acceptodds
Withdrawn from ICLR 2027

Revisiting Co-Salient Object Detection via Explicit Semantic Consensus

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Co-salient object detection (CoSOD) aims to locate and segment the common salient object across a group of images. Existing methods infer group-level consensus in visual feature space and carry this implicit representation into object localization and segmentation. Yet visual similarity can cause semantically different distractors to enter the implicit consensus, reducing its reliability for segmentation. This motivates us to reconsider whether the essential shared information must remain encoded in a rich implicit visual representation, or can instead be preserved in a cleaner and more explicit form. We propose CoSEC (Co-Salient Explicit Consensus), a fully training-free framework that represents group-level consensus as a compact textual concept and uses it as the sole interface between group-level reasoning and per-image localization. A Semantic Consensus Estimator identifies the common object from the image group using a multi-image vision-language model, while a Consensus-Guided Segmenter independently grounds this concept in each image using a frozen text-conditioned segmenter. Without task-specific training or fine-tuning, CoSEC outperforms existing methods across all three benchmarks. On the challenging CoCA benchmark, it achieves an MAE of and an of , while requiring only s/image on a single GPU. These results suggest that once the shared semantic identity is reliably established, a compact explicit semantic consensus can replace a rich implicit visual representation for strong CoSOD performance. Code will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Co-Salient Object Detection, Vision-Language Models, Salient Object Detection, Open-Vocabulary Segmentation, Training-Free

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_qYc1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CoSEC, a training-free framework for co-salient object detection that represents group consensus as a short textual concept. A frozen vision-language model identifies the shared object, which guides independent image segmentation followed by adaptive mask selection and merging.

Critical strengths
  1. CoSEC achieves strong co-saliency detection with efficient inference and an interpretable textual interface, without task-specific training.

  2. Evaluations across multiple VLMs and consensus strategies, together with ablation and oracle studies, support the framework’s flexibility and clarify its behavior.

Critical weaknesses
  1. Limited methodological novelty.

The paper does not convincingly establish a substantive methodological advance beyond assembling existing foundation-model capabilities. SCE is essentially a fixed object-naming prompt applied to a frozen VLM, while CGS combines text-conditioned SAM 3 inference with a confidence-threshold heuristic and mask union. Given established MLLM-to-SAM 3 pipelines such as SAM 3 Agent, inferring one shared object name and reusing it across an image group appears to be an incremental task-specific adaptation. The manuscript provides insufficient evidence that this constitutes a significant advance in consensus modeling. The experiments also fail to isolate a distinct contribution from the proposed representation. Every CoSEC variant retains SAM 3, whereas the competing methods use different foundation models and representations, confounding the claimed advantage of explicit consensus with differences in pretrained capabilities. Substituting VLMs demonstrates compatibility across models but does not address this issue. Moreover, the “single-image” variants still perform cross-image voting, leaving independent per-image naming and segmentation without group consensus untested. Without controlled comparisons using the same underlying models, the strong benchmark results do not establish that the proposed consensus formulation contributes a sufficiently novel scientific insight.

  1. Insufficient analysis of error propagation from VLM consensus.

The authors explicitly acknowledge in Section 4.6 that an incorrect VLM prediction propagates to all images in a group without self-correction. However, the paper does not systematically examine how this failure mode affects downstream segmentation. Appendix A.3 reports consensus accuracy and categorizes naming errors, while Table 4 compares predicted consensus with ground-truth category names at the dataset level. These analyses establish that consensus quality matters, but do not characterize segmentation performance within the groups where consensus prediction fails. Although the reported end-to-end results already reflect the effects of VLM errors, aggregate scores do not reveal whether an incorrect shared prediction causes limited degradation or widespread segmentation failures across the corresponding group. A breakdown of segmentation performance by consensus correctness and error type, together with the proportion of affected images within each group, would clarify the severity of this limitation. Selectively correcting erroneous consensus predictions while keeping the segmenter fixed would further help distinguish failures caused by target misidentification from those caused by the segmenter itself. Without such analysis, the consequences of relying on a single, uncorrected textual prediction for an entire image group remain insufficiently characterized.

Questions for authors
  1. How does the method behave when several object categories are shared across the group, or when target and distractor instances belong to the same category?

