{"id":"93e83cdd-60f7-4cd4-bffe-56dc63cac908","arxiv_id":"1908.01975","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A boundary-weighted loss and a hierarchical global attention module improve salient object segmentation on six benchmarks.","lead":"This paper introduces a new way to train a deep network for finding the most eye-catching object in an image, adding extra emphasis on object edges and a global attention step. The method reports the best published accuracy on six benchmarks while running at 26 frames per second.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of boundary-aware improvement is not directly tested: only global Fβ/MAE are reported, leaving open that Contour Loss gains come from non-boundary effects. A boundary-specific metric would settle it.","rationale":"The reader's weakest assumption concerned the specific K=5 spatial weighting in Eq. 4. I agree partially: the exact value of K is not the main risk; what is untested is whether a static boundary weighting actually translates into improved boundary accuracy. The paper reports only global Fβ and MAE, so the core 'Contour Loss' contribution could be helping for reasons unrelated to boundaries. A boundary-specific metric directly tests the stated mechanism. Because the reader already returned a conditional verdict, and this concern strengthens rather than overturns that verdict, the final recommendation remains UNCHANGED.","tokens_in":11255,"tokens_out":4359,"duration_ms":49488,"concrete_test":"On DUTS-TE and DUT-O, recompute the Table 2 comparison (B vs B+C vs B+H vs B+C+H) using a boundary-focused metric: standard boundary F-score (Fb) or F-measure within a 5-pixel band around ground-truth contours, plus MAE restricted to that boundary band. If B+C does not improve the boundary metric over B by a margin comparable to the reported global Fβ gain, the central 'boundary-aware' claim is unsupported. If it does, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Contour Loss makes the network 'boundary-aware' and that this drives the reported improvements. The evidence, however, is indirect. All quantitative comparisons (Table 1 and the ablation in Table 2) report whole-image maximum Fβ and MAE. These metrics can improve when interior pixels are corrected or when the predicted object is slightly enlarged or shifted; they do not isolate boundary accuracy. Equation 4 is a static, dataset-independent weighting of ground-truth boundary bands: the weight map is computed from dilated and eroded masks with K=5 and is never adapted to prediction difficulty. If the observed gains on DUTS-TE and DUT-O come primarily from HGAM or from global map statistics rather than from improved localization of contour pixels, the core contribution 'Contour Loss' would not be supported. The absence of any boundary-specific evaluation (e.g., boundary F-measure, trimap F-measure, or contour precision/recall), combined with no released code and no statistical significance testing, leaves this possibility open. This is load-bearing because the method's name and framing rest on boundary awareness; global Fβ is compatible with non-boundary explanations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two components for salient object segmentation: Contour Loss, a reweighted cross-entropy loss that up-weights pixels near salient-object boundaries using a spatial weight map derived from dilated/eroded ground-truth masks, and a hierarchical global attention module (HGAM) that captures global context at multiple resolutions. The architecture is an FPN-style VGG-16 network with residual decoder blocks; the final output is generated from the guided first-level feature. The method is evaluated on six benchmarks against 13 prior methods using maximum F-beta and MAE, and an ablation on DUTS-TE and DUT-O compares the baseline with Contour Loss alone, HGAM alone, and their combination. The authors report consistent improvements and a real-time inference speed of 26 fps.","tokens_in":11480,"tokens_out":4806,"duration_ms":51126,"significance":"If the claims hold, the paper makes a modest but useful contribution: a simple boundary-weighting loss and a global-context attention module that together improve salient object segmentation over VGG-16-based competitors. The strengths are the clear formulation of the two components, the breadth of the benchmark comparison, the inclusion of PR curves and visualizations, and the explicit ablation study. The central boundary-aware claim, however, is not directly tested: only whole-image F-beta and MAE are reported, and the gains over the closest baselines are small and shown without repeated runs or significance testing. The reader's concern that the Contour Loss benefit might come from non-boundary effects therefore lands, and it is the main load-bearing issue for the paper's framing.","major_comments":[{"comment":"The central claim that Contour Loss makes the network 