{"id":"0fba0a9e-a347-43f6-afcc-59e026187247","arxiv_id":"2508.07170","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LMFNet is a lightweight saliency detector built from depthwise separable dilated convolutions, but its results and implementation do not match the paper's central claims.","lead":"This paper proposes LMFNet, a 0.81M-parameter network for salient object detection that combines depthwise separable and dilated convolutions. Its reported results do not support the claimed state-of-the-art status, and the described fully connected layer is not what the network actually runs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LMF layer's defining 'fully connected' operation (Eq. 8) is not implemented in the forward pass (Eq. 11), so the experiments test a different, less-connected architecture than the one proposed.","rationale":"The central claim has two load-bearing parts: (1) a novel fully connected LMF mechanism and (2) an efficiency/accuracy advantage. The reader's weakest_assumption targets part (1), and I agree: Eq. 8 and Eq. 11 are mutually inconsistent about the number of input feature maps each LMF layer receives, so the experiments cannot validate the proposed mechanism. This is the single most load-bearing concern because it does not depend on comparing against external SOTA results or on statistical significance; it follows directly from the paper's own equations. The performance contradiction in Table 1 (LMFNet is not SOTA and trails HVPNet/SAMNet and all traditional models) is real and independently supports rejection, but the untested-mechanism issue is more fundamental. The ablation table's test-set selection and absent error bars are aggravating but secondary. No formal verification or external reproduction is provided, so the internal mathematics and the released code are the only checks; hence the concrete test is a code inspection plus a minimal re-training experiment. Because my identified concern matches the reader's weakest_assumption and confirms the REJECT verdict, no verdict adjustment is needed.","tokens_in":16109,"tokens_out":6077,"duration_ms":55121,"concrete_test":"Download the released code (github.com/Shi-Yun-peng/LMFNet), instantiate LMFNet, and instrument each LMF layer's forward pass to count how many distinct input tensors the Eq.-8 entries I0..Im contain. If every LMF layer receives exactly one tensor at runtime, the fully connected LMF layer is absent. As a stronger test, implement a true multi-input LMF layer (each branch Ki receives, say, the previous three feature maps) and retrain on DUTS-TR; if DUTS-TE metrics do not change materially versus the single-input model, the claimed fully connected connectivity is not responsible for the reported performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's claimed novelty is the LMF layer in which each depthwise-separable dilated branch Ki processes every input feature map Ij (Eq. 8), then fuses via 1x1 conv (Eq. 10). This cross-input connectivity is what 'fully connected' means in the title and abstract. But Eq. 11 defines the encoder as F1 = f_LMF1(I), Fi = f_LMFP_i(F_{i-1}) for i=2..5, where each LMF layer receives exactly one tensor (I or F_{i-1}); there is no vector of feature maps I=[I0,...,Im] with m>0 at any layer. Eq. 8 therefore degenerates to Fi = Ki(I0), a per-branch single-input convolution. The decoder (Eq. 13) likewise feeds one concatenated tensor into each f_LMF. Thus the fully connected mechanism—the paper's central methodological contribution—is not exercised by any experiment in Tables 1, 2, 4, or 5. Even if the reported numbers are correct, they validate only a depthwise-separable dilated-conv stack with 1x1 fusion, not the proposed LMF mechanism. This is an internal inconsistency, not a matter of contested consensus. A secondary issue is that Table 2 selects dilation rates and loss terms using the same five test datasets used for final evaluation, with no held-out validation split and no error bars; this makes the reported configuration vulnerable to selection bias, but the untested mechanism alone is sufficient to undermine the paper's central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Lightweight Multi-scale Feature (LMF) layer based on depthwise separable dilated convolutions arranged in a 'fully connected' structure, and an SOD network LMFNet with 0.81M parameters. The paper claims state-of-the-art or comparable results on five benchmark datasets and also reports image-classification experiments. The central idea is that each dilated branch K_i processes a vector of input feature maps I=[I_0,...,I_m] and fuses the results. However, the forward equations in the actual encoder/decoder only ever pass a single tensor to each LMF layer, so the defining multi-input connectivity is not implemented. In addition, the reported configuration is selected by tuning hyperparameters on the test sets, and the abstract's performance claim is contradicted by the paper's own Table 1.","tokens_in":16476,"tokens_out":5158,"duration_ms":48660,"significance":"If the proposed fully connected LMF layer were