{"id":"b3b3b4d8-8125-4118-a962-2d6973d71404","arxiv_id":"1908.07274","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new high-resolution salient object detection dataset plus a global-local fusion network that outperforms prior methods on high-resolution images.","lead":"This paper introduces HRSOD, a new dataset of 2,010 high-resolution images with pixel-level saliency annotations, and a three-branch network that combines global semantic cues, attention-based local refinement, and fusion to detect salient objects in images above 1200 pixels. It addresses a practical gap because most saliency models are trained and tested only on low-resolution images, limiting their usefulness in image editing and 4K pipelines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'large margin' on high-resolution benchmarks is not yet disentangled from an evaluation asymmetry: baselines are only tested with low-resolution inputs and no HRSOD training, so the gap may reflect pixels seen or training data, not the GSN-LRN-GLFN design.","rationale":"I read the paper's strongest claim as a comparison claim: the proposed GSN-LRN-GLFN framework outperforms prior art by a large margin on high-resolution saliency while staying competitive on low-resolution benchmarks. The reader's evaluation-fairness concern is real, but the Ours-D row already shows a large gain over all prior methods without training on HRSOD, so 'baselines were not trained on HRSOD' cannot by itself explain the headline margin. The more damaging uncontrolled factor is that no baseline is given high-resolution inputs at test time: prior methods are compared at 384×384 while Ours uses a 384×384 global input plus original-resolution/local patches and a 1024×1024 fusion network. This conflates the proposed architecture with the benefit of higher input resolution. Concretely, a dense/overlapping patch-based version of DGF or DGRL, using official weights only, would test whether the extra margin comes from seeing high-resolution pixels. If it does not close the gap, the paper's case is strong; if it does, the verdict should move toward rejection or at least major revision. Because this test has not been run, the current CONDITIONAL verdict is appropriate; no change to the reader's decision is needed.","tokens_in":13101,"tokens_out":9016,"duration_ms":98831,"concrete_test":"Run the strongest baseline (DGF or DGRL) with its official weights in an overlapping patch-based high-resolution inference mode on HRSOD-Test: crop 384×384 patches with 50% overlap, average the predicted maps in overlapping regions, and compute Fβ/MAE. As a second arm, fine-tune the same baseline on DUTS-TR plus HRSOD-Train. If either arm reaches Fβ close to Ours-D (0.857) or Ours-DH (0.888), the claimed margin is substantially a resolution/data advantage rather than evidence for the proposed framework; if both stay near the original 0.79–0.80, the architecture claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 1, where Ours-DH reaches Fβ 0.888 on HRSOD-Test versus 0.795 for the strongest prior (DGF). The decisive question is whether that margin is caused by the proposed framework or by the comparison protocol. Every prior method is evaluated with its standard low-resolution pipeline: existing models cannot directly ingest high-resolution images, and Sections 5.1.3 and 5.2 report no baseline that is fine-tuned on HRSOD or given a high-resolution test-time pathway. Part of the training-data confound is mitigated by the Ours-D row, which trains only on DUTS and still achieves 0.857, so the architecture is doing something. However, no control baseline sees high-resolution pixels at test time. Section 1 itself lists 'partitioning inputs into patches' as the intuitive alternative, yet no patch-based version of a prior method is evaluated. Without that control, the large margin could reflect the simple benefit of processing original high-resolution content rather than the specific GSN, APS, LRN, or GLFN contributions. The paper's own ablations compare only variants inside its framework, not prior methods equipped with a comparable high-resolution mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper tackles salient object detection in very high-resolution images, a regime where existing FCN-based methods downscale inputs and lose boundary details. The authors contribute a new dataset, HRSOD (1610 training and 400 test images, shortest edge over 1200 pixels, with careful pixel-level annotation), and propose a multi-branch framework: a Global Semantic Network (GSN) produces a coarse saliency map from a 384x384 downscaled image; an Attended Patch Sampling (APS) scheme selects patches around uncertain regions; a Local Refinement Network (LRN) refines those patches under semantic guidance from GSN; and a lightweight Global-Local Fusion Network (GLFN) fuses the coarse and refined predictions using the original