'boundary-aware' is not directly tested. Table 2 reports only whole-image maximum F-beta and MAE, and Table 1 likewise uses only these global metrics. These metrics can improve from better interior labeling or from a slight shift in the predicted object extent, rather than from more accurate contour localization. Since the loss is named and motivated by boundary perception, the paper should report a boundary-specific metric (for example boundary F-measure, trimap F-measure, or contour precision/recall) on at least DUTS-TE and DUT-O. Without such a metric, the comparison B+C versus B in the ablation does not establish that the improvement is due to improved boundary awareness.","section":"§3.2, §4.3, Table 2"},{"comment":"The reported improvements are small and are presented without error bars or repeated runs. For instance, on DUTS-TE the maximum F-beta values are 0.848 for B, 0.861 for B+C, 0.860 for B+H, and 0.872 for the full model; the corresponding DUT-O values are 0.787, 0.806, 0.801, and 0.825. Differences of 0.001 to 0.005 between B+C and B+H are within typical run-to-run variation for a single training run. The authors should provide mean and standard deviation over multiple runs, or a statistical significance test, to justify the claim that each component contributes independently and that the combined gain is reliable.","section":"§4.3, Table 2"},{"comment":"The Contour Loss depends on the empirically chosen K=5 and a fixed 5x5 dilation/erosion mask, while HGAM depends on lambda=0.1 in Eq. (6). No sensitivity analysis is given for these hyperparameters. Because the loss's effect is central to the paper and the reported gains are small, the authors should show that the conclusions are stable over a reasonable range of K, or describe a validation-based selection procedure for K and lambda. Without this, the transferability of the boundary-weighting scheme to other datasets or settings remains unclear.","section":"§3.2, Eq. (4), §4.1"}],"minor_comments":[{"comment":"The text spells 'Gaussian' as 'Guass' in Eq. (4); also define the structuring element S and the border handling used in the dilation and erosion operations.","section":"§3.2, Eq. (4)"},{"comment":"The caption lists 'DRFI[5]' but the reference list and Table 1 use DRFI[9]; the citation should be corrected.","section":"Fig. 7 caption"},{"comment":"The heading 'HGAM Visualizaiton' contains a typo and should read 'Visualization'.","section":"§4.4"},{"comment":"The caption refers to 'guidedRes1' before the notation ResG_i is defined in Eq. (8); please unify the notation for clarity.","section":"Fig. 2 caption"},{"comment":"The compared methods use different backbones (VGG-16, VGG-19, ResNet-50); the 'state-of-the-art' claim should be qualified by noting which comparisons are most direct, and the reader should be pointed to the closest VGG-16 competitors.","section":"Table 1"},{"comment":"The paper does not mention code or model release; a reproducibility statement or a promise of releasing the code would strengthen the contribution.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the proposed components are reasonable, but the central boundary-aware claim currently rests on small, unreplicated gains measured only with global metrics. I would support publication after the authors add a boundary-specific evaluation, repeated-run statistics, and a sensitivity analysis for the key hyperparameters. No citation or ethics concerns beyond the missing code availability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest empirical paper, not a breakthrough. The new pieces are a boundary-weighted cross-entropy loss (Contour Loss) and a global-contrast attention module (HGAM), both riding on an FPN/VGG-16 baseline. On six standard saliency benchmarks the full model tops prior VGG-16 methods by 1-2 points Fβ, and runs at 26 fps. I believe the results are probably real, but the paper doesn't prove the boundary mechanism it claims.\n\nWhat's good: the ablations in Table 2 show each component helps, on two large datasets, and the gains are consistent. The visual examples, while cherry-picked, do show finer contours with Contour Loss. The design is simple enough to reproduce, and the loss could transfer to other dense-prediction tasks. The writing is clear, and the comparisons are fair on the whole.\n\nWhere I'd push back: (1) No boundary-specific metric. The paper's title and central claim are about boundaries, but all numbers are global Fβ and MAE. Those can improve from non-boundary effects like better interior filling or a slight shift of the predicted region. The authors never report boundary F-measure, trimap F, or contour precision/recall, so the mechanism is not actually confirmed. A single boundary metric would have settled it. (2) No error bars, no repeated runs, no released code. That's common in this subfield, but the gains are small enough (1-2%) that I'd like to know they're stable. The ablation is also only on DUTS-TE and DUT-O. (3) The attention map in Eq. 6 is under-specified: δ was earlier defined as conv+ReLU, but the attention map is used for multiplicative gating; if it's not bounded to [0,1] that's a problem. A typo ('Guass') and the lack of a weight for the final P loss in Eq. 9 are minor.