actually realized and validated, it would be a useful contribution to lightweight multi-scale feature extraction. The paper's highly compact model (0.81M parameters) and the attempt to ablate dilation rates and loss variants are strengths, and the code link is provided. However, the mismatch between Eq. (8) and the implemented forward pass means that the experiments, as reported, validate only a depthwise separable dilated convolution stack with 1x1 fusion, not the fully connected LMF mechanism. The performance claims are also not supported by the tables: the reported F-measures are below several compared methods, and the text itself concedes lower performance than HVPNet and SAMNet. Because the defining mechanism is absent from the tested model, the central claim cannot be accepted on the present evidence.","major_comments":[{"comment":"The proposed LMF layer is defined with a vector input I=[I_0,...,I_m], and each branch K_i is applied to every input feature map: F_i = Concat(K_i(I_0),...,K_i(I_m)). In the encoder, however, Eq. (11) sets F_1 = f_LMF1(I) and F_i = f_LMFP_i(F_{i-1}), so every LMF layer receives exactly one tensor. The decoder Eq. (13) likewise passes a single tensor (possibly after an outer Concat) into each f_LMF layer. Consequently, m=0 in every executed LMF layer and Eq. (8) degenerates to F_i = K_i(I_0). The 'fully connected' cross-input connectivity that motivates the title and abstract is never tested. This is an internal inconsistency, not a matter of interpretation.","section":"§3.3 Eq. (8); §4.1 Eqs. (11), (13); Fig. 1"},{"comment":"The abstract states that LMFNet achieves 'state-of-the-art or comparable results' and outperforms several models, but Table 1 shows, for example, ECSSD Fβ=0.897 for LMFNet versus 0.941 for U2Net, and DUTS-TE Fβ=0.782 versus 0.876 for CDMNET. The text itself concedes that LMFNet is slightly worse than HVPNet and SAMNet and reports an average F-measure of 0.820 versus 0.877 for CDMNET. The stated claim is not supported by the paper's own quantitative results and should be corrected to 'competitive among lightweight models' at most.","section":"Abstract; Table 1; §5.2"},{"comment":"The dilation-rate vector and the loss function are selected by comparing variants on the same five test datasets used for the final evaluation (ECSSD, DUTS-TE, HKU-IS, PASCAL-S, DUT-OMRON), with no held-out validation split and no error bars or significance tests. Table 2 evaluates all rows on the test sets, and §5.3 reports tuning d[1] on the same five datasets. The reported configuration is therefore the result of fitting hyperparameters to the test data, and the resulting numbers do not provide an unbiased estimate of generalization performance.","section":"Table 2; §5.3"},{"comment":"The related-work section states, 'Additionally, we incorporate Deformable Modules to further improve the network's adaptability to object deformations.' The methodology and architecture description in Sections 3–4 and Figure 3 do not mention or implement deformable convolutions, and no experiment or ablation involving them is reported. As written, this introduces an unsupported methodological component that is not part of the actual model.","section":"§2.1 (last paragraph) vs. §§3–4"}],"minor_comments":[{"comment":"The heading reads 'Relate Work'; it should be 'Related Work'.","section":"§2 heading"},{"comment":"'Sigmod' should be 'Sigmoid'.","section":"Eq. (14)"},{"comment":"The text says the F-measure curve is 'as shown in formula (15)', but Eq. (15) is the loss function; the reference should be to Eq. (17).","section":"§5.1.3"},{"comment":"Eq. (1) defines d_i as 'stride', while the rest of the paper uses d as the dilation rate. The notation should be reconciled.","section":"Eq. (1)"},{"comment":"The section contains an unresolved cross-reference: 'As shown in Table ??'. Also, 'In Table3-10, we removed all dilated convolutions...' is ambiguous and should be clarified.","section":"§5.3"},{"comment":"The Table 1 caption says gray marks 'Results inferior to our model' for traditional networks, while §5.2 says 'we highlight the second-best values in gray and the best values in red'. These conventions should be made consistent.","section":"Table 1 caption and §5.2"},{"comment":"The comparison with lightweight models refers to 'Table IV', but the paper only has Tables 1–5; this should be corrected to the appropriate table.","section":"§5.2"},{"comment":"Row 10 of Table 2 uses '*' without explaining what configuration it represents.","section":"Table 2"}],"recommendation":"reject","confidential_remarks":"The core issue is not simply an overstatement: the proposed LMF layer as defined in Eq. (8) is not exercised by the implemented architecture in Eqs. (11)–(13). This is an internal inconsistency that cannot be repaired by local editing; it would require either redefining the contribution as a single-input depthwise separable dilated convolution stack or re-running experiments with a genuinely multi-input fully connected layer. The test-set hyperparameter selection and the unsupported 'state-of-the-art' claim