high-resolution input. Experiments compare the method with nine prior saliency detectors on HRSOD-Test, a high-resolution subset of DAVIS (DAVIS-S), and three standard low-resolution benchmarks (DUTS-Test, HKU-IS, THUR), reporting F-beta, S-measure, MAE, and Boundary Displacement Error. Two versions are evaluated: Ours-D (trained on DUTS only) and Ours-DH (trained on DUTS + HRSOD). Ablation studies compare the full model against GSN alone, GSN+CRF, GSN+RPS+LRN, and GSN+APS+LRN+CRF.","tokens_in":13271,"tokens_out":3458,"duration_ms":37057,"significance":"If the results hold, the paper makes a timely contribution by defining a high-resolution saliency benchmark and showing that a modular global-local framework can be trained end-to-end on high-resolution images with modest GPU memory. The HRSOD dataset fills a real gap: existing benchmarks cap image sizes around 400-500 pixels, which is unrealistic for modern camera outputs. The APS uncertainty-guided patch sampling is a practical idea, and the ablations (APS vs RPS, GLFN vs CRF) give evidence that the specific design choices matter. The authors also report BDE on high-resolution data, which is more sensitive to boundary quality than standard metrics. The central accuracy claim, however, is weakened by an evaluation asymmetry: baselines are not trained on HRSOD and do not receive high-resolution test-time inputs, so the large margin on HRSOD-Test (F-beta 0.888 vs 0.795) is not cleanly attributable to the proposed architecture. The strengths of the paper are the dataset, the modular framework, and the generally careful ablation; the weakness is the comparison protocol for the headline claim.","major_comments":[{"comment":"The central claim of a \"large margin\" on high-resolution datasets is not fully disentangled from an evaluation asymmetry. Every baseline (RFCN, DHS, UCF, Amulet, NLDF, DSS, RAS, DGRL, DGF) is evaluated with its standard low-resolution pipeline and is not fine-tuned on HRSOD, while Ours-DH is trained on HRSOD training images. The Ours-D row (trained only on DUTS) mitigates part of this concern, but even Ours-D sees high-resolution pixels at train and test time, whereas no baseline is given a high-resolution pathway (either fine-tuning on HRSOD or a patch-based high-resolution inference mode). To support the claim that the margin comes from the GSN-LRN-GLFN design rather than merely from training data scale or input resolution, the authors should include at least one control: for example, fine-tune a strong baseline (e.g., DGRL or DGF) on the HRSOD training set, or apply a patch-based version of a baseline such that it also processes original high-resolution content, and compare on HRSOD-Test. If such a control is infeasible, the text should explicitly temper the \"large margin\" claim.","section":null},{"comment":"The paper reports a single run of each method and gives no error bars, confidence intervals, or significance tests. On HRSOD-Test the margins are large, but on low-resolution benchmarks several differences are small (e.g., HKU-IS F-beta 0.886 for Ours-DH vs 0.895 for DSS; DUTS-Test F-beta 0.791 vs 0.791 for DSS); without repeated runs or significance testing it is unclear whether these differences are meaningful. The authors should report variance over at least three training runs (or, if this is prohibitive, at least for the main model and the closest baseline) or use significance tests such as paired bootstrap over images.","section":"§5.2, Table 1"},{"comment":"The claim that the comparison is \"fair\" in §5.2 is not substantiated for the high-resolution setting. The compared methods' official implementations or provided saliency maps are used with recommended settings, which means they are optimized for low-resolution benchmarks. This is fine for standard benchmarks, but on HRSOD-Test and DAVIS-S it conflates task adaptation with method quality. The paper should either provide adapted baselines or explicitly frame the comparison as \"off-the-shelf methods vs. a method trained for high resolution,\" which is a weaker claim than the current wording.","section":"§5.2"}],"minor_comments":[{"comment":"The statement \"our method is the fasted among all the compared methods\" is inaccurate: in Table 3, DGF and UCF also run at 0.05 s at 384x384, and Ours* (0.39 s at 1024x1024) is slower than several 384x384 baselines. The time comparison mixes input sizes; it should be clearly stated that Ours* uses a different (higher) input resolution and the claim should be restricted accordingly.","section":"§5.3.2, Table 3"},{"comment":"The hyper-parameters T1, T2, D, n, and r are said to be set by grid search with \"results not sensitive to their specific choices,\" but no sensitivity analysis is shown (e.g., a plot of F-beta or MAE vs. each parameter). Adding this analysis, even in the supplementary material, would support the claim of robustness.","section":"§4.3"},{"comment":"The