\n\nOverall: the stress-test note is right but a bit severe. The evidence is suggestive, not conclusive. This isn't a desk-reject paper; it's a reasonable conference submission that would benefit from a boundary-aware evaluation and some robustness checks. I'd send it to review and ask for that.","headline":"A modest, plausible saliency-detection paper whose Contour Loss claim is real but not directly proven; deserves a referee, not a desk reject.","tokens_in":11991,"tokens_out":2665,"would_cite":true,"duration_ms":28809,"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 Contour Loss and a hierarchical global attention module, combined in an FPN-style VGG-16 network, outperform 13 prior methods on six salient-object-segmentation benchmarks.","keywords":["salient object segmentation","Contour Loss","boundary-aware learning","hierarchical global attention module","global context","feature pyramid network","deep supervision","VGG-16"],"falsifier":"A decisive test: retrain the full network twice on DUTS-TR, once with every Contour Loss weight fixed to 1 (plain cross-entropy) and once with HGAM connections severed; if the full model does not beat both ablated variants by roughly 0.01 to 0.02 in max $F_\\beta$ on DUTS-TE and DUT-O, the paper's mechanism is not the source of the claimed improvement.","tokens_in":11076,"feed_emoji":"🎯","tokens_out":12577,"duration_ms":109366,"temperature":0.7,"pith_summary":"This paper sets out to show that salient object segmentation improves when the training signal itself points at object boundaries and when the decoder is steered by global context at every scale. To do that, the authors introduce Contour Loss, a reweighting of binary cross-entropy that gives extra weight to pixels near edges via a dilated-minus-eroded ground-truth map, and the hierarchical global attention module (HGAM), which propagates unsupervised contrast-based attention maps down the feature pyramid. These are combined in an FPN-style network with a VGG-16 backbone. If the claim holds, the result is sharper boundaries and cleaner background suppression than 13 earlier methods across six benchmark datasets, at a speed of 26 frames per second.","feed_headline":"Boundary loss plus global attention tops six saliency benchmarks","feed_subtitle":"Contour Loss reweights edge pixels while HGAM spreads global context; the model runs at 26 fps on a single GPU.","key_machinery":"The load-bearing objects are two. The first is Contour Loss: a spatial weight map $M_C=\\mathrm{Gauss}(K\\cdot((Y;S)_+-(Y;S)_-))+1$, with $K=5$, built from the difference between dilated and eroded ground-truth masks, multiplies the per-pixel binary cross-entropy term so boundary pixels dominate training. The second is HGAM: each module concatenates pooled, encoded, and previous-message features, forms an attention map $H^{\\mathrm{Atten}}=\\delta((F_{\\mathrm{In}}-\\operatorname{Aver}(F_{\\mathrm{In}}))/\\sqrt{\\operatorname{Var}(F_{\\mathrm{In}})+\\epsilon})+\\lambda_1$, and multiplies that map onto residual decoder features, passing messages coarse-to-fine so each level sees global contrast rather than a softmax-sparse set of pixels.","core_discovery":"On the authors' own terms, the central discovery is that boundary-focused supervision and global contextual attention are complementary and jointly push saliency segmentation past earlier methods. Contour Loss alone raises the baseline's maximum $F_\\beta$ on DUTS-TE from 0.848 to 0.861, HGAM alone raises it to 0.860, and the combination reaches 0.872, with analogous MAE gains and consistent improvements on DUT-O. The paper argues that these improvements come from the two modules addressing different error types: Contour Loss makes predicted boundaries match the mask edges, while HGAM removes background distractions that softmax-based attention misses. The reported system also keeps inference at 26 fps, so the accuracy gain is not bought with extra test-time computation.","pith_inferences":["The authors leave implicit that Contour Loss is not saliency-specific; the same dilation-minus-erosion weight map could be applied to semantic segmentation or medical imaging, where boundary errors dominate.","A testable extension is to swap HGAM into other FPN-style decoders in place of softmax attention; the paper's mechanism predicts the benefit should grow with image resolution.","Because HGAM's attention maps are unsupervised and sometimes emphasize background, they may act as an implicit regularizer; that predicts HGAM helps most on smaller or noisier training sets."],"forward_implications":["Any binary segmentation