{"id":"0696adc8-c754-4254-95d4-da29d0106456","arxiv_id":"2412.16609","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ConceptCoSOD extracts a shared text embedding from an image group and guides co-salient object segmentation with it, outperforming five baselines on three clean and five corrupted datasets.","lead":"A computer vision method called ConceptCoSOD learns a text-based concept of the common object from a group of photos, then uses that concept to segment the object across the whole group. It applies semantic priors from diffusion models to co-salient object detection, an approach the authors say is new.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The learned concept token is not constrained to encode the co-salient object; Eq. (5)'s reconstruction loss can capture shared background or appearance, so the semantic-anchor mechanism central to the claim is unvalidated.","rationale":"The empirical evaluation issues the reader identifies (test-set tuning of α and λ, only five baselines, no code) are real and contribute to CONDITIONAL, but they do not target the conceptual core. The concept-alignment assumption is more fundamental: the paper's novelty is that high-level semantics from a diffusion model anchor detection, and Eq. (5) provides no guarantee that the learned token corresponds to the co-salient object. Because the paper claims to be the first concept-guided Co-SOD, the method must demonstrate that the concept is actually the semantic anchor. The proposed IoU-vs-oracle test would directly adjudicate this. If the test fails, the method reduces to generic TI+LD-ZNet with test-time tuning, weakening novelty and the claim. If it passes, the central claim is supported. Thus the reader's CONDITIONAL verdict is appropriate, with the condition extended to include concept-alignment validation.","tokens_in":12644,"tokens_out":7854,"duration_ms":73280,"concrete_test":"For each group in CoCA/Cosal2015, extract the cross-attention map of c* from AttnExtract and compute IoU with the GT mask; compare against (a) a random/unrelated token and (b) the oracle class-name token. If learned c* attention is not significantly above (a) and does not approach (b) on groups with repetitive backgrounds or small objects, the concept-alignment assumption fails. Additionally, build synthetic groups sharing a fixed background with different foreground objects; if ConceptCoSOD segments the background rather than the object, the concern lands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that c* from Eq. (5) encodes the shared salient object. But Eq. (5) is a Textual Inversion reconstruction loss over the whole image group: it minimizes pixel-level noise prediction error for the prompt 'a photo of S*', with no co-saliency, objectness, or cross-image-consistency term. Any common, easily reconstructable signal—a repetitive background, a dominant color cast, or a high-frequency texture—can be absorbed into c*. The diffusion cross-attention module then localizes whatever c* denotes, so if c* denotes background or generic appearance, the 'semantic anchor' is not the co-salient object and the mechanism collapses. Nothing in the paper verifies alignment: there is no nearest-token analysis, no correlation between attention maps of c* and GT, and no failure study on groups with shared backgrounds or small objects. Figure 4(a) itself shows learned concepts underperform native class-name tokens, so the gap between the ideal semantic anchor and the actual token is acknowledged. The generic Limitation section (5.6) does not address this. The reported averages could still hold, but the claimed conceptual explanation for the gains is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ConceptCoSOD proposes a two-stage co-salient object detection framework: first, a learnable concept token c* is obtained by applying Textual Inversion to the input image group with a fixed prompt, and second, this token is used as semantic guidance in a diffusion-based segmentation module derived from LD-ZNet, whose cross-attention maps are extracted and refined into a final binary mask. A timestep-resampling strategy is introduced so that middle diffusion timesteps are sampled more heavily when learning the concept. The method is evaluated on Cosal2015, CoSOD3k, and CoCA, on corrupted versions of Cosal2015 under five corruptions, and against five prior Co-SOD methods, reporting top or near-top scores on most metrics.","tokens_in":12826,"tokens_out":7303,"duration_ms":60796,"significance":"If the central mechanism is validated, the paper's idea of importing diffusion-model textual personalization into co-salient object detection is novel and represents a plausible route to semantic robustness: a concept token could in principle anchor detection so that object parts are not segmented as wholes and distractors are excluded. The consistency of the reported improvements across three clean datasets and five corrupted settings, the use of standard components (Textual Inversion and LD-ZNet), and the explicit ablation of the resampling ratio and threshold are all strengths. However, the paper's main explanatory claim—that the learned token is a reliable semantic anchor for the co-salient object—is not directly verified, and several evaluation choices currently prevent the results from being taken at