{"id":"3a493734-4e42-4db4-97d9-5b0ad3382828","arxiv_id":"2507.20809","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SCANet with the new Split Coordinate Attention module achieves 91.61% and 75.49% IoU on the WHU and Massachusetts building datasets, slightly beating prior SOTA.","lead":"This paper introduces a new attention module, Split Coordinate Attention, that combines coordinate and split attention to improve building footprint extraction from satellite images. The resulting network, SCANet, reports slightly higher accuracy than previous state-of-the-art methods on two public datasets.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on single-run IoU gaps of 0.45–0.46% with no error bars; these gaps are within plausible seed-to-seed variance, so the reported ordering may not be statistically reliable.","rationale":"The reader and I converge on the same weak point: the empirical superiority is asserted from single runs with no error bars and a uniform training protocol that may not suit all baselines. I do not see a mathematical flaw in Eqs. (2)–(6); the module is a reasonable combination of coordinate attention and split attention, and the parameter counts in Table 3 are internally consistent. The issue is that the decisive differences (0.45–0.46) are exactly the size of typical seed noise, so the central claim is under-determined by the reported evidence. This does not change the conditional verdict: the paper should be treated as promising but not yet reliable until repeated-seed statistics and clarified baseline training are provided. The secondary concern about 'ResNet-14/26' naming in Table 4 does not by itself invalidate the main comparison, since Tables 1–2 use standard backbones, but it should be corrected in revision.","tokens_in":8386,"tokens_out":5929,"duration_ms":69712,"concrete_test":"Run SCANet, BCTNet, ResNeSt+UNet++, and ResNet-101+CA on both WHU and MASS with 5 random seeds using the protocol in Sec. 3.2, and report mean±std IoU and the per-seed SCANet-minus-BCTNet difference. Also run BCTNet and TransFuse-L under their originally published training recipes, or with a small LR/batch search, for one seed each. If the mean SCANet−BCTNet gap on either dataset is smaller than the larger of the two stds, or if a paired test across the 5 seeds gives p≥0.05, the SOTA claim is not established by the present tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the abstract's SOTA claim is that the observed ordering in Tables 1–3 reflects the proposed module rather than training noise or unequal baselines. The key evidence is a set of single-run numbers with no error bars: SCANet beats BCTNet by 0.46 IoU on WHU and 0.45 on MASS, and beats the SA/CA variants by roughly 1.0 IoU in Table 3. Section 3.2 states a consistent protocol is used for all methods, including batch size 4 on MASS; for segmentation models on these benchmark splits, seed-to-seed IoU variation of 0.3–0.8 is common, so a 0.45-point margin is not decisive without repeated runs. The uniform training protocol is itself a risk: transformer-based baselines such as BCTNet and TransFuse-L typically need different learning rates and longer schedules, so a one-size-fits-all recipe can systematically lower their reported IoU. The code is not commit-pinned, so the exact configuration behind the reported numbers cannot be independently verified. The module's internal equations are coherent; the vulnerability is empirical and statistical.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Split Coordinate Attention (SCA), a plug-and-play attention module that decomposes the global pooling used in split attention into horizontal and vertical pooled descriptors, applies per-group dense transformations, and combines horizontal and vertical soft-attention weights to reweight cardinal-group features. Inserting SCA into ResNet/ResNeSt backbones and decoding with UNet++ yields SCANet. The authors report state-of-the-art IoU values of 91.61% on the WHU Building Dataset and 75.49% on the Massachusetts Building Dataset, together with ablations, complexity comparisons, Grad-CAM++ visualizations, and a diagonal pixel-similarity analysis. The module equations in Section 2.2 are coherent, and the parameter analysis in Table 4 indicates that the gains are not obtained simply by enlarging the model. The central weakness is empirical: the state-of-the-art claim rests on single-run IoU differences of 0.45-0.46 percentage points over strong baselines, with no error bars, no repeated seeds, and no evidence that the uniform training protocol is appropriate for transformer-based baselines.","tokens_in":8635,"tokens_out":5029,"duration_ms":52739,"significance":"If the reported numbers are reproducible and the comparisons are fair, the paper would provide a useful lightweight attention module for building footprint extraction: it is simple, adds no parameters relative to ResNeSt in the main configuration, is evaluated on two standard public benchmarks, and the code is released. The interpretability analyses (Grad-CAM++ and diagonal similarity) are a positive complement, and the complexity analysis is a useful check that the reported gains are not simply a consequence of larger models. However, the quantitative improvement over BCTNet is modest (0.45-0.46 IoU), and the paper's principal contribution is empirical. Without statistical