GLFN model size is reported as 11.9 kB, which is unusually small for a network with conv layers and dense connectivity; it should be clarified whether this is parameter storage after quantization, a compressed format, or a misprint, and how the size is computed.","section":"§4.4"},{"comment":"The text in §5.2 says \"Figure 8 shows a visual comparison of our method with respect to others,\" but the quantitative visual comparison appears to be Figure 7 (labeled \"Visual comparison\"); §5.3.1 uses Figure 8 for the CRF comparison. The figure numbering is inconsistent and should be fixed.","section":"Figures 7 and 8"},{"comment":"The paper states \"Source code will be released\" but does not provide a public repository at the time of this review; given the paper's reliance on a new dataset and a multi-stage training protocol, making the code available is important for reproducibility and should be stated with an actual link.","section":"§5.1.3"},{"comment":"The BDE is evaluated only on HRSOD-Test and DAVIS-S, with the justification that other benchmarks have poor boundary annotations. This is reasonable, but the paper does not report whether BDE values are stable across thresholds or boundary extraction methods; a brief explanation of the boundary pixel extraction would help.","section":"§5.3.2"},{"comment":"Minor language issue: \"the above defect becomes more severe\" and several other informal phrasings (e.g., \"very high-resolution\") could be tightened; also 'fasted' in §5.3.2 should be 'fastest'.","section":"Abstract / §1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation asymmetry in Table 1 is the main correctness risk. A reader who takes the headline claim at face value could over-attribute the margin to the architecture. The Ours-D row already shows the framework is strong even without HRSOD training, but a baseline fine-tuned on HRSOD is the missing control. The editor may also want to ask the authors to clarify the relationship of HRSOD to any concurrent high-resolution saliency datasets, since the first-dataset claim is likely to be scrutinized. The dataset itself is a valuable resource and the method is plausibly effective; the revision should strengthen rather than reject the work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the dataset. HRSOD is the first high-resolution saliency benchmark, with more than 2000 images, shortest edge over 1200 pixels, and careful annotation. That alone is worth having. The method is a sensible stack: a global net on downscaled input, attention-guided patch sampling for uncertain regions, a local refinement net with global guidance, and a lightweight fusion network. The ablations show each piece adds something, and the comparison against CRF is a fair way to argue that learned refinement beats post-processing. The DAVIS-S results give independent support: even Ours-D, trained only on DUTS, beats the best prior by a clear margin on a high-resolution set it never trained on.\n\nThe soft spot is exactly what the stress-test note says. Table 1 compares against prior methods run with their standard low-resolution pipelines. None are fine-tuned on HRSOD, and none are given a high-resolution pathway at test time. So the large margin bundles together high-resolution processing, training data, and architecture. The Ours-D row helps separate training data from architecture, and the internal ablations show the components matter, but there is still no control where, say, DGRL or DGF is fine-tuned on HRSOD or fed a patch-based pipeline. The paper itself lists patch-based processing as the intuitive alternative, yet no prior method is evaluated that way. That is a real gap, but it does not sink the paper. The central claim that processing high-resolution content and refining uncertain regions helps is supported. The larger claim that this specific framework beats SOTA by a large margin is not yet cleanly established.\n\nMinor issues: no error bars or significance tests; code is promised but not released; the hyperparameters are chosen empirically and reported as insensitive, which is fine but a sensitivity table would be better. The running-time claim is a bit loose: Ours* at 0.39s is not actually the fastest in Table 3. Citation pattern looks honest; they cite the patch-based and global-local methods they build on.