{"id":"559c8c3c-adca-409b-8153-d631779b3d34","arxiv_id":"2501.04582","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A text-driven BLIP-GroundingDINO-SAM pipeline produces pseudo-labels and a new 260k-image dataset for SOD, with claimed SOTA results that are weakened by a likely PASCAL-S overlap and missing artifacts.","lead":"Using text prompts plus large vision models, the authors generate saliency masks cheaply and assemble a 260,000-image training set, BDS-TR. They report state-of-the-art weakly supervised salient object detection, but one evaluation benchmark may not be independent of the training data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main Table I 'Ours' results appear to be trained on DUTS-TR, not the proposed BDS-TR: the numbers match the 'DUTS' row of Table III and Section IV-A names an undefined 'T-DUTS' training set. The headline SOTA claim therefore does not evaluate the paper's central dataset contribution.","rationale":"The reader's verdict is REJECT, and I agree, but for a different primary reason. The strongest claim in the abstract is that the proposed method (including BDS-TR) significantly outperforms SOTA on five benchmarks. For that claim to hold, the main table must report a model trained with the proposed BDS-TR pipeline. The evidence says otherwise: Section IV-A names 'T-DUTS' as the training set, never defined; Table III explicitly compares 'DUTS' and 'BDS' configurations, and Table I's 'Ours' row is numerically identical to the 'DUTS' row across all overlapping metrics. This means the headline SOTA comparison is for the DUTS-TR-trained baseline, while the BDS-TR-trained model appears only in a three-dataset ablation. The reader's VOC-overlap concern is genuine—PASCAL-S is VOC-derived and BDS-TR draws from VOC2012, with no overlap removal reported—and it further weakens the empirical case. But even if all overlap were removed, the main table still would not demonstrate the contribution of BDS-TR. I therefore center the attack on the mismatch between the claimed dataset contribution and the configuration actually evaluated. A concrete check is to have the authors identify the exact training set for Table I and to regenerate Table I with the BDS-TR model; the numbers should change to the BDS row of Table III if the claim is accurate. No code, data, or error bars are provided, so this cannot be checked independently, which strengthens the need for the requested clarification. Verdict remains REJECT, so verdict_should_be is UNCHANGED relative to the reader's assessment.","tokens_in":14787,"tokens_out":8076,"duration_ms":70391,"concrete_test":"Ask the authors to state explicitly which training set produced Table I. Cross-check: if Table I 'Ours' was trained on DUTS-TR, then Table I should be regenerated with BDS-TR and compared against the same SOTA methods; if it was trained on BDS-TR, Table I's numbers should equal Table III's 'BDS' row (they do not). Additionally, inspect the released training configs or logs to verify the dataset name; if 'T-DUTS' refers to DUTS-TR, the abstract's claim that BDS-TR is central to the SOTA result fails. A minimal resolution is to rerun the five-benchmark comparison with the BDS-TR-trained model and report both configurations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BDS-TR, the new 260k-image pseudo-labeled dataset, plus the DEDecoder, yields SOTA SOD. But the paper's main comparison, Table I, is not actually tied to BDS-TR. Section IV-A says the second-phase model is trained 'using T-DUTS as the training dataset,' and T-DUTS is never defined. More decisively, every overlapping metric in Table I's 'Ours' row exactly matches the 'DUTS' row of Table III (e.g., DUT-O: S .860, Fm .841, Eξ .899, M .045; ECSSD: S .933, Fm .958, Eξ .937, M .028; PASCAL-S: S .883, Fm .894, Eξ .889), not the 'BDS' row (DUT-O: S .875, Fm .855, Eξ .910, M .043; ECSSD: S .935, Fm .959, Eξ .941, M .027). Thus the headline 'Ours' model was trained on DUTS-TR (or some undefined T-DUTS), and the BDS-TR benefit appears only in a three-dataset ablation (Table III). If Table I is the basis for 'comprehensive experiments on five benchmarks,' the contribution of BDS-TR is not part of the SOTA comparison; if the authors intended Table I to use BDS-TR, then the reported numbers are inconsistent. The reader's concern about VOC/PASCAL-S overlap is real but secondary; the more immediate problem is that the main experimental table does not evaluate the claimed dataset. Without clarification, the paper's strongest claim—that the BDS-TR pipeline outperforms SOTA—is unsupported by the main table.