face value: hyperparameters are selected on the test benchmarks, the resampling distribution as written is not normalized, the segmentation module's training protocol is unspecified, and the threshold-selection claim is contradicted by the paper's own table.","major_comments":[{"comment":"The default hyperparameters α=1.5 and λ=0.5 are selected by evaluating the model on Cosal2015, CoSOD3k, and CoCA, which are the same datasets used for the final comparison against baselines. This makes the reported gains at least partially a result of test-set tuning rather than predictive performance. The timestep interval boundaries in Fig. 4(b) appear to be chosen in the same way. Please re-run with a held-out validation split, or report results for defaults fixed a priori.","section":"§5.1 and Tables 3–4"},{"comment":"The central claim is that c* encodes the shared salient object and serves as a semantic anchor, but Eq. (5) is a per-image reconstruction loss with no co-saliency, objectness, or cross-image consistency term. Shared background, a color cast, or a common texture can be absorbed into c*, and nothing in the paper verifies alignment between c* and the co-salient ground truth. The paper itself shows in Fig. 4(a) that native class-name tokens outperform the learned concept, so the learned token is not obviously a better semantic anchor. Please provide evidence such as nearest-token analysis, correlation between c* cross-attention maps and ground truth, or a failure study on groups with shared backgrounds.","section":"§4.2, Eq. (5), and Fig. 4(a)"},{"comment":"The resampling distribution is claimed to be normalized, but integrating the three intervals gives |T1|·(1/(2|T1|(1+α))) + |T2|·(α/(|T2|(1+α))) + |T3|·(1/(2|T3|(1+α))) = (1+2α)/(2(1+α)), which equals 0.8 for the default α=1.5, not 1. Please correct the formula or the normalization statement, and clarify whether the experiments used the written distribution or a renormalized version.","section":"§4.4, p_alpha(t)"},{"comment":"The concept-guided segmentation module is not fully specified. The paper says LD-ZNet is employed and FineSeg takes z_t, c*, and the attention map as inputs, but it does not state the training loss, the training data, or whether FineSeg is fine-tuned on the three benchmark datasets or used off-the-shelf. Without this information, the reader cannot determine whether the evaluation protocol is fair or whether the segmentation module itself has seen the test groups.","section":"§4.3 and §5.1"},{"comment":"The text states that 'threshold values of 0.5 provide optimal performance across six metrics,' but Table 4 contradicts this claim. On CoCA, MAE at λ=0.5 is 0.0886 and decreases to 0.0737 at λ=0.8; on Cosal2015, Sm at λ=0.6 (0.8485) is higher than at λ=0.5 (0.8467). The selection rationale should be clarified, or the claim should be corrected to reflect that 0.5 is a reasonable compromise rather than optimal on every metric.","section":"Table 4 and §5.5"}],"minor_comments":[{"comment":"The notation t∼U([0,1]) in Eq. (5) conflicts with the later use of discrete timesteps in [0,1000) in §5.1; please unify the notation.","section":"§4.2, Eq. (5)"},{"comment":"The abstract and experiments describe three clean datasets plus five corrupted settings, while the contribution list says 'seven datasets'; please rephrase so that corrupted settings are not counted as separate datasets.","section":"§1, contribution list"},{"comment":"Only five baselines from 2020–2021 are compared. Given the active progress in Co-SOD, the abstract's claim of 'significantly outperforming existing methods' should be tempered, or recent competing methods should be included.","section":"§5.1"},{"comment":"The native class-name token condition is described as a 'theoretical upper bound,' but class-name tokens are not a formal upper bound for the task; please rephrase this description.","section":"Fig. 4(a)"},{"comment":"The section heading 'Border Impact' appears to be a typo for 'Broader Impact'; please correct it.","section":"§5.7"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting idea with consistently strong reported numbers, but the evaluation protocol needs substantive changes before publication: test-set hyperparameter selection must be eliminated, the resampling formula bug fixed, and the segmentation training protocol clarified. The unverified concept-alignment claim is the deepest issue because it is the paper's main explanatory mechanism; the authors should be asked to provide direct evidence or substantially soften the claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new: learn a shared text token from the image group via Textual Inversion, then use it to guide diffusion-based co-saliency segmentation. The empirical results are consistently in its favor across three clean benchmarks and five corruptions. But the central mechanism is under-validated: nothing ensures the learned token c* encodes the co-salient object rather than shared background or low-level appearance. That's not a fatal flaw, but it is the load-bearing assumption, and the paper doesn't test it.\n\nWhat's good: The combination of personalization (Textual Inversion) with LD-ZNet segmentation for Co-SOD is not in the cited literature. The timestep resampling analysis (Fig. 4b) is a useful observation about where informative signal lives in diffusion training. The gains over the five 2020-2021 baselines are fairly consistent, especially on CoCA, which is the hard dataset. The corruption robustness story, if true, is a real plus.