\n\nWho is this for? Saliency researchers and anyone doing high-resolution dense prediction will want the dataset. The method is a reasonable baseline. It deserves peer review, but the referee should ask for a fair high-resolution evaluation of prior methods, ideally with code and error bars.","headline":"The HRSOD dataset is a genuine contribution and the high-res processing story is plausible, but the headline 'large margin' over prior methods is confounded by an evaluation protocol that gives baselines neither high-resolution inputs nor HRSOD training.","tokens_in":13867,"tokens_out":2607,"would_cite":true,"duration_ms":28498,"reading_group":"yes","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 introduces the first high-resolution saliency dataset, HRSOD, and a global-local network that lifts Fβ on its test set from 0.795 (best prior) to 0.888 while keeping pace on standard low-resolution benchmarks.","keywords":["salient object detection","high-resolution images","HRSOD dataset","global-local network","attended patch sampling","boundary refinement","dilated convolution","saliency benchmark"],"falsifier":"Retrain a strong baseline such as DGRL or DGF on the HRSOD training set (or on DUTS+HRSOD) using its own training protocol, then evaluate it on HRSOD-Test with the same metrics; if its Fβ approaches or exceeds 0.888, the paper's architectural claim is not the source of the large margin.","tokens_in":12827,"feed_emoji":"🖼️","tokens_out":7626,"duration_ms":62558,"temperature":0.7,"pith_summary":"This paper tackles salient object detection on very high-resolution images (shortest edge over 1200 pixels), where existing deep saliency models typically down-sample inputs and lose boundary detail. It introduces HRSOD, the first high-resolution saliency dataset, with 1610 training and 400 test images, and a three-part framework: a Global Semantic Network reads a down-sampled whole image, a Local Refinement Network crops uncertain regions at full resolution guided by the global prediction, and a lightweight Global-Local Fusion Network merges the two into a spatially consistent mask. On HRSOD-Test the method reports an Fβ score (a precision-recall harmonic mean weighted toward precision) of 0.888 versus 0.795 for the best prior baseline, with lower boundary displacement error, while staying competitive on standard low-resolution benchmarks. If the result holds, high-resolution saliency maps become accurate enough for downstream tasks such as image editing and background replacement, without post-processing or excessive GPU memory.","feed_headline":"Saliency score jumps from 0.795 to 0.888 on first high-res benchmark","feed_subtitle":"A new dataset and a global-local network make boundary-accurate saliency maps practical for images above 1200 pixels.","key_machinery":"The load-bearing mechanism is coarse-to-fine refinement with uncertainty-driven attention. Attended Patch Sampling (APS) takes the coarse global saliency map from GSN, marks pixels whose predicted saliency falls between thresholds T1=50 and T2=200 as uncertain, and crops sub-images centered on those pixels; the Local Refinement Network then sees those patches at full local resolution, with the corresponding GSN map patch concatenated into its decoder as global guidance. A lightweight Global-Local Fusion Network, built from densely connected layers and dilated convolutions (which enlarge receptive field without down-sampling) and with only 11.9 kB of parameters, takes the fused GSN/LRN output together with the original high-resolution image and produces the final spatially consistent mask. This hierarchy is what lets the method keep global context at low resolution while spending computation only where boundaries and ambiguous regions need detail.","core_discovery":"On its own terms, the paper claims two contributions. First, HRSOD is the first saliency detection dataset whose images have a shortest edge above 1200 pixels, with 1610 training and 400 test images and pixel-level annotations checked by multiple subjects, designed to expose boundary detail that low-resolution datasets lack. Second, the proposed GSN-LRN-GLFN framework combines a global semantic network that reads a 384×384 down-sampled image, a local refinement network that crops attended uncertain sub-regions at higher resolution, and a global-local fusion network that merges both into a spatially consistent prediction. The authors report Ours-DH reaching Fβ 0.888, S-measure 0.897, and MAE 0.030 on HRSOD-Test, against best prior values of 0.795, 0.853, and 0.053, and a similar lead on DAVIS-S, while remaining within a few points of the best methods on THUR, HKU-IS, and DUTS-Test.","pith_inferences":["Because Ours-DH is trained on HRSOD while none of the nine baselines is, part of the large margin may come from seeing high-resolution training data; fine-tuning a baseline such as DGRL or DGF on HRSOD would tell how much remains architectural.","APS is a general answer to the memory-versus-detail conflict: any dense prediction task with limited GPU memory could crop only the uncertain regions, guided by a cheap global pass, rather than processing the whole image at full resolution.","If the margin persists under fair retraining, the practical implication is that saliency in consumer images can be computed at full resolution without heavy CRF post-processing, simplifying pipelines for background replacement and depth-of-field