{"id":"c5a67a99-60d2-43f9-be00-9ebccf56e510","arxiv_id":"2508.07585","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A lightweight SOD model that supervises decoder stages with coarse center maps versus fine boundary maps, and fuses features with two new modules, achieves new SOTA among lightweight image and video models.","lead":"GAPNet is a lightweight neural network for finding salient objects in both images and video, using different 'granularities' of supervision for different network stages. It reports top accuracy among lightweight models on five image datasets and most tested video datasets, pointing toward efficient real-time saliency detection on edge devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Video SOD claim contradicted by Table 2: on DAVSOD, GAPNet trails JL-DCF-Light on all three metrics, yet abstract claims SOTA among lightweight video SOD.","rationale":"I read the paper in good faith. The image SOD results appear strong and the method is coherent. The most load-bearing issue I find is the overbroad claim about video SOD SOTA: Table 2 directly contradicts it. The reader's CONDITIONAL verdict already flags the DAVSOD loss as a minor issue, but I think it deserves more weight because it targets the headline claim directly. A concrete check—re-running the DAVSOD evaluation—would settle whether this is a genuine overclaim or a reporting artifact. I am not alleging dishonesty; it may be a definitional issue (e.g., 'SOTA' meaning overall or on most datasets). My recommendation is to keep the CONDITIONAL verdict, with the condition being a qualified claim or an aggregate justification. The test-set hyperparameter tuning is a separate concern that the reader correctly identified; I partially agree with that but do not rely on it.","tokens_in":24336,"tokens_out":8554,"duration_ms":81976,"concrete_test":"Reproduce the video SOD evaluation using the released code/saliency maps and official evaluation scripts on DAVSOD for GAPNet and JL-DCF-Light. If the metrics are confirmed (GAPNet lower on all three), require the authors to revise the SOTA claim to name the specific benchmarks where GAPNet is SOTA or to provide an aggregate metric demonstrating overall state-of-the-art. If the metrics are not confirmed, the claim may stand pending corrected tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2 reports GAPNet on DAVSOD with Sα=0.706, Fmaxβ=0.597, MAE=0.089, while JL-DCF-Light achieves Sα=0.728, Fmaxβ=0.630, MAE=0.088. GAPNet is worse on every reported metric. The text in Sec. 4.2.1 concedes this ('Although the lightweight JL-DCF achieves better performance than our model on the DAVSOD dataset, it falls short on other datasets...'), but the abstract and conclusion state an unqualified 'new state-of-the-art performance among lightweight image and video SOD models.' DAVSOD is a standard video SOD benchmark, so this is a direct counterexample to the headline claim. Either 'state-of-the-art' must be qualified to the specific datasets where GAPNet wins, or an aggregate justification (e.g., mean rank across datasets) must be supplied. This is not an artifact of test-set hyperparameter selection; it is an internal inconsistency between the reported numbers and the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GAPNet, a lightweight encoder-decoder for image and video salient object detection. The decoder uses granularity-aware connections: low-level features are supervised by a boundary/others map, high-level features by a center map, and the final output by the full saliency map. Two fusion modules are introduced: GPC (pyramid atrous convolution with pooled self-attention) for low-level fusion and CSA (cross-scale attention) for high-level fusion, with a lightweight self-attention global feature extractor on top of MobileNet-V2. In video, RGB and FlowNet-2 optical flow features are fused in a two-stream variant. The authors report state-of-the-art results among lightweight models on five image SOD datasets and four video SOD datasets, at high speed (571/349 FPS), and provide code.","tokens_in":24634,"tokens_out":7358,"duration_ms":83118,"significance":"If the reported results are reproducible, GAPNet is a useful contribution to edge-deployable SOD: it offers a simple, modular recipe for matching granularities of supervision to decoder scales, with clear ablations (Tables 3–6) and thorough comparison on standard benchmarks. The image-SOD claim is well supported: GAPNet improves over EDN-Lite on all five datasets across all six metrics, and the gains are not restricted to one benchmark. The paper is also transparent about evaluation: comparisons use official saliency maps where available, and code is promised. The central conceptual claim—that coarse supervision at high levels and fine supervision at low levels improves lightweight decoders—is plausible and falsifiable. However, the video SOTA claim is not supported as stated, and the image margin is attenuated by test-set hyperparameter selection and lack of statistical confidence. I view the core idea as sound and the manuscript as close, but needing revision.","major_comments":[{"comment":"The unqualified claim of 'a new state-of-the-art performance among lightweight image and video SOD models' (Abstract, echoed in Conclusion) is contradicted by the paper's