reliability and a clearly fair baseline protocol, the claimed state-of-the-art performance is not established. The module itself is a reasonable combination of coordinate attention and split attention, but its practical value depends on resolving these empirical concerns.","major_comments":[{"comment":"The abstract and Section 3.3 claim state-of-the-art performance based on single-run IoU differences of 0.46% (WHU) and 0.45% (MASS) over BCTNet, but no standard deviations, confidence intervals, or multiple-seed experiments are reported. On these benchmark splits, seed-to-seed IoU variation for segmentation models is commonly in the 0.3-0.8 range, so the observed margins are within plausible run-to-run noise. Please report means and standard deviations over at least three independent runs for SCANet and the nearest baselines, and state whether the ordering is consistent across seeds. This is load-bearing because the central claim of the paper is empirical superiority.","section":"Section 3.3, Tables 1 and 2"},{"comment":"The comparison may be unfair to transformer-based baselines. Section 3.2 states that 'consistent hyperparameter settings' are used for all experiments, including batch size 4 on MASS, AdamW with learning rate 1e-3, and IoU-based learning-rate halving. Methods such as BCTNet, TransFuse-L, and Swin generally require different learning rates, warmup schedules, and training lengths; a one-size-fits-all recipe can systematically depress their reported IoU. Please either run each baseline with its official or recommended training configuration, or provide evidence that the chosen protocol is near-optimal for every baseline. In addition, Section 2.3 does not specify where SCA is inserted into the backbone: which stages, which residual blocks, and how many SCA blocks are used. This information is needed to reproduce SCANet.","section":"Sections 3.2 and 2.3"},{"comment":"Table 4 lists 'ResNet-14' and 'ResNet-26' as backbones. Standard ResNet depths are 18, 34, 50, 101, and 152; ResNet-14 and ResNet-26 are non-standard and their block configurations are not described. Without stating the number of blocks per stage, stem, stride pattern, and how 'ResNet-14/26' relate to standard ResNets, the parameter counts and IoU values in this table cannot be verified or compared. Please clarify these configurations and, if they are custom variants, define them precisely.","section":"Table 4"},{"comment":"The ablation table mixes changes in the attention module and the decoder, and the parameter counts are not self-explanatory: Baseline UNet++ has 68.0M, +SA has 73.2M, +CA has 79.2M, +SCA UNet++ has 73.2M, and +SCA UNet has 55.3M. If +SCA and +SA have identical parameter counts, the table should state explicitly what backbone and insertion points are used for each row. If the baseline row uses a different backbone or insertion configuration than the +SCA row, the comparison does not isolate the effect of the module. Please report the backbone, insertion locations, and decoder for every row, and add a row with SCA inserted into the same backbone and decoder as the baseline.","section":"Section 3.4, Table 3"}],"minor_comments":[{"comment":"The symbols S_h_c and S_w_c are described as 'Dense connections' but are not formally defined; please specify the input/output dimensions, the number of layers, and the growth rate of these dense connections.","section":"Section 2.2, Eq. (5)"},{"comment":"Figure 1 contains labels 'PreQuantization Stage 1' and 'PostQuantization Stage N' that are not defined anywhere in the text; please define them or remove them.","section":"Figure 1"},{"comment":"There are several typographical and grammar issues: 'featuremap Group' in Section 2.1, 'the c-th channel is calculated' in Section 2.2, and 'two spatial range of pooling kernels' in the abstract. These should be corrected.","section":"Throughout"},{"comment":"Equation (4) uses [·, ·] for joining elements, and the phrase 'activation function applied after convolution' for G_c^i is vague; please specify whether G_c^i is a convolutional block with batch norm and ReLU, and give the kernel size.","section":"Section 2.2, Eq. (4)"},{"comment":"The code URL is given but no commit hash is provided, so the exact configuration cannot be pinned. Reference [12] is cited only as 'arXiv preprint (2017)' with no identifier; please add the arXiv number.","section":"References and code"},{"comment":"Table 4 reports IoU for Swin UNETR and Swin-B, but these methods do not appear in Tables 1-2; the text does not describe the experimental setup for these additional comparisons. Please clarify whether these numbers come from the same protocol and where the corresponding details are reported.","section":"Table 4 versus Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":"The author B. Zhao is also a co-author of ConvBNet (reference [21]), one of the compared baselines. The comparison itself is not circular, but this relationship should be disclosed to the editor and, ideally, to readers. The central risk is the lack of statistical evidence for the claimed superiority; the paper needs repeated runs and clearer baseline configurations before the state-of-the-art claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SCANet is a tidy, honest engineering paper. The Split Coordinate Attention module is exactly what it says: coordinate attention with split attention's cardinal-group weighting folded in. That specific combination is new as far as the cited literature goes, and the internal ablation shows it earning its keep. On WHU, SCA beats plain CoordAttention by about 1 IoU point (91.61 vs 90.62) while using fewer parameters (73.2M vs 79.2M), and it also beats split attention. The module math is simple and coherent, and the authors report parameter counts across several backbones, which is more than many papers in this space do. The qualitative visualizations are a nice touch, not proof, but supportive.