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly supervised salient object detection (SOD) pipeline in which BLIP is fine-tuned on a small manually annotated text subset and then used, together with GroundingDINO and SAM, to generate pseudo-labels for a new large-scale training set called BDS-TR (about 260k images from COCO, OpenImages, and VOC2012). The authors also introduce DEDecoder, an edge-preserving dynamic upsampling decoder, and report results on ECSSD, DUT-O, PASCAL-S, DUTS-TE, and HKU-IS, claiming state-of-the-art performance among weakly supervised methods and competitiveness with fully supervised ones.","tokens_in":15136,"tokens_out":5333,"duration_ms":52971,"significance":"If the claims were fully supported, the work would be significant: a foundation-model-based pseudo-label pipeline that produces near-manual-quality SOD training data at low annotation cost, plus a large-scale dataset that improves generalization, would be a useful contribution to weakly supervised SOD. The proposed BDS-TR dataset is substantially larger and more category-diverse than DUTS-TR, and the decoder design is reasonable. However, the experimental validation as written does not establish the central claims: the main comparison table appears to evaluate a model trained on DUTS-TR rather than BDS-TR, the PASCAL-S test set may overlap with the VOC2012 portion of BDS-TR, the loss formulation for the IoU term is incorrect as written, and the reported ablations show very small differences without error bars. The strengths of the paper are the clear pipeline description and the potential value of BDS-TR, but the current evidence is insufficient.","major_comments":[{"comment":"The main SOTA comparison in Table I is not connected to the proposed BDS-TR dataset. Section IV-A says the model is trained on an undefined set called 'T-DUTS', and the 'Ours' row in Table I numerically matches the 'DUTS' row of Table III (e.g., DUT-O S .860, Fm .841, Eξ .899, M .045; ECSSD S .933, Fm .958, Eξ .937, M .028; PASCAL-S S .883, Fm .894, Eξ .889), not the 'BDS' row (e.g., DUT-O S .875, Fm .855, Eξ .910, M .043; ECSSD S .935, Fm .959, Eξ .941, M .027). Thus the headline claim that a BDS-TR-trained model outperforms SOTA is unsupported by the main table. Please define T-DUTS, state explicitly which training set produced Table I, and provide a five-benchmark table for the BDS-TR-trained model.","section":"Section IV-A and Table I vs. Table III"},{"comment":"Test independence is not established for PASCAL-S. Section III-B says BDS-TR is constructed by selecting images from COCO, OpenImages, and VOC2012, while Section IV-B evaluates on PASCAL-S, which is a subset of PASCAL VOC images. The manuscript never reports whether the VOC2012 images used in BDS-TR overlap with PASCAL-S test images. If they do, the PASCAL-S metrics in Tables I and III are not a valid generalization test. Please provide an overlap/de-duplication analysis between BDS-TR and all evaluation sets, or remove any overlapping images from training.","section":"Section III-B and Section IV-B"},{"comment":"The IoU loss is formulated incorrectly. Eq. (3) defines LIoU as the IoU ratio (intersection over union), and Eq. (4) adds this term to the BCE losses for minimization. Minimizing a positive IoU ratio would drive the prediction to have lower overlap with the ground truth, which is the opposite of the intended behavior. The standard formulation is 1 - IoU (or equivalently -IoU if maximizing). Please correct the equation or clarify the actual loss used in the experiments.","section":"Section III-D, Eq. (3)"},{"comment":"The comparison with fully supervised methods is confounded by backbone and training-data scale. The proposed model uses a DINO ViT-S backbone and, for the BDS-TR results, roughly 260k training images, while most fully supervised baselines in Table I use different backbone architectures and train on the 10.5k-image DUTS-TR. The claim that pseudo-labels approach manual quality because the method 'surpasses' fully supervised methods is therefore not established. Please report a same-backbone baseline trained on DUTS-TR and, where possible, a same-data-scale comparison.","section":"Section IV-C and Table I"},{"comment":"The ablations supporting the adjective-based text and the edge decoder show very small numerical differences and no variance estimates. For example, Table IV reports DUT-O S .855 vs. .860 and ECSSD S .933 vs. .933 with and without the decoder, and Table II reports PASCAL-S S .883 vs. .883 with and without adjectives. Without multiple runs, error