own Table 2 on DAVSOD: GAPNet trails JL-DCF-Light on all three reported metrics (S-alpha 0.706 vs 0.728, Fmax-beta 0.597 vs 0.630, MAE 0.089 vs 0.088). Sec. 4.2.1 concedes this explicitly. DAVSOD is a standard video SOD benchmark, so the headline must be qualified to the datasets where GAPNet actually leads, or an aggregate justification (e.g., average rank with significance testing) must be supplied.","section":"Sec. 4.2.1, Table 2; Abstract"},{"comment":"The two hyperparameters that define the GPC module—adaptive pooling size m=7 and pyramid split ratios [1/8, 1/8, 1/4, 1/2]—are selected by ablation on DUTS-TE, which is also the main test dataset (Sec. 4.1). The headline margin over EDN-Lite on DUTS-TE Fmax-beta is only 1.1% (0.867 vs 0.856); tuning on the test set can inflate this margin. Since the improvement also appears consistently across the four other datasets, this does not overturn the image claim, but an independent validation split or a sensitivity analysis is needed to establish that the configuration generalizes rather than being optimized to DUTS-TE.","section":"Sec. 4.3.1, Table 3; Sec. 4.3.2, Table 4"},{"comment":"All comparisons are single-run, with no error bars or significance tests, and several decisive margins are very small (e.g., DAVIS S-alpha 0.893 vs 0.892 and Fmax-beta 0.864 vs 0.863 against JL-DCF-Light; DUT-OMRON MAE 0.057 vs 0.058). For a state-of-the-art claim, at least a standard deviation over multiple seeds, or a paired test over samples for the closest competitor, should be reported. This is especially important in the video results, where the DAVIS advantage over JL-DCF-Light is within 0.001–0.004 on two metrics.","section":"Tables 1–2, Sec. 4.2.1"}],"minor_comments":[{"comment":"Typo: 'newstate-of-the-art' should be 'new state-of-the-art'.","section":"Section 5"},{"comment":"Axis labels read 'DA VIS' and the caption cites 'DAVIS [93]'; should be 'DAVIS' and the correct reference [100].","section":"Fig. 6"},{"comment":"The text says 'six lightweight models' in the comparison, but Table 1 lists seven lightweight competitors (HVPNet, CSNet, SAMNet, EDN-Lite, ELWNet, LARNet, ADMNet+).","section":"Sec. 4.2"},{"comment":"In Sec. 2, 'lightweight backbones like EfficientNet-B0 [26]' cites ref. [26], which is a 2013 salient-region-detection paper, not EfficientNet-B0. The correct Tan & Le (2019) EfficientNet reference appears to be missing.","section":"References [26]"},{"comment":"The explanation of the dot-product and L1-norm symbols is garbled ('·' and '·' are both rendered as middle dots). Spell out the operations in words or use distinct symbols.","section":"Eq. (11)"},{"comment":"Cross-scale attention is described only in words. Since E3 and E4 have different spatial resolutions (1/16 and 1/32), the flatten-and-concatenate step should be stated explicitly, including how the different sequence lengths are handled in the attention computation.","section":"Sec. 3.2.2, Eq. (10)"},{"comment":"The video feature-fusion mechanism is described verbally but no equations, parameter counts, or ablations are given for the two-stream variant. A schematic or pseudocode would aid reproducibility.","section":"Sec. 3.2.3"},{"comment":"For ELWNet and LARNet, the paper states that numbers were extracted from published papers because no official code/maps are available. This limits the controlled comparison on those rows, although it does not affect the main conclusions.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The unqualified 'state-of-the-art' wording in the abstract is a factual mismatch with Table 2 on DAVSOD; the authors already acknowledge it in the text, so the fix is local (qualify the claim or provide an aggregate justification). The test-set hyperparameter tuning in Sec. 4.3 is a methodological concern that is common in this literature; the cross-dataset consistency of the image results leads me to believe the inflation is modest. No circularity is apparent: evaluation uses external benchmarks and pretrained backbones. The self-citations to EDN and MobileSal are appropriate as baselines. I recommend major revision, mainly to repair the video-SOTA claim and add a validation protocol for the GPC hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful lightweight SOD paper. The granularity-aware supervision idea—coarse center maps supervising high-level side-outputs and boundary-others maps supervising low-level side-outputs—is new to me in this form, and the GPC/CSA modules are simple and efficient. The image experiments are extensive and the gains over EDN-Lite (1.1–2.3 Fmax points) plus the speed/accuracy Pareto plots look credible. The paper also deserves credit for including video and for building a two-stream version that is genuinely fast.\n\nWhere it wobbles: the abstract and conclusion claim 'state-of-the-art among lightweight image and video SOD models,' but Table 2 shows GAPNet losing to JL-DCF-Light on DAVSOD on all three metrics (S, Fmax, MAE). The body concedes this, so the unqualified claim should be fixed, or the 'SOTA' should be dataset-qualified. The other issue is that the two key hyperparameters—GPC pooling size m=7 and the pyramid split ratios—are selected using ablations run on the DUTS-TE test set (Sec 4.3). That makes the reported margins over baselines a bit optimistic; an independent holdout or nested validation would shore this up. No error bars, but that's typical for SOD. Also, Table 6 shows the granularity-aware supervision only helps for their specific assignment; other combinations are worse than uniform supervision. So 'paradigm' is a bit of an overstatement—it's a tuned recipe, not a demonstrated general principle.