\n\nThe soft spots are real but not fatal. The headline SOTA claim against BCTNet rests on single-run IoU gaps of 0.45–0.46, which is within seed-to-seed variance for these benchmarks. That is the stress-test's point, and it is correct. What the stress-test misses is that the more important comparison is the ablation against CA and SA, where the margin is roughly one full IoU point, not half a point. Still, that is also single-run, and the Massachusetts test set has only ten images, making any single-run number there especially shaky. The uniformly applied training recipe is another concern: transformer baselines like BCTNet and TransFuse-L often need different learning rates and longer schedules, so a one-size-fits-all protocol can quietly handicap them. The paper also needs to explain what \"ResNet-14\" and \"ResNet-26\" in Table 4 are; those are not standard ResNet depths and the configurations are undefined. The GitHub link is welcome, but without a commit pin or detailed training logs, the reported numbers are not independently verifiable.\n\nNone of this sinks the paper. The central claim—that splitting coordinate attention along cardinal groups helps building extraction—is plausible and internally consistent. The work is a modest, useful step, not a breakthrough. The right outcome is peer review with requests for multiple seeds, error bars, clarified backbone definitions, and a fairer comparison protocol for transformer baselines. I would not cite it in my own work this year, but I would send it to a reviewer who cares about attention modules for remote sensing.\n\nRecommendation: send to peer review, with expectation of revision.","headline":"A coherent attention-module hybrid with consistent internal gains, but the SOTA claim needs error bars and clarified backbones before it is trustworthy.","tokens_in":9125,"tokens_out":1816,"would_cite":false,"duration_ms":22586,"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":"A split coordinate attention module that pools each channel along x and y axes improves building footprint extraction on two public datasets.","keywords":["building footprint extraction","image segmentation","remote sensing","attention mechanism","coordinate attention","convolutional neural network","UNet++","split attention"],"falsifier":"Run SCANet and BCTNet on the WHU Building Dataset with five random seeds each, record mean and standard deviation of IoU; if the best-baseline margin overlaps zero or the ordering flips, the claimed state-of-the-art result is not supported.","tokens_in":8191,"feed_emoji":"🏙️","tokens_out":5333,"duration_ms":54268,"temperature":0.7,"pith_summary":"Building footprint extraction from remote sensing images is usually done with convolutional networks whose attention modules either summarise each channel globally or lose positional detail. This paper claims that a new plug-and-play module, Split Coordinate Attention (SCA), fixes that by pooling each channel separately along the horizontal and vertical axes and then splitting each feature group into weighted branches. Inserted into a ResNet or ResNeSt encoder and paired with a UNet++ decoder, the resulting SCANet reaches 91.61% IoU on the WHU Building Dataset and 75.49% on the Massachusetts Building Dataset, the best among the compared methods. The authors argue this shows that a targeted CNN attention design can outperform transformer-based extraction models while using no more parameters.","feed_headline":"SCANet hits 91.61% IoU for building footprint extraction","feed_subtitle":"A plug-in CNN attention module beats transformer-based rivals on two public building datasets.","key_machinery":"The load-bearing object is the Split Coordinate Attention (SCA) module. It decomposes the global pooling of split attention into two pooling kernels, $(H,1)$ along the horizontal direction and $(1,W)$ along the vertical direction, so each channel is encoded with its row and column context rather than a single scalar. The two feature maps are concatenated, reduced by a shared dense layer with batch normalisation and ReLU, split back into horizontal and vertical branches, and each branch passes through its own dense layer and sigmoid. A cardinal group's output is the input times the sum over splits of the product of the two directional weights; this product is the operation that lets the module weight long-range spatial interactions per channel.","core_discovery":"The paper's central claim is that replacing the global average pooling used in split attention with two directional pooling operations makes channel attention retain positional information. For each cardinal group, SCA pools along (H,1) and (1,W), concatenates the two direction-aware descriptors, passes them through a shared dense layer, splits the result back into height and width