bars, or significance tests, these differences cannot be distinguished from training noise. Please report repeated-run statistics or otherwise demonstrate that the gains are reliable.","section":"Section V, Tables II-IV"}],"minor_comments":[{"comment":"The title contains a typo: 'Distillated' should be 'Distilled'. In addition, the abstract says 'Traditional models rely on manually annotated pseudo labels'; this should presumably be 'pixel-level ground-truth labels', since pseudo-labels are the model-generated labels proposed in the paper.","section":"Title and Abstract"},{"comment":"The notation for the generated bounding box set is malformed: 'BI = blogit1 1 ∪ blogit2 2 ∪ ...' does not parse as a set of boxes. Please define bi and logiti with clear notation and specify how multiple detected boxes are merged or ranked.","section":"Section III-A"},{"comment":"The caption says the red background indicates GroundingDINO and the blue background indicates SAM, but this color coding is not apparent in the printed figure and there is no legend. Please add a clear legend or labels.","section":"Figure 3"},{"comment":"The caption and surrounding text should state explicitly that the DUTS and BDS rows use the same backbone, decoder, and all other training settings except the training dataset, so that the isolated effect of BDS-TR can be assessed.","section":"Table III"},{"comment":"The text says 'using T-DUTS as the training dataset' but T-DUTS is never defined anywhere in the manuscript. Please define it or replace it with the correct dataset name.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the BDS-TR dataset could be a valuable community resource, but the experimental section currently does not support the paper's strongest claims. I would like the editor to require the authors to clarify which model produced Table I, provide a corrected five-benchmark comparison for BDS-TR, address the PASCAL-S/VOC overlap issue, and fix the IoU loss formulation before any further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the headline comparison never actually uses the paper's new dataset. The 'Ours' row in Table I matches the DUTS-TR row in Table III, not the BDS-TR row, and Section IV-A says training used something called 'T-DUTS' that is never defined. So the central claim—that BDS-TR plus the pseudo-label pipeline produces SOTA SOD—is not tested in the main experiment. The authors instead show BDS-TR's benefit only in a three-dataset ablation.\n\nThat is a real problem, but the paper is not worthless. The BDS-TR dataset itself is a genuine contribution: 260k images, 960 categories, built by fine-tuning BLIP on a small text subset, then using GroundingDINO and SAM to produce masks. The recipe is plausible and cheap, and the category diversity looks like a real improvement over DUTS-TR. The ablations show consistent gains on DUT-O and PASCAL-S when switching to BDS-TR, and the edge decoder gives a modest boost. That is enough to take the dataset idea seriously.\n\nNow the soft spots. The training-set mismatch is the big one; without fixing it, the SOTA claim is unsupported. Second, BDS-TR is sourced partly from VOC2012, and PASCAL-S is VOC-derived; the paper never checks or removes overlap. That could inflate one of the five benchmarks. Third, comparisons to fully supervised methods use a different backbone and vastly more training data, so those wins are not clean. No code, data, or error bars either.\n\nWho should read this? People working on weakly supervised SOD and large-scale pseudo-labeling. The dataset could be a useful resource if released with decontaminated splits. A serious referee should engage with it, because the idea is sound and the main flaw is fixable by simply training the main model on BDS-TR and reporting those numbers. But as it stands, the paper is not ready; the authors need to reconcile Table I with Table III.\n\nRecommendation: send to peer review rather than desk reject, with the expectation of major revision. If the numbers cannot be reproduced after the fix, reject.","headline":"The headline comparison never actually uses the paper's own dataset; the 'Ours' row matches the DUTS-TR ablation, so the central claim is untested.","tokens_in":15741,"tokens_out":2664,"would_cite":false,"duration_ms":24519,"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":"The paper claims that pseudo-labels generated by a BLIP–GroundingDINO–SAM pipeline guided by fine-tuned text prompts are accurate enough that a weakly supervised salient-object-detection model trained on the resulting 260,000-image