\n\nSoft spots: First, the stress-test concern is valid. Eq. (5) is a reconstruction loss over the whole group with no co-saliency or objectness constraint. Figure 4(a) even shows native class tokens outperform the personalized token, which suggests c* is a weak anchor. The paper needs failure analysis: nearest-token visualization, attention-GT correlation, or at least a qualitative look at groups with shared backgrounds. Second, the evaluation is tuned on the test set: α (Table 3) and λ (Table 4) are chosen by best test metrics. That's mild, but it inflates the reported margins. Third, the baselines are 2020-2021; there are newer Co-SOD methods that should be compared. Fourth, no code is provided, and test-time concept learning per group (up to 2000 steps) is a real compute asymmetry versus single-pass baselines. Fifth, the abstract's 'seven datasets' doesn't match the actual three-plus-corruptions setup.\n\nBottom line: This deserves a serious referee, not a desk reject. The idea is worth publishing if the authors add concept-validation experiments, move hyperparameter selection to a validation split, add newer baselines, and release code. The central insight—semantic prior from diffusion concept learning—is plausible, but the paper currently argues it more than demonstrates it.","headline":"A genuinely new idea with consistent empirical wins, but the semantic anchor's validity is unverified and the evaluation has test-set tuning.","tokens_in":13385,"tokens_out":3612,"would_cite":true,"duration_ms":28761,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Co-salient object detection improves when the model first learns a text concept naming the shared object and uses that concept to guide segmentation.","keywords":["co-salient object detection","diffusion models","textual inversion","semantic priors","vision-language models","timestep resampling","segmentation","corruption robustness"],"falsifier":"Take a group of images taken in the same room with a consistent background but different salient objects, run the concept-learning step on that group, and inspect whether the resulting segmentation highlights the background or the intended objects; if the background wins or the token reconstructs the room, the central assumption is false.","tokens_in":12388,"feed_emoji":"🎯","tokens_out":7681,"duration_ms":60762,"temperature":0.7,"pith_summary":"Co-salient object detection asks a model to find the same salient object across a group of related images. This paper argues that existing methods fail when they rely only on low-level visual patterns, because a bottom-up view of a guitar can be reduced to its strings and a lemon can be absorbed into an apple. The proposed framework, ConceptCoSOD, first learns a text-based concept token that names the shared object from the group itself, using a diffusion model personalization approach, and then conditions segmentation on that token. The paper reports that this semantic anchor outperforms five recent co-saliency methods on three benchmark datasets and on five corrupted settings, including blur, noise, frost, and adversarial camouflage.","feed_headline":"A text concept learned from the image group sharpens co-salient detection","feed_subtitle":"A diffusion-learned semantic anchor cuts part/whole errors and holds against blur, noise, frost and attacks.","key_machinery":"The central object is the learned concept token $c^*$, a learnable embedding that replaces the placeholder 'S*' in the prompt 'a photo of S*' and is trained with the latent diffusion denoising loss (Eq. 5). This token carries the semantic identity of the co-salient object and acts as the anchor that the segmentation module conditions on. The supporting mechanism is the timestep resampling distribution $p_\\alpha(t)$, which increases sampling probability in the middle interval of the diffusion timestep range and decreases it in the head and tail, based on the paper's empirical study showing that middle timesteps are the most informative for concept quality. A coarse-to-fine segmentation submodule, consisting of attention extraction followed by a fine-grained segmentation network with an extra cross-attention input, converts the concept and image latent into the final saliency map.","core_discovery":"On the paper's own terms, the central claim is that high-level semantic knowledge, extracted automatically from the image group, is what lets co-salient detection distinguish essential object structure from incidental appearance. The method learns a concept token $c^*$ by optimizing a denoising diffusion objective (Eq. 5) over the group's images, so that the token functions like the name of the shared object. That token is then injected into a segmentation module built on the same diffusion architecture, which extracts an attention map as coarse localization and refines it into a binary saliency map. The paper also claims that the choice of diffusion timestep matters for concept quality: sampling more densely from the middle timestep interval $[300,800)$ improves final detection, and the resulting method reaches top reported numbers on Cosal2015, CoSOD3k, and CoCA, and under