branches, and applies per-branch dense connections with sigmoid gating. Each split then votes with the product of its horizontal and vertical weights, so the network can emphasise spatially remote but semantically related pixels. With this module inside a CNN encoder, SCANet reports state-of-the-art intersection-over-union on both benchmark datasets and, in the ablation study, improves over both split attention and coordinate attention at equal or smaller parameter counts.","pith_inferences":["The diagonal-similarity visualisation suggests SCA enlarges the effective receptive field of the final encoder stage; a direct measurement of receptive field size across the compared backbones would test that explanation.","The same two-axis pooling design could be applied to other dense prediction tasks, such as road or water mapping, where buildings occupy a small fraction of the image; that test is not in the paper.","Because the Massachusetts dataset has lower resolution and noisier labels than WHU, the consistent margin there hints the module is robust to label noise, an idea the paper does not pursue.","The reported gaps are single runs; multi-seed evaluation would tell whether the improvement is stable enough for practical deployment."],"forward_implications":["Plugging SCA into an existing CNN encoder can raise building footprint IoU without increasing parameter count, so it may be a drop-in upgrade for deployed segmentation models.","Because SCA beats transformer-based baselines, the paper implies that long-range context in remote sensing can be obtained with cheap directional pooling rather than self-attention.","The module's parameter parity with split attention and coordinate attention means its gains come from where the pooling is applied, not from added capacity.","UNet++ as decoder is needed to realise the full gain: with a plain UNet decoder the same module reaches 90.41% IoU on WHU, still above baselines but below the reported peak."],"supporting_citations":[{"why":"Supplies the two-axis pooling decomposition that SCA adapts from coordinate attention.","marker":"[6]"},{"why":"Provides the split-attention group architecture and the ResNeSt baseline that SCA modifies.","marker":"[23]"},{"why":"Motivates channel attention and the global pooling that SCA replaces with directional pooling.","marker":"[7]"},{"why":"Gives the ResNet backbone into which SCA is inserted to form SCANet.","marker":"[5]"},{"why":"Supplies the UNet++ decoder used in the full SCANet model.","marker":"[27]"},{"why":"Defines the WHU Building Dataset, the primary evaluation benchmark.","marker":"[9]"},{"why":"Defines the Massachusetts Building Dataset, the secondary evaluation benchmark.","marker":"[13]"},{"why":"BCTNet, the strongest transformer-based baseline that SCANet claims to surpass.","marker":"[20]"},{"why":"ConvBNet, a dedicated building extraction baseline used in the comparison.","marker":"[21]"}],"fun_headline_variants":["SCANet: a plug-in attention module for precise building extraction","Direction-aware pooling boosts building footprint IoU to 91.61%","Split coordinate attention: SOTA building extraction without transformers","SCA module sharpens building footprints: 91.61% IoU","New attention module for CNNs beats SOTA on building datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage rests on the assumption that the 0.45-0.46 percentage-point IoU margins over the strongest baselines are stable, reproducible effects rather than noise from a single training run.","fun_headline_variants_meta":{"raw":{"variants":["SCANet: a plug-in attention module for precise building extraction","Direction-aware pooling boosts building footprint IoU to 91.61%","Split coordinate attention: SOTA building extraction without transformers","SCA module sharpens building footprints: 91.61% IoU","New attention module for CNNs beats SOTA on building datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3168,"prompt_tokens":871,"completion_tokens":2297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":2207}},"tokens_in":487,"tokens_out":2297,"duration_ms":18235,"temperature":1.0,"reasoning_tokens":2207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:13:24.721470+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SCANet and BCTNet on the WHU Building Dataset with five random seeds each, record mean and standard deviation of IoU; if the best-baseline margin overlaps zero or the ordering flips, the claimed state-of-the-art result is not supported.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the two-axis pooling decomposition that SCA adapts from coordinate attention."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Provides the split-attention group architecture and the ResNeSt baseline that SCA modifies."},{"cited_title":"IEEE Transactions on Geoscience and Remote Sensing57(1), 574–586 (2018)","cited_arxiv_id":null,"evidence_quote":"Defines the WHU Building Dataset, the primary evaluation benchmark."},{"cited_title":"IEEE Transactions on Geoscience and Remote Sensing (2023)","cited_arxiv_id":null,"evidence_quote":"BCTNet, the strongest transformer-based baseline that SCANet claims to surpass."},{"cited_title":"IEEE Geoscience and Remote Sensing Letters 20, 1–5 (2023)","cited_arxiv_id":null,"evidence_quote":"ConvBNet, a dedicated building extraction baseline used in the comparison."}],"review_version":1}