task can reuse Contour Loss as a drop-in loss term, since it needs only a mask and morphological operations.","HGAM gives high-resolution dense predictors an attention mechanism that retains background information instead of collapsing onto a few pixels.","The 26 fps inference means the accuracy gains come without slowing deployment, making the approach usable in real-time systems.","The ablations imply the two modules attack different error modes, so combining boundary weighting with global-context attention is a reusable recipe rather than a dataset-specific fix."],"supporting_citations":[{"why":"Provides the focal-loss idea of reweighting hard examples, which Contour Loss adapts to spatial boundary weighting.","marker":"[18]"},{"why":"Supplies the Feature Pyramid Network baseline architecture that the proposed decoder and HGAM extend.","marker":"[17]"},{"why":"Defines the VGG-16 backbone used as the encoder and fine-tuned at reduced learning rate.","marker":"[27]"},{"why":"Supplies the deeply supervised weighted-sum loss formulation used for the hierarchical predictions.","marker":"[6]"},{"why":"Suggests the max-pool and average-pool context extraction that HGAM adopts for global features.","marker":"[33]"},{"why":"PiCANet is the attention-based baseline whose global-context behavior HGAM is designed to improve on.","marker":"[20]"},{"why":"Reverse Attention is a baseline using predicted maps as attention, which HGAM contrasts with by keeping background information.","marker":"[2]"},{"why":"Contour Knowledge Transfer is a multi-task boundary-aware baseline that motivates the boundary-loss approach and is compared against.","marker":"[14]"},{"why":"NLDF is a non-local multi-task baseline that the paper compares against and whose boundary handling motivates Contour Loss.","marker":"[22]"}],"fun_headline_variants":["Boundary-aware loss and global attention top six saliency benchmarks","Saliency segmentation: Contour Loss plus HGAM hits 26 fps","Edge loss and hierarchical attention sharpen salient object boundaries","Contour Loss and HGAM: real-time saliency with higher F-measure","Boundary-aware loss and global attention: 26 fps on six benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The boundary-aware gain rests on the assumption that the fixed weight map built from dilation, erosion, and Gaussian smoothing with $K=5$ correctly identifies the pixels whose misclassification matters most; if that proxy does not generalize, the reported improvements would shrink or vanish outside the six datasets.","fun_headline_variants_meta":{"raw":{"variants":["Boundary-aware loss and global attention top six saliency benchmarks","Saliency segmentation: Contour Loss plus HGAM hits 26 fps","Edge loss and hierarchical attention sharpen salient object boundaries","Contour Loss and HGAM: real-time saliency with higher F-measure","Boundary-aware loss and global attention: 26 fps on six benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3073,"prompt_tokens":829,"completion_tokens":2244,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":2167}},"tokens_in":445,"tokens_out":2244,"duration_ms":16998,"temperature":1.0,"reasoning_tokens":2167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:56:54.213677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: retrain the full network twice on DUTS-TR, once with every Contour Loss weight fixed to 1 (plain cross-entropy) and once with HGAM connections severed; if the full model does not beat both ablated variants by roughly 0.01 to 0.02 in max $F_\\beta$ on DUTS-TE and DUT-O, the paper's mechanism is not the source of the claimed improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the focal-loss idea of reweighting hard examples, which Contour Loss adapts to spatial boundary weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Feature Pyramid Network baseline architecture that the proposed decoder and HGAM extend."},{"cited_title":"Hou, M.-M","cited_arxiv_id":null,"evidence_quote":"Supplies the deeply supervised weighted-sum loss formulation used for the hierarchical predictions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Suggests the max-pool and average-pool context extraction that HGAM adopts for global features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PiCANet is the attention-based baseline whose global-context behavior HGAM is designed to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reverse Attention is a baseline using predicted maps as attention, which HGAM contrasts with by keeping background information."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contour Knowledge Transfer is a multi-task boundary-aware baseline that motivates the boundary-loss approach and is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NLDF is a non-local multi-task baseline that the paper compares against and whose boundary handling motivates Contour Loss."}],"review_version":1}