- 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
-
CoSEC achieves strong co-saliency detection with efficient inference and an interpretable textual interface, without task-specific training.
-
Evaluations across multiple VLMs and consensus strategies, together with ablation and oracle studies, support the framework’s flexibility and clarify its behavior.
Critical weaknesses
- 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.
- 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
-
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?
-
How were the threshold grid and candidate-count settings selected without using test-set performance for configuration choice?
-
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)