{"id":"e994f2de-84b2-4dc6-a9bd-12d32601607b","arxiv_id":"2501.00360","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A shape-guided transformer with row and column global attention reaches state-of-the-art instance segmentation accuracy on three remote sensing datasets.","lead":"This paper introduces SGTN, a deep learning network that detects and outlines individual objects in satellite and aerial images. It combines a transformer that connects far-apart features with a module that sharpens object boundaries, and reports the highest scores on three public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'highest AP' claim rests on a 0.2 AP margin over an in-house CenterMask reimplementation with no error bars; if that reimplementation is even slightly undertuned, the headline result is not established.","rationale":"I read the paper in good faith. The architecture is plausible, the ablations are consistent, and the method likely works. The reader's weakest_assumption is exactly the comparison fairness and stability, and I agree that this is the most load-bearing uncertainty. I considered the internal efficiency contradiction (LSwin runtime 66.3s vs Swin-S 56.0s in Table V contradicts the 'same level of efficiency' claim), but that would weaken a secondary claim, not the headline SOTA claim. The headline claim is what the abstract and introduction emphasize, and it depends on the 0.2 AP margin being real and stable. Since the authors reimplemented all baselines and give no variance estimates, the claim is not yet established. A conditional verdict with a request for official-baseline comparison and repeated runs is appropriate; the result does not require rejection.","tokens_in":17791,"tokens_out":6663,"duration_ms":63757,"concrete_test":"Run the WHU comparison (9420/1537/3848 split) with the official CenterMask implementation from its authors, using the same augmentation and 30-epoch schedule as SGTN, and repeat both CenterMask and SGTN five times with different random seeds. Report the mean and standard deviation of AP for each method. If official CenterMask reaches 72.4 or the 0.2 gap is within one standard deviation, the 'highest AP' claim on WHU is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B.2 states that for a fair comparison all baselines were re-implemented in PyTorch by the authors, rather than using official released code or published checkpoints. The principal SOTA evidence is Table I: SGTN with ResNet-50 reaches 72.4 AP on WHU, while the second-best CenterMask reaches 72.2 AP, a 0.2 margin. The ablation in Table IV attributes 1.3 AP on WHU to the SGM module, so the entire winning margin is far smaller than the effect of a single component. No repeated runs, standard deviations, or significance tests are reported for any table, so it is impossible to distinguish a genuine improvement from run-to-run noise or small differences in training schedules, data augmentation, or hyperparameters. Because the comparison is in-house, a plausible alternative explanation is that the baselines are slightly under-tuned relative to their published configurations, which would erase the 0.2 margin and invalidate the abstract's claim of the highest AP. This concern applies to the BITCC and NWPU VHR-10 results too, though their margins are larger (1.0 and 2.8 AP); the WHU result is the fragile one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SGTN, a two-stage instance segmentation network for remote sensing images, combining a new transformer encoder LSwin and a shape guidance module (SGM). LSwin augments Swin-S with vertical and horizontal 1D global self-attention blocks, with learnable fusion weights α and β. SGM produces category-agnostic foreground, edge, and corner predictions from a fused shallow-and-deep feature, and its foreground map is used to refine predicted instance masks via a Hadamard product. The authors report state-of-the-art AP on WHU, BITCC, and NWPU VHR-10, and present ablations showing consistent gains from both LSwin and SGM across all three datasets.","tokens_in":18077,"tokens_out":3928,"duration_ms":40204,"significance":"If the empirical claims are reliable, the combination of local shifted-window attention with 1D global self-attention is a practical way to improve global context modeling for remote sensing instance segmentation without the full cost of 2D global attention, and the SGM's foreground-refinement mechanism is a simple and transferable module. The ablations are systematic and consistent across three datasets, which lends internal support to the contribution of each component. However, the significance is currently limited by the fragility of the headline results: the main SOTA margin on the WHU dataset is 0.2 AP against an in-house reimplementation of CenterMask, no variance or significance information is provided, the runtime claim for LSwin is contradicted by Table V, and no code or official baseline checkpoints are made available to reproduce the comparison.","major_comments":[{"comment":"The central claim of achieving the highest AP on the WHU dataset rests on a 0.2 AP margin over an in-house reimplementation of CenterMask (72.4 vs 72.2, Table I), and no standard deviations, repeated runs, or significance tests are reported for any result. The WHU ablation in Table IV attributes 1.3 AP to the SGM alone, so the entire cross-method winning margin is smaller than the effect of a single component; this makes the SOTA claim vulnerable to run-to-run noise, data augmentation differences, or slight under-tuning of the reimplemented baselines. Please provide multiple-seed runs with variance estimates for at least the main comparisons, and either use official released code or checkpoints for all baselines or justify in detail that the in-house implementations reproduce the published configurations.","section":"Section IV-B.2, Tables I–III"},{"comment":"The abstract and Section IV-D state that LSwin is better than ResNet and Swin at the same level of efficiency, but Table V reports 66.3 sec for SGTN with LSwin versus 52.6 sec with ResNet-50 and 56.0 sec with Swin-S on the BITCC dataset, which is roughly a 26% increase over the ResNet-50 variant. The text that describes these numbers as 'almost equally efficient' is an overstatement. Please either correct the efficiency claim to reflect the measured runtime differences or provide an explanation of the throughput conditions under which the claim holds.","section":"Section IV-F.1 and Table V"},{"comment":"The mask refinement step Mi = Mc ⊙ Ms is presented as a key use of the SGM, and the ablation in Table IV measures the effect of the whole SGM including the auxiliary losses and the fusion rule together. Since the whole-module gain of 1.3 AP on WHU exceeds the 0.2 AP SOTA margin, the paper should isolate whether the improvement comes from the multi-label shape supervision, the Hadamard-product refinement, or the simple concatenation of shallow and deep features. An additional ablation removing only the mask-refinement step (keeping the auxiliary losses) would clarify this and make the design justification load-bearing rather than implicit.","section":"Section III-A and Table IV"}],"minor_comments":[{"comment":"The 1D vertical/horizontal self-attention idea is introduced as a 'new paradigm' and credited only to the authors' previous work [8], but similar axial/1D attention schemes have been used in vision transformers; the related work discussion should acknowledge this prior art and position LSwin as an integration rather than a wholly new mechanism.","section":"Section II-C and III-B"},{"comment":"The notation in equations (1) and (2) is inconsistent (for example, z^l_sw and z_lr^l+1 are used without a shared convention for layer indexing), which makes the block equations difficult to follow; please unify the superscript and subscript notation.","section":"Section III-B, Eqs. (1)–(2)"},{"comment":"The text says 'YOLCAT' instead of 'YOLACT' in the WHU experiment description, and the reference numbers in Table II do not match the bibliography order (e.g., YOLACT is cited as [13] instead of [14]); these should be corrected.","section":"Section IV-C.1"},{"comment":"The caption reads 'from SG-LRN', which appears to be a leftover from a previous draft and should be replaced with 'from SGTN'.","section":"Fig. 13 caption"},{"comment":"The subsection title says 'Long-Range Correction blocks', but the term used elsewhere is 'Long-Range Correlation'; please make the terminology consistent.","section":"Section IV-F.2"},{"comment":"The manuscript promises 'Code will be available' but no code or model checkpoints are provided; given that all baselines were reimplemented in-house, public code would be particularly important for verifying the comparison.","section":"Section IV-B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's main empirical claim is plausible but not yet demonstrated convincingly because the SOTA margin on WHU (0.2 AP) is within the range of typical run-to-run variation for instance segmentation and the baselines are all reimplemented in-house. The authors should be encouraged to add multi-seed results, use official baseline code, or soften the 'highest AP' claim. The efficiency contradiction in Table V is also a substantive inconsistency. I do not see grounds for rejection, since the ablations are internally consistent and the components are reasonable, but the required evidence is substantial enough to warrant a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid applied paper, not a breakthrough. The combination of a 1D axial-attention-boosted Swin encoder (LSwin) and a shape guidance module (SGM) gives consistent AP gains across three remote sensing benchmarks, and the ablations are thorough enough to show both components earn their place. I especially like the transfer experiments where LSwin improves CenterMask and DANCE — that suggests the encoder is genuinely useful beyond the authors' own head.\n\nWhat's new is modest but real. The LSwin block interleaves standard shifted-window attention with vertical and horizontal 1D self-attention, which is a sensible way to get global context without quadratic cost. SGM adds foreground/edge/corner supervision and then uses the foreground map to refine predicted masks. Neither idea is wildly original, but the specific combination and the careful ablation across three datasets is a legitimate contribution for the remote sensing instance segmentation community.\n\nNow the soft spots. The headline 'highest AP' on WHU is fragile: SGTN ResNet-50 beats CenterMask by 0.2 AP. All baselines were re-implemented in-house, no official checkpoints or code are released, and there are no error bars or multiple-seed runs. That 0.2 margin could easily vanish with a slightly better-tuned baseline. The BITCC and NWPU margins (1.0 and 2.8 AP) are more believable, so the paper's overall case is not riding on the one fragile number, but the abstract's absolute claim is. The efficiency statement also conflicts with Table V: LSwin is 66.3s vs Swin-S 56.0s, so saying it is 'at the same level of efficiency' is misleading. Minor issue: the 1D attention idea is credited only to the authors' own prior work, omitting earlier axial attention literature; that inflates novelty but doesn't affect the architecture's validity.\n\nWho is this for? People working on instance segmentation in aerial/satellite imagery who want a ready-to-try architecture for a few extra points of AP. It deserves a serious referee, not because it breaks new conceptual ground, but because the engineering is sound, the ablations are informative, and the claimed results are plausible pending verification.\n\nMy recommendation: send it to peer review, but require the authors to release code (or at least training configuration details), report multiple runs or error bars, and fix the efficiency claim. With those changes this is a citable applied paper.","headline":"Useful applied architecture for RSI instance segmentation with consistent ablations, but the SOTA claim rests on a 0.2 AP margin over in-house reimplemented baselines with no error bars.","tokens_in":18553,"tokens_out":1700,"would_cite":true,"duration_ms":18217,"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":"This paper claims that a two-stage instance segmentation network, SGTN, combining a long-range-correlation Swin-style encoder (LSwin) and a shape guidance module, achieves the highest mask average precision on the WHU, BITCC, and NWPU…","keywords":["instance segmentation","remote sensing images","transformer encoder","long-range correlation","shape guidance","self-attention","building instance segmentation","average precision"],"falsifier":"Re-run the same baselines with the authors' training setup and multiple random seeds, or with their published official code and checkpoints; if CenterMask or DANCE matches or exceeds SGTN's average precision on WHU, BITCC, or NWPU VHR-10 once noise is accounted for, the claimed superiority is not established. A narrower test is to ablate Swin-S versus LSwin inside SGTN over several seeds and check whether the reported 0.5 to 2.2 average-precision gaps exceed the run-to-run spread.","tokens_in":17624,"feed_emoji":"🛰️","tokens_out":10324,"duration_ms":93567,"temperature":0.7,"pith_summary":"This paper is trying to establish that instance segmentation in remote sensing images can be improved by attacking two bottlenecks at once: relating instances scattered across large scenes, and