further weaken the manuscript. The authors could potentially resubmit a corrected lightweight SOD paper with more modest claims, but the present version is not suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's headline mechanism—the 'fully connected' LMF layer—is not actually present in the implemented network. The generic definition in Eq. 8 has each dilated branch K_i operate on every input feature map I_j and concatenate the results. That is the cross-connectivity the title promises. But in the encoder (Eq. 11) every LMF layer receives exactly one tensor: F1 = f_LMF1(I), Fi = f_LMFP_i(F_{i-1}), and so on. Same in the decoder. So the experiments test a stack of depthwise-separable dilated convolutions with 1x1 fusion—a standard combination, not the proposed fully connected layer.\n\nThat is the main problem, and it is decisive. The paper's own tables show LMFNet ranking below HVPNet and SAMNet on nearly every metric, and well below the large models. The abstract's 'state-of-the-art' claim is not supported by the evidence in the paper. The hyperparameters (dilation vector, loss weights) were tuned on the same test sets used for final comparison, with no held-out validation and no error bars, which makes the reported configuration look selected rather than predictive.\n\nTo give credit where it's due: the paper is honest about underperforming the top lightweight models—it explicitly says it is 'slightly worse' than HVPNet and SAMNet—and it does provide ablation studies for dilation rates and loss functions. The code is public. The actual architecture, while not novel, is a clean, minimal design that gets 0.81M parameters and reasonable results. That is not worthless.\n\nBut the central claim—that a fully connected multi-scale layer is the reason for the performance—is untested. The internal inconsistency between Eq. 8 and Eq. 11 is the kind of error a careful referee would catch quickly, and it breaks the paper's contribution. The CIFAR extension is routine and adds little. I would not cite this work as-is, and I would not bring it to the reading group except as a cautionary example about matching equations to implementation.\n\nRecommendation: send to peer review, but the outcome should be rejection. If the authors can actually implement the fully connected topology and show it helps, or if they reframe the paper as a simple lightweight multiscale network, there might be a salvageable result. As it stands, the paper does not test what it claims.","headline":"The LMF layer's defining cross-input connectivity is not implemented in the forward pass, so the experiments test a different architecture than the one proposed.","tokens_in":16990,"tokens_out":4707,"would_cite":false,"duration_ms":45890,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes LMFNet, a 0.81M-parameter network built on a fully connected multi-scale layer of depthwise separable dilated convolutions, and claims competitive salient object detection on five benchmarks.","keywords":["lightweight networks","multi-scale feature learning","salient object detection","depthwise separable dilated convolution","LMF layer","receptive field","model efficiency","CIFAR"],"falsifier":"Inspect the released code and check the forward pass of an LMF layer: if every LMF layer receives only one feature map (as in Eq. 11) and never a vector of maps (as in Eq. 8), then the claimed fully connected structure is not what was tested; alternatively, re-run the DUTS-TE evaluation with the released weights to verify the reported MAE and F-measure numbers.","tokens_in":15961,"feed_emoji":"👁","tokens_out":4321,"duration_ms":40655,"temperature":0.7,"pith_summary":"The paper tries to establish that multi-scale feature extraction for salient object detection can be done with an extremely small network by stacking a new layer, the LMF layer, built from depthwise separable dilated convolutions with varying dilation rates. If true, this would give edge devices a practical saliency detector with 0.81M parameters and 3.8G FLOPs, and the same encoder would transfer to image classification. The authors claim competitive or state-of-the-art results, though the reported tables show performance slightly below some lightweight peers like HVPNet and SAMNet.","feed_headline":"Salient-object detection shrinks to 0.81M parameters","feed_subtitle":"A fully connected LMF layer of dilated depthwise convolutions delivers competitive saliency maps at a fraction of the compute.","key_machinery":"The LMF layer: a set of depthwise separable dilated convolutions K = [K0,...,Kn] with dilation factors d, where each Ki is applied to input feature maps and the outputs are concatenated and fused by a 1x1 convolution. It provides multiple receptive fields in one layer at low parameter cost, with the paper's stated connectivity rule (Eq. 8) as its defining feature.","core_discovery":"The central discovery claimed is that a fully connected arrangement of depthwise separable dilated convolutions with dilation rates such as [1,4,12,36,108] yields diverse receptive fields that capture multi-scale information at very