\n\nOverall, the central image-SOD claim holds up well enough to deserve a serious referee. The video claim needs to be reworded, and the authors should clarify the test-set tuning. I'd send it to peer review, but I'd flag the overclaim and ask for a re-evaluation on a properly held-out validation split.","headline":"Useful lightweight SOD with a clever granularity supervision split, but video SOTA claim overreaches and ablations tune on the test set.","tokens_in":25109,"tokens_out":2487,"would_cite":true,"duration_ms":28170,"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":"GAPNet claims that a lightweight salient object detector beats prior lightweight models and rivals heavy ones when supervision granularity is matched to feature scale—coarse center maps for high-level features, fine boundary maps for low-le","keywords":["salient object detection","lightweight model","granularity-aware paradigm","multi-scale feature fusion","video salient object detection","cross-scale attention","deep supervision"],"falsifier":"Retrain GAPNet from the released code, but select the pooling size and split ratios on a held-out validation split of the training set rather than on DUTS-TE; if the $F_\\beta^{\\max}$ margin over EDN-Lite drops below roughly 0.005 or the ranking against other lightweight baselines changes, the quantified advantage of the granularity-aware paradigm is not replicated.","tokens_in":24268,"feed_emoji":"🎯","tokens_out":19023,"duration_ms":168255,"temperature":0.7,"pith_summary":"GAPNet is a lightweight network for salient object detection, the task of locating the most attention-grabbing object or region in an image or video frame. Its central idea is that a small network should supervise each decoder stage with a saliency target of matching granularity: coarse object-center maps for high-level features and fine boundary maps for low-level features, rather than giving every stage the full saliency map. The paper reports that this granularity-aware supervision, implemented with granular pyramid convolution and cross-scale attention modules plus a compact global self-attention head, sets a new state of the art among lightweight image and video SOD models—for instance, best F-measure 0.867 versus 0.856 for EDN-Lite on DUTS-TE—while running at hundreds of frames per second. If the claim is right, it means that well-chosen supervision signals can substantially close the accuracy gap between tiny networks and much heavier ones, which matters for deployment on phones and edge devices.","feed_headline":"Outperforms prior lightweight models on five saliency benchmarks","feed_subtitle":"Matching supervision to feature granularity closes the gap to heavy models at hundreds of FPS.","key_machinery":"Granularity-aware paradigm: decompose the ground-truth saliency foreground into boundary (within 5 pixels of background), center (top 20% of distance to background), and others, then supervise high-level decoder outputs with the coarse center map, low-level outputs with the fine boundary-plus-others map, and the final output with the full map. The implementation rests on two fusion blocks: Granular Pyramid Convolution (GPC), an attention-refined multi-scale atrous convolution with channel splits [1/8,1/8,1/4,1/2] and an attention branch that pools to m=7 before computing a compact self-attention; and Cross-Scale Attention (CSA), where the query is computed from both scales but the keys and v","core_discovery":"On the paper's own terms, the central claim is that supervising all decoder side-outputs with the full saliency map wastes the limited feature richness of lightweight backbones. GAPNet decomposes the ground truth into center, boundary, and others regions by pixel distance to the nearest background, then applies granularity-aware deep supervision: the high-level side-output D2 is trained with the coarse center map, the low-level side-output D1 with the boundary-plus-others map, and the final output D3 with the full map. With two compact modules—granular pyramid convolution for low-level fusion and cross-scale attention for high-level fusion—plus a self-attention global extractor, a MobileNetV","pith_inferences":["A direct test of the thesis would be to retrofit other lightweight U-Net style decoders (e.g., SAMNet, HVPNet) with granularity-matched side-output supervision while leaving their architecture unchanged; the paper's ablations suggest part of the gain comes from the supervision scheme, but the paper does not isolate that attribution outside its own modules.","Because the hyperparameters were selected using the DUTS-TE test set, the published margins could be optimistic under independent re-evaluation; tuning on a held-out split during ablation would make the state-of-the-art claim more robust.","The center/boundary/others decomposition uses fixed thresholds (5 pixels, top 20% of distance). Varying these thresholds or making the decomposition learnable would clarify