASA's relationship to conventional text adapters and SAM3-I in the revised manuscript.

1. LASA versus Conventional SAM 3 Adaptation

Following your suggestion, we compared Text-Encoder LoRA and Decoder Fine-tuning using the same SAM 3 backbone. The additional adaptation baselines follow a common training protocol and are evaluated across three random seeds.

Table R2. Adaptation methods on Scene C (mIoU/oIoU, %).

MethodSalientCamouflagedAdaptation Params.
Text-Encoder LoRA78.10/77.4765.05/63.313.15M
Decoder Fine-tuning75.40/70.3158.03/53.0613.87M
Full CoRefNet78.14/78.7965.97/66.033.22M

The 3.22M count refers specifically to the trainable parameters of the LASA text-adapter module. Compared with Text-Encoder LoRA, CoRefNet achieves numerically higher results on all four Scene-C metrics: salient-object mIoU/oIoU improves by 0.04/1.32 percentage points, and camouflaged-object mIoU/oIoU improves by 0.92/2.72 percentage points. The advantage is more pronounced for camouflaged objects, particularly in oIoU. Full CoRefNet also outperforms Decoder Fine-tuning, while LASA uses approximately 76.8% fewer trainable parameters than that decoder configuration. Its parameter budget is comparable to Text-Encoder LoRA's, indicating that the performance advantage is not simply associated with a substantially larger adaptation module. Together with the existing component ablations, these results support the effectiveness of the proposed adaptation framework, particularly in challenging coexistence scenes.

2. ADM versus Simple Dual-Prompt Fusion

To further assess ADM, we compared the complete decoding strategy with independent predictions and conventional fusion alternatives.

Table R3. Dual-prompt fusion strategies on Scene C (mIoU/oIoU, %).

MethodSalientCamouflaged
Independent predictions75.28/74.5064.62/53.58
Independent predictions + pixel-score fusion77.23/76.3865.17/61.77
Same dual-query candidates + pixel-score fusion77.21/76.4465.35/61.82
Full ADM78.15/78.8065.97/66.03

Compared with pixel-score fusion using the same dual-query candidates, full ADM improves salient-object mIoU/oIoU by 0.94/2.36 percentage points and camouflaged-object mIoU/oIoU by 0.62/4.21 percentage points. This supports the effectiveness of the complete ADM decoding strategy relative to the tested simple fusion methods. We additionally conducted a direct pixel-level cross-attribute confusion analysis.

Confusion DirectionWithout ADMWith ADM
Salient to Camouflaged10.10%2.69%
Camouflaged to Salient20.48%7.51%

ADM reduces the two cross-attribute error rates by 7.41 and 12.97 percentage points, respectively. These results directly support its effectiveness in alleviating pixel-level confusion between salient and camouflaged predictions. We will describe ADM more precisely as an attribute-competition-aware candidate aggregation and decoding strategy, emphasizing its coordination of competing salient and camouflaged predictions.

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

The paper builds UniRefDet, a referring-segmentation benchmark of 6,000 images taken from USC12K and split evenly into scenes with only salient objects, only camouflaged objects, both, or neither. Each target gets a long expression, drafted by GPT-5.5 and checked by another VLM and three annotators, covering eight cues such as position, scene context, and saliency or camouflage cues. The paper also proposes an approach, CoRefNet, builds on SAM-3. CoRefNet adds adapters to the text encoder, and a module (ADM) that aggregates the top-K mask proposals per attribute and suppresses pixels claimed by both attributes. It gets 82.5 overall mIoU against 80.2 for LISA, and also reports zero-shot gains on COD and RefCOCO+ test sets.

Critical strengths
  1. The paper proposes a reasonable new setting with a balanced benchmark. Scenes with salient objects, camouflaged objects, both, and neither are covered, and the same category can appear with opposite attributes in one image. Textual expressions are long and cover eight cues, which gives a testbed for separating two same-category objects that differ in context. Annotation combines a VLM draft, an independent VLM check and two rounds of review by three annotators.

  2. Thirteen baselines from SOD/COD, unified and referring-segmentation families are retrained on the same 4,400 images and scored under one three-class protocol. CoRefNet is the best on both mIoU and oIoU in every reported column of Table 1, results hold on ten complementary metrics, and cross-dataset transfer is reported on five external sets.

Critical weaknesses
  1. The scene type is known at inference. Depending on the scene, the input is the salient expression, the camouflaged expression, both together, or the fixed string "background", so the model never decides whether a target exists or which attribute it has. The introduction presents exactly this decision as the goal of the benchmark, and results with one uniform prompt format for all images would show how much is left to infer.