low parameter cost. Each dilation branch processes input feature maps, and a 1x1 convolution fuses the results. The authors also introduce a design rule: adjacent dilation rates should keep a ratio smaller than the kernel size to avoid information loss. On DUTS-TR training and five test sets, LMFNet reaches the reported metrics, and the encoder achieves competitive CIFAR accuracy.","pith_inferences":["The paper's Eq. 8 defines a fully connected layer where each dilation branch processes every input map, but the forward pass in Eq. 11 feeds each LMF layer a single feature map; a direct consequence is that the implemented model differs from the described architecture, so the claimed benefit of full connectivity is untested.","The abstract's 'state-of-the-art' claim is stronger than the evidence in Table 1, where LMFNet trails HVPNet and SAMNet on most metrics; a more defensible claim is a competitive efficiency-performance trade-off.","The dilation-ratio rule could generalize as a heuristic for any dilated-convolution stack, and testing it on other lightweight backbones would clarify its scope.","The CIFAR results suggest the LMF encoder is a reusable lightweight backbone; probing it on ImageNet or dense prediction tasks would illuminate its generality."],"forward_implications":["LMFNet achieves competitive salient object detection with 0.81M parameters and 3.8G FLOPs, making deployment on resource-constrained devices plausible.","The rule that the ratio of adjacent dilation rates must stay below the kernel size offers a practical design principle for layered dilated convolutions.","The same encoder reaches competitive CIFAR-10/100 accuracy with fewer than 0.66M parameters, suggesting the LMF layer is transferable beyond SOD.","Ablation results indicate that increasing parameter count does not significantly improve performance, pointing to redundancy in the design.","The hybrid SSIM+BCE+IoU loss measurably improves performance over each component alone."],"supporting_citations":[{"why":"Supplies depthwise separable convolution, the base parameter-reduction operation the LMF layer builds on.","marker":"[23]"},{"why":"Supplies dilated/atrous convolution for receptive-field expansion without adding parameters.","marker":"[24]"},{"why":"Provides the hybrid loss (SSIM+BCE+IoU) used by LMFNet during training.","marker":"[38]"},{"why":"A lightweight SOD baseline LMFNet is compared against; defines the efficiency-performance context.","marker":"[17]"},{"why":"DUTS is the training dataset for all SOD experiments, providing the supervision signal for the reported results.","marker":"[28]"}],"fun_headline_variants":["Lightweight LMFNet hits saliency benchmarks with 0.81M params","Multi-scale saliency with just 0.81M parameters","LMF layer squeezes multi-scale saliency into 0.81M","Tiny saliency net with fully connected dilated convs","0.81M params for state-of-the-art saliency maps"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Each LMF layer is assumed to be truly fully connected, with every dilated branch processing every input feature map; in the implemented forward pass, each layer receives a single feature map, so the fully connected connectivity described in Eq. 8 is not present in the tested model.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight LMFNet hits saliency benchmarks with 0.81M params","Multi-scale saliency with just 0.81M parameters","LMF layer squeezes multi-scale saliency into 0.81M","Tiny saliency net with fully connected dilated convs","0.81M params for state-of-the-art saliency maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001095,"raw_usage":{"total_tokens":4379,"prompt_tokens":689,"completion_tokens":3690,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":3597}},"tokens_in":433,"tokens_out":3690,"duration_ms":25361,"temperature":1.0,"reasoning_tokens":3597,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:17:30.226380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code and check the forward pass of an LMF layer: if every LMF layer receives only one feature map (as in Eq. 11) and never a vector of maps (as in Eq. 8), then the claimed fully connected structure is not what was tested; alternatively, re-run the DUTS-TE evaluation with the released weights to verify the reported MAE and F-measure numbers.","supporting_citations":[{"cited_title":"IEEE transactions on pattern analy- sis and machine intelligence 40(4), 834–848 (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies dilated/atrous convolution for receptive-field expansion without adding parameters."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the hybrid loss (SSIM+BCE+IoU) used by LMFNet during training."},{"cited_title":"IEEE Transactions on Image Processing 30, 3804–3814 (2021)","cited_arxiv_id":null,"evidence_quote":"A lightweight SOD baseline LMFNet is compared against; defines the efficiency-performance context."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"DUTS is the training dataset for all SOD experiments, providing the supervision signal for the reported results."}],"review_version":1}