  2. How were the threshold grid and candidate-count settings selected without using test-set performance for configuration choice?

  3. Could the authors verify the VCP results in Table 1? VCP’s CoSOD3k (E^max) is listed as 0.868, which matches its (F^max) score, whereas the original paper reports (E^max=0.918) for VCP-I. Was the F-measure value inadvertently duplicated in the E-measure column? Also, why was the stronger VCP-II configuration omitted? (“Visual Consensus Prompting for Co-Salient Object Detection,” CVPR 2025)

Reviewer_MFF6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that existing co-salient object detection (CoSOD) methods primarily establish implicit consensus by measuring visual feature similarities across images. However, visual similarity may cause semantically different distractors to be incorrectly included in the consensus representation, thereby reducing segmentation accuracy. To address this issue, the paper proposes CoSEC, a training-free framework based on explicit semantic consensus. Specifically, the method first employs the vision-language model Qwen2.5-VL to jointly analyze an image group, identify the common salient object, and generate a concise textual description of its semantic identity. This textual description is then used as a prompt for SAM 3 to independently segment the target object in each image. Finally, an adaptive thresholding strategy is applied to select and merge candidate masks, producing the co-salient object segmentation results for each image.

Critical strengths

The paper shows strong quantitative and qualitative results. The reported inference speed is also faster than that of the compared methods.

Critical weaknesses

The fundamental motivation of this paper is questionable. The authors argue that computing visual feature similarities across images may introduce semantically different distractors into the implicit consensus, thereby reducing segmentation reliability. To address this issue, they employ Qwen2.5-VL to identify the common salient object and generate its textual description. However, when jointly processing multiple images, Qwen2.5-VL can still model cross-image visual dependencies through its internal attention mechanisms. Therefore, the proposed framework does not fundamentally eliminate the need for cross-image visual information interaction. Instead, it largely delegates the core challenge of CoSOD—identifying the common salient object across multiple images—to a powerful pretrained vision-language model. Simply relying on an existing foundation model to solve this challenging problem should not be regarded as a methodological contribution of the proposed framework. The strong performance may primarily result from the pretrained semantic reasoning capabilities of Qwen2.5-VL and the segmentation capabilities of SAM 3, rather than from the proposed explicit semantic consensus formulation. Consequently, the authors have not sufficiently demonstrated that their method fundamentally addresses the limitations of visual similarity-based consensus estimation. More convincing analysis and controlled experiments are needed to establish the independent contribution and novelty of the proposed approach.

The paper contains some basic writing errors, such as the repetition in the first sentence of the abstract.

Questions for authors

The proposed framework employs two large pretrained models, yet the reported inference time is surprisingly low. The authors should carefully verify the correctness and fairness of the inference speed comparisons. In particular, the input image resolution should be explicitly specified, and the authors should clarify whether all methods were evaluated under consistent experimental settings. Furthermore, a comparison of computational complexity in terms of FLOPs would provide a more objective assessment of efficiency, reducing the influence of differences in hardware configurations and implementation optimizations.

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

This paper investigates training-free co-salient object detection and proposes Co-Salient Explicit Consensus (CoSEC), which combines a VLM-based semantic consensus estimator with a SAM3-based consensus-guided segmenter. Experimental results demonstrate improvements on benchmark datasets.

Critical strengths
  • The potential novelty lies in the adaptive thresholding strategy, but it appears to be a simple heuristic for selecting among predefined thresholds.
Critical weaknesses
  • The methodological novelty appears incremental, as the framework primarily integrates a pretrained VLM with SAM3 for co-salient object detection.
Questions for authors
  • Lines 320–323 and Table 1: What is the purpose of comparing different frozen VLMs beyond reporting performance differences? Could the authors analyze which models are better at identifying salient objects and estimating semantic consensus?
  • Table 2: Were the parameters analyzed and selected using the test datasets or a separate validation set? If test datasets were used, how did the authors avoid test-set tuning?
  • Table 3: The prompt ablation is conducted using Qwen, whereas the baseline prompt is applied to all VLMs in Table 1. Could model-specific prompt sensitivity affect the fairness of this comparison? Would the conclusions remain consistent if prompts were evaluated across all VLMs?

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

Related papers

Discussion (0)

Sign in to comment.