whether the specific definitions matter or just the coarse-vs-fine distinction.","The video pipeline depends on offline optical flow from an external estimator, which is itself a cost and an error source; an online variant that learns motion jointly would show how much of the video improvement is attributable to the granularity paradigm rather than to the optical-flow preprocessing."],"forward_implications":["Lightweight SOD models can be made more accurate by replacing uniform full-map deep supervision with granularity-matched supervision; this design rule is likely to transfer to other small backbones and decoders.","The margin over EDN-Lite on DUTS-TE ($F_\\beta^{\\max}$ 0.867 vs 0.856) indicates that the earlier lightweight state of the art had remaining decoder capacity, not that lightweight encoders are the primary bottleneck.","The same decoder, fed with RGB plus optical flow, achieves competitive video results ($S_\\alpha$ 0.893 on DAVIS) at roughly 350 FPS, extending the paradigm to spatio-temporal saliency.","The reported hyperparameters (m=7, split ratios [1/8,1/8,1/4,1/2]) offer concrete defaults for future lightweight SOD training; ablations show both smaller and larger pooling sizes reduce accuracy."],"supporting_citations":[{"why":"Supplies the EDN-Lite baseline that GAPNet beats and the extreme-downsampling global module used in the global-feature ablation.","marker":"[18]"},{"why":"Provides the lightweight backbone encoder (MobileNetV2) that all GAPNet experiments build on.","marker":"[27]"},{"why":"The prior label-decoupling method whose iterative refinement GAPNet contrasts with its single-stage granularity supervision.","marker":"[68]"},{"why":"Supplies the DUTS-TR training set and DUTS-TE test set used for training and for all image ablations and benchmark comparisons.","marker":"[93]"},{"why":"One of the video datasets used for training/fine-tuning and for reporting the video SOD results.","marker":"[100]"},{"why":"Generates the offline optical flow required by the two-stream video architecture.","marker":"[92]"},{"why":"A lightweight video baseline (MMN) that GAPNet compares against and re-implements in a lightweight setting.","marker":"[82]"},{"why":"A lightweight video baseline (JL-DCF) that GAPNet outperforms on speed and most video datasets.","marker":"[91]"}],"fun_headline_variants":["Granularity-aware supervision makes lightweight SOD state of the art","Lightweight SOD gets a precision boost via granularity-aware training","Splitting saliency by granularity lifts lightweight SOD to SOTA","GAPNet: granularity-matched supervision closes gap to heavy SOD"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The reported results rest on hyperparameters (pooling size $m=7$ and split ratios $[1/8,1/8,1/4,1/2]$) tuned on the DUTS-TE test set, so the claim of a new state of the art depends on those choices not being overfit to that test set.","fun_headline_variants_meta":{"raw":{"variants":["Granularity-aware supervision makes lightweight SOD state of the art","Lightweight SOD gets a precision boost via granularity-aware training","Splitting saliency by granularity lifts lightweight SOD to SOTA","GAPNet: granularity-matched supervision closes gap to heavy SOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1211,"prompt_tokens":765,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":509,"tokens_out":446,"duration_ms":4960,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:00:12.323697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain GAPNet from the released code, but select the pooling size and split ratios on a held-out validation split of the training set rather than on DUTS-TE; if the $F_\\beta^{\\max}$ margin over EDN-Lite drops below roughly 0.005 or the ranking against other lightweight baselines changes, the quantified advantage of the granularity-aware paradigm is not replicated.","supporting_citations":[{"cited_title":"Label decoupling framework for salient object detection,","cited_arxiv_id":null,"evidence_quote":"The prior label-decoupling method whose iterative refinement GAPNet contrasts with its single-stage granularity supervision."},{"cited_title":"Learn- ing to detect salient objects with image-level supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the DUTS-TR training set and DUTS-TE test set used for training and for all image ablations and benchmark comparisons."},{"cited_title":"A benchmark dataset and eval- uation methodology for video object seg- mentation,","cited_arxiv_id":null,"evidence_quote":"One of the video datasets used for training/fine-tuning and for reporting the video SOD results."},{"cited_title":"Flownet 2.0: Evolution of optical flow estimation with deep networks,","cited_arxiv_id":null,"evidence_quote":"Generates the offline optical flow required by the two-stream video architecture."},{"cited_title":"Motion-aware mem- ory network for fast video salient object detection,","cited_arxiv_id":null,"evidence_quote":"A lightweight video baseline (MMN) that GAPNet compares against and re-implements in a lightweight setting."},{"cited_title":"Siamese network for rgb-d salient object detection and beyond,","cited_arxiv_id":null,"evidence_quote":"A lightweight video baseline (JL-DCF) that GAPNet outperforms on speed and most video datasets."}],"review_version":1}