  2. ADM adds little where it is meant to help and recombines standard parts. Without it the overall mIoU is still 80.69, above every baseline, and the mIoU on camouflaged targets in coexistence scenes is unchanged at 66.11 versus 65.97.

  3. The expression analyses do not support the claim that object-environment relations matter. In coexistence scenes, removing position cues costs 6.3 mIoU while removing scene context costs 0.03, and since expressions were drafted with the target contour overlaid, they seem to mainly point to the target.

Questions for authors
  1. The scene type is given through the input, the expressions were written with the target contour visible and carry explicit cues, and overlap between training and zero-shot test images is not addressed. Please report results with one uniform prompt format, with short user-style prompts and for Scene D alone, and confirm there is no image overlap. If the benchmark holds up under these checks, I would raise my Rating.

  2. The large gain over the frozen model in Table 3 is not separated from the effect of fine-tuning in general. Could you compare the Text Adapter with other ways of fine-tuning SAM 3 on the same expressions, such as LoRA or tuning the decoder?

3 comments

We sincerely thank you for the valuable and constructive feedback. Below, we address your concerns point by point. We would be grateful if you could let us know whether our responses adequately address your questions and concerns.

Q1. Scene Information Leakage and Prompt Robustness

UniRefDet primarily investigates attribute-conditioned referring segmentation, where the challenge lies in grounding language descriptions to objects with competing visual attributes. For instance, two animals of the same category may appear in one image, one salient and the other camouflaged. Although language specifies the desired target, the model must still distinguish the corresponding instances through visual evidence, particularly when their predicted regions overlap. This setting has potential applications in wildlife monitoring and UAV-assisted search and rescue.

We acknowledge that the original protocol provides explicit attribute cues. To assess their influence, we conducted end-to-end retraining after removing Attribute IDs from Scenes A/B/D, followed by an additional evaluation with explicit attribute words removed from the referring expressions.

Table R1. Performance with reduced explicit attribute cues (mIoU/oIoU, %).

SettingA: Sal.B: Cam.C: Sal.C: Cam.
Without Attribute ID87.82/88.6883.15/86.3880.67/80.3573.83/76.59
Further removing attribute words87.82/88.5881.51/84.0280.65/80.3773.83/76.44

Performance remains strong after removing explicit attribute cues. Notably, Scene C camouflaged-object segmentation achieves 73.83% mIoU and 76.44% oIoU without explicit attribute words, suggesting that the model can exploit additional appearance and contextual information rather than relying solely on explicit attribute terms.

We also evaluated shorter, user-style expressions. The average mIoU decreases from 82.53% (original) to 77.09% (medium) and 72.59% (short), indicating sensitivity to linguistic detail while retaining meaningful performance under shorter prompts.

For Scene D, replacing the fixed "background" token with a complete natural-language background-selection instruction yielded 100% background mIoU/oIoU in the corresponding evaluation. To further assess no-target rejection, we tested ordinary referring expressions describing absent objects on all 400 background-only images. The model produced strictly empty foreground masks for 359/400 images (89.75%). The background mIoU/oIoU reached 99.4924%/99.6551%, with a global false-positive foreground pixel ratio of only 0.3449%. These results demonstrate substantial no-target rejection capability on background-only images, while indicating that further evaluation with absent referents in object-containing scenes remains valuable. We will clarify these evaluation settings and their respective conclusions in the revision.

Q2. Training–Test Image Overlap

We appreciate the reviewer's concern regarding potential data leakage. We performed an image-level overlap audit and confirmed no overlap between UniRefDet's 4,400 training images and 1,600 held-out test images. We additionally verified that the training images do not overlap with any external COD/RIS test sets used in our cross-dataset evaluation. These checks are documented in the supplementary material, supporting the validity of the reported cross-dataset generalization results.

Q3. Effectiveness and Independent Contribution of ADM

We acknowledge that ADM does not improve all metrics uniformly. In the original ablation, Scene C camouflaged-object mIoU changes slightly from 66.11% to 65.97%, whereas oIoU increases substantially from 57.05% to 66.03% (+8.98 points). This discrepancy suggests that ADM's benefits are more pronounced in aggregated pixel-level overlap, reflecting improvements on larger camouflaged regions.To further isolate ADM's contribution, we compared it with simpler dual-prompt fusion strategies under controlled settings.

Table R2. Scene C comparison of fusion strategies (mIoU/oIoU, %).