all five corruption settings.","pith_inferences":["A natural extension the authors do not test: when the image group shares a dominant background while the co-salient object is small or peripheral, the learned token may encode the background instead of the object; such groups would be a useful stress test for the method.","The same group-level concept extraction could be applied to video object co-localization or weakly supervised semantic segmentation, where low-level consensus is also fragile, though the paper does not explore these tasks.","The timestep resampling insight likely transfers beyond Co-SOD: other personalization-based segmentation or detection methods that rely on text embeddings could benefit from the same middle-range bias, which is a testable hypothesis rather than a claim in the paper.","Because concept learning is per-group optimization, latency is a practical bottleneck; an amortized encoder that predicts $c^*$ directly from the group would make the approach scalable, but that is not proposed here."],"forward_implications":["If the central claim is correct, co-salient detection can be driven by a semantic prior learned from the group itself, without any manual text annotations for new object categories.","The reported gains on corrupted inputs imply that concept-guided detection is more resistant to blur, noise, frost, and adversarial camouflage than low-level consensus methods.","The timestep analysis suggests that diffusion-based concept learning for segmentation should oversample the middle denoising range rather than treat all timesteps uniformly.","Because the pipeline uses an existing segmentation module and a pretrained text-to-image diffusion model, the approach can be assembled from off-the-shelf components and does not require training a new detector from scratch."],"supporting_citations":[{"why":"Supplies the Textual Inversion method used in the implementation to learn the concept token from the image group.","marker":"[12]"},{"why":"Supplies the personalization formulation that the concept-learning module draws on.","marker":"[11]"},{"why":"Provides the latent diffusion model that the concept learning and segmentation modules are built on.","marker":"[30]"},{"why":"Provides the LD-ZNet segmentation approach that the concept-guided segmentation module adapts.","marker":"[29]"},{"why":"Provides the Cosal2015 benchmark used for clean and corrupted evaluation.","marker":"[44]"},{"why":"Provides the CoSOD3k benchmark used for evaluation.","marker":"[8]"},{"why":"Provides the CoCA benchmark, the hardest case where the method reports the largest gains.","marker":"[50]"},{"why":"Defines the common corruption types (frost, motion blur, defocus blur, Gaussian noise) used in robustness tests.","marker":"[16]"},{"why":"Defines the Jadena adversarial camouflage attack used to test robustness.","marker":"[14]"}],"fun_headline_variants":["Concept token sharpens co-salient masks","Diffusion-learned concept anchors co-saliency","Semantic guide improves co-salient detection","Concept-guided Co-SOD tops benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the concept token learned by reconstructing the group encodes the shared salient object rather than the shared background or other common appearance, because nothing in the objective explicitly forces the token to match the co-salient target.","fun_headline_variants_meta":{"raw":{"variants":["Concept token sharpens co-salient masks","Diffusion-learned concept anchors co-saliency","Semantic guide improves co-salient detection","Concept-guided Co-SOD tops benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1342,"prompt_tokens":872,"completion_tokens":470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":415}},"tokens_in":488,"tokens_out":470,"duration_ms":4573,"temperature":1.0,"reasoning_tokens":415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:24:45.653848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a group of images taken in the same room with a consistent background but different salient objects, run the concept-learning step on that group, and inspect whether the resulting segmentation highlights the background or the intended objects; if the background wins or the token reconstructs the room, the central assumption is false.","supporting_citations":[{"cited_title":"H.; Chechik, G.; and Cohen-or, D","cited_arxiv_id":null,"evidence_quote":"Supplies the personalization formulation that the concept-learning module draws on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the latent diffusion model that the concept learning and segmentation modules are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LD-ZNet segmentation approach that the concept-guided segmentation module adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Cosal2015 benchmark used for clean and corrupted evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CoSOD3k benchmark used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CoCA benchmark, the hardest case where the method reports the largest gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the common corruption types (frost, motion blur, defocus blur, Gaussian noise) used in robustness tests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Jadena adversarial camouflage attack used to test robustness."}],"review_version":1}