BDS-TR…","keywords":["salient object detection","weakly supervised learning","text-guided pseudo-labeling","large foundation models","BDS-TR dataset","edge-preserving decoder","dynamic upsampling","knowledge distillation"],"falsifier":"Compute the mean IoU between the SAM-generated pseudo-labels and fresh human pixel masks on a random sample of BDS-TR images; if it is far below the level implied by beating fully supervised models, the near-manual-quality claim fails, and a check of image overlap with PASCAL-S would further test whether the benchmark improvements reflect training-set leakage rather than generalization.","tokens_in":14539,"feed_emoji":"🎯","tokens_out":7790,"duration_ms":70677,"temperature":0.7,"pith_summary":"The paper tries to establish that the costly pixel-by-pixel manual labeling of salient-object-detection training data can be replaced by pseudo-labels generated from large foundation models guided by short text prompts. Its pipeline fine-tunes BLIP on a small set of manually written noun-phrase captions, uses GroundingDINO to turn those captions into bounding boxes, and lets SAM segment the boxes into masks. From this it builds BDS-TR, a training set of roughly 260,000 images spanning about 960 categories, and pairs it with an edge-preserving dynamic-upsampling decoder. On ECSSD, DUT-O, PASCAL-S, DUTS-TE, and HKU-IS the trained model reports higher scores than previous weakly supervised methods and matches or exceeds several fully supervised ones. If the result holds, high-quality SOD training data no longer requires expensive manual annotation.","feed_headline":"Weakly supervised saliency now rivals fully supervised models","feed_subtitle":"A text-to-mask pipeline turns short captions into 260,000 pseudo-labels that match manual quality on five benchmarks.","key_machinery":"The load-bearing machinery is the text-to-mask annotation pipeline: a fine-tuned BLIP captioner, GroundingDINO as the box proposer, and SAM as the mask generator, together with the new BDS-TR dataset (roughly 260,000 images across about 960 categories) and the DEDecoder, an edge-preserving decoder that uses content-aware dynamic upsampling to restore resolution and a residual channel attention block to suppress non-edge information. The structured prompt format \"adjective + noun\" is what lets the pipeline separate the salient object from equally salient-looking distractors. Training uses binary cross-entropy, partial cross-entropy, and IoU losses with equal weights, and the edge supervision comes from Canny edges computed on the pseudo-labels.","core_discovery":"The central claim is that knowledge already stored in large multimodal models can be distilled into an SOD trainer through language: a fine-tuned captioner supplies \"adjective + noun\" descriptions of the salient object, an open-vocabulary detector grounds those descriptions in bounding boxes, and a segmenter converts the boxes into precise masks. The paper further claims that scaling this recipe to about 260,000 images, with roughly 960 categories and over 3,000 subcategories drawn from COCO, OpenImages, and VOC2012, raises generalization enough that a weakly supervised model trained on these pseudo-labels outperforms existing weakly supervised methods and also surpasses several fully supervised methods on all five test sets. The quality of the generated pseudo-labels, not just their quantity, is asserted to be the reason: the same model trained on the older DUTS-TR images with the same pipeline scores lower, and adjective-free prompts produce visibly noisier masks. A dynamic edge-preserving decoder trained with Canny-based edge supervision from these clean labels is presented as the architectural complement that keeps boundaries sharp while feature resolution is restored.","pith_inferences":["If the pipeline transfers, text-prompt annotation could lower the cost of building saliency datasets enough to make open-vocabulary, long-tail, and video saliency training sets practical; the paper does not test those settings.","The comparison against fully supervised models mixes two variables at once, pseudo-label fidelity and BDS-TR's much larger scale, so a size-matched experiment using the same 260,000 images with human masks versus pseudo-labels would isolate how much of the gain is label quality rather than data volume.","Because PASCAL-S is derived from PASCAL VOC and BDS-TR includes VOC2012 images, the paper leaves an overlap check unstated; until that separation is verified, the PASCAL-S results should be read as conditional.","Language-mediated saliency also raises a subtle bias