effects."],"forward_implications":["High-resolution saliency detection becomes tractable on a single 1080 Ti: GSN and LRN run at 384×384 and GLFN at 1024×1024, with reported 0.39 s per 1024×1024 image.","Boundary accuracy improves substantially: BDE drops from 32.91 for the weakest baseline to 17.57 on HRSOD-Test, putting the method in range for downstream editing and compositing applications.","The framework is backbone-agnostic; GSN and LRN can be swapped for any saliency model, so future improvements in low-resolution saliency should transfer directly to high-resolution settings.","The HRSOD dataset gives the community a training and evaluation target where boundary detail and annotation quality, not just global overlap scores, are the differentiators."],"supporting_citations":[{"why":"Pretrained VGG-16 supplies the backbone for both GSN and LRN, initialized from ImageNet weights.","marker":"[30]"},{"why":"Dense connectivity in DenseNet shapes the lightweight GLFN architecture used for fusion.","marker":"[15]"},{"why":"DUTS provides the low-resolution training set used for Ours-D and as part of the Ours-DH training mixture.","marker":"[33]"},{"why":"Dense CRF is the post-processing baseline that LRN and GLFN are shown to outperform in the ablation study.","marker":"[19]"},{"why":"DSS is both a compared baseline and the source of the CRF parameter setting used in the ablation.","marker":"[14]"},{"why":"DAVIS is the source of the collected DAVIS-S high-resolution test set.","marker":"[27]"},{"why":"HKU-IS is one of the standard low-resolution benchmarks and also illustrates the resolution gap motivating HRSOD.","marker":"[21]"},{"why":"Structure-measure is one of the evaluation metrics used to compare methods.","marker":"[9]"}],"fun_headline_variants":["New dataset and network boost high-res saliency detection","First high-res saliency dataset: Fβ 0.888 vs 0.795","Global-local fusion lifts saliency accuracy on high-res images","High-resolution saliency: new benchmark and a 12% Fβ gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the nine comparison methods, none of which was trained on high-resolution saliency data, are a fair baseline for a method that was; if that training-data gap were removed, the claimed large margin might shrink.","fun_headline_variants_meta":{"raw":{"variants":["New dataset and network boost high-res saliency detection","First high-res saliency dataset: Fβ 0.888 vs 0.795","Global-local fusion lifts saliency accuracy on high-res images","High-resolution saliency: new benchmark and a 12% Fβ gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1543,"prompt_tokens":1018,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":634,"tokens_out":525,"duration_ms":5771,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:20:30.872624+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain a strong baseline such as DGRL or DGF on the HRSOD training set (or on DUTS+HRSOD) using its own training protocol, then evaluate it on HRSOD-Test with the same metrics; if its Fβ approaches or exceeds 0.888, the paper's architectural claim is not the source of the large margin.","supporting_citations":[{"cited_title":"Very deep convo- lutional networks for large-scale image recognition","cited_arxiv_id":null,"evidence_quote":"Pretrained VGG-16 supplies the backbone for both GSN and LRN, initialized from ImageNet weights."},{"cited_title":"Densely connected convolutional net- works","cited_arxiv_id":null,"evidence_quote":"Dense connectivity in DenseNet shapes the lightweight GLFN architecture used for fusion."},{"cited_title":"Learning to de- tect salient objects with image-level supervision","cited_arxiv_id":null,"evidence_quote":"DUTS provides the low-resolution training set used for Ours-D and as part of the Ours-DH training mixture."},{"cited_title":"Efficient inference in fully connected crfs with gaussian edge potentials","cited_arxiv_id":null,"evidence_quote":"Dense CRF is the post-processing baseline that LRN and GLFN are shown to outperform in the ablation study."},{"cited_title":"Deeply supervised salient ob- ject detection with short connections","cited_arxiv_id":null,"evidence_quote":"DSS is both a compared baseline and the source of the CRF parameter setting used in the ablation."},{"cited_title":"A benchmark dataset and evaluation methodology for video object segmentation","cited_arxiv_id":null,"evidence_quote":"DAVIS is the source of the collected DAVIS-S high-resolution test set."},{"cited_title":"Visual saliency based on multi- scale deep features","cited_arxiv_id":null,"evidence_quote":"HKU-IS is one of the standard low-resolution benchmarks and also illustrates the resolution gap motivating HRSOD."},{"cited_title":"Structure-measure: A New Way to Evaluate Fore- ground Maps","cited_arxiv_id":null,"evidence_quote":"Structure-measure is one of the evaluation metrics used to compare methods."}],"review_version":1}