MethodSalientCamouflaged
Independent predictions75.28/74.5064.62/53.58
Independent + pixel-score fusion77.23/76.3865.17/61.77
Same dual-query candidates + pixel-score fusion77.21/76.4465.35/61.82
Full ADM78.15/78.8065.97/66.03

Compared with pixel-score fusion using the same dual-query candidates, ADM improves salient-object mIoU/oIoU by 0.94/2.36 points and camouflaged-object mIoU/oIoU by 0.62/4.21 points, demonstrating its advantage over the tested conventional fusion strategies.

More importantly, we conducted a direct pixel-level attribute-confusion analysis. Without ADM, the Salient→Camouflaged and Camouflaged→Salient confusion rates are 10.10% and 20.48%, respectively. With ADM, they decrease to 2.69% and 7.51%, corresponding to reductions of 7.41 and 12.97 percentage points. These results directly support ADM's effectiveness in reducing cross-attribute confusion.

We acknowledge that individual operations such as score calibration and candidate aggregation are established techniques. Accordingly, we will clarify ADM's contribution as an attribute-competition-aware candidate aggregation and decoding strategy, rather than attributing its benefits to the novelty of individual operations.

Q4. Linguistic Cues and Object–Environment Relationships

We appreciate the reviewer's observation that the original linguistic ablation did not adequately establish the role of object–environment relationships. We re-audited 3,000 Scene C expressions and examined the original ablation implementation. We found that the previous w/o Scene Context operation did not consistently remove environmental relationship information: an audit identified residual relationship cues in 310 of 800 validation expressions, often embedded within saliency/camouflage descriptions.

For example, describing an object's texture as matching the surrounding vegetation conveys both camouflage attributes and object–environment relationships. Thus, removing the Scene Context category alone did not eliminate all environmental information, explaining the previously observed negligible 0.03-point mIoU decrease.

We therefore performed additional controlled analyses. Removing position cues decreases mIoU by 6.30 points confirming the importance of spatial information. Furthermore, compared with category-only descriptions, incorporating complete salient-related and camouflage-related descriptions improves mIoU by 11.31 and 10.23 points, respectively.

These findings demonstrate that fine-grained attribute-related language provides substantial information beyond object categories alone. However, because attribute and environmental cues are semantically intertwined, these results do not independently quantify the contribution of each relationship cue. We will correct the original ablation interpretation and clarify this distinction.

Q5. Comparison with Conventional SAM 3 Fine-tuning

Following the reviewer's suggestion, we conducted additional comparisons with Text-Encoder LoRA and Decoder Fine-tuning, using the same SAM 3 backbone and a common training protocol, with evaluations across three random seeds.

Table R3. SAM 3 adaptation comparison on Scene C (mIoU/oIoU, %).

MethodSalientCamouflagedTrainable Params.
Text-Encoder LoRA78.10/77.4765.05/63.313.15M
Decoder Fine-tuning75.40/70.3158.03/53.0613.87M
Full CoRefNet78.14/78.7965.97/66.033.22M

Compared with Text-Encoder LoRA, CoRefNet achieves numerically higher results on all four Scene C metrics, particularly for camouflaged objects, where mIoU/oIoU improves by 0.92/2.72 points. It also substantially outperforms Decoder Fine-tuning. Notably, LASA requires approximately 76.8% fewer trainable parameters than the decoder fine-tuning configuration, while remaining comparable in parameter budget to Text-Encoder LoRA.

These results, together with the existing component ablations, support the effectiveness of our proposed adaptation framework, particularly under challenging coexistence scenarios. We will further distinguish the system-level improvements of CoRefNet from the individual contribution of LASA in the revised manuscript.

Thank you for the rebuttal and the extra experiments. The train/test overlap check resolve my concern. I also appreciate that you rechecked the cue ablation and admitted the original "scene context" result does not support the object–environment claim. The new baselines show ADM does reduce cross-attribute confusion.

But there are still two concerns:

  1. The retraining without attribute IDs and attribute words is helpful, but it does not remove the scene-dependent query problem. Scene C still uses separate salient and camouflaged expressions, while Scene D uses a background-selection instruction. Short-prompt results are reported only as an overall average. The Scene D test shows the ability of no-target rejection, but it doesn't cover the case where the image has targets, and the query asks for one that isn't in it.

  2. Table R3 compares full CoRefNet with Text-Encoder LoRA, not LASA alone. mIoU of scene C gains are 0.04/0.92 points for salient/camouflaged targets. So LASA's distinct contribution is still not isolated. In Table R2, ADM gains 0.94/0.62 mIoU points over pixel-score fusion using the same candidates. The larger oIoU and confusion-rate improvements, but the mIoU gain is small.

Overall, I keep my rating at 2

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.