question: the captioner, not the image, decides what counts as salient, so visually prominent but rarely described objects may be systematically missed; a caption-free visual prompt baseline would test this."],"forward_implications":["A SOD model can be trained from short text prompts and foundation-model masks, cutting annotation effort from pixel-level drawing to a few percent of textual captions, plus a light manual review.","Training on BDS-TR's 260,000 images and 960 categories generalizes better than training on DUTS-TR under the same pipeline, with the largest gains on the harder datasets DUT-O and PASCAL-S.","The same pseudo-labels that drive overall detection also support edge supervision, so the decoder recovers boundaries instead of smoothing them away.","The structured \"adjective + noun\" prompt format is a real ingredient: dropping adjectives degrades pseudo-label precision and downstream scores.","The BDS-TR dataset can serve as a large, category-diverse foundation set for future SOD research, not only for this model."],"supporting_citations":[{"why":"Supplies the captioning model that is fine-tuned on manually annotated text to produce salient-object descriptions.","marker":"[24]"},{"why":"Grounds the generated text in object bounding boxes, bridging language and image location.","marker":"[25]"},{"why":"Segments the detected boxes into the final pixel-level pseudo-labels.","marker":"[26]"},{"why":"Provides DUTS-TR, the seed images and manual text subset used for fine-tuning, and defines the baseline training-set scale.","marker":"[16]"},{"why":"Source of images selected for the expanded BDS-TR training set.","marker":"[31]"},{"why":"Additional source of images for BDS-TR, contributing scale and category diversity.","marker":"[32]"},{"why":"Third image source for BDS-TR; also creates the potential overlap with PASCAL-S that the paper does not check.","marker":"[33]"},{"why":"Supplies the dynamic upsampling idea on which the DEDecoder's progressive resolution recovery is based.","marker":"[34]"},{"why":"Provides the self-supervised DINO ViT backbone the SOD network is built on.","marker":"[51]"}],"fun_headline_variants":["Weak supervision surpasses full supervision in saliency","Text prompts yield 260k saliency masks rivaling manual labels","Caption-driven distillation boosts weak saliency to strong levels","Pseudo-labels from captions outmatch fully supervised saliency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The BDS-TR training images selected from COCO, OpenImages, and VOC2012 are assumed to have no overlap with the five evaluation benchmarks, especially PASCAL-S, whose test images come from PASCAL VOC; the paper never states or verifies this image-level separation, so if VOC2012 training images reappear in PASCAL-S, the reported gains there would not measure generalization.","fun_headline_variants_meta":{"raw":{"variants":["Weak supervision surpasses full supervision in saliency","Text prompts yield 260k saliency masks rivaling manual labels","Caption-driven distillation boosts weak saliency to strong levels","Pseudo-labels from captions outmatch fully supervised saliency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1955,"prompt_tokens":997,"completion_tokens":958,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":890}},"tokens_in":613,"tokens_out":958,"duration_ms":9406,"temperature":1.0,"reasoning_tokens":890,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:29:29.792269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the mean IoU between the SAM-generated pseudo-labels and fresh human pixel masks on a random sample of BDS-TR images; if it is far below the level implied by beating fully supervised models, the near-manual-quality claim fails, and a check of image overlap with PASCAL-S would further test whether the benchmark improvements reflect training-set leakage rather than generalization.","supporting_citations":[{"cited_title":"Learning to detect salient objects with image-level supervision,","cited_arxiv_id":null,"evidence_quote":"Provides DUTS-TR, the seed images and manual text subset used for fine-tuning, and defines the baseline training-set scale."},{"cited_title":"Learning to upsample by learning to sample,","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic upsampling idea on which the DEDecoder's progressive resolution recovery is based."},{"cited_title":"Emerging properties in self-supervised vision transformers,","cited_arxiv_id":null,"evidence_quote":"Provides the self-supervised DINO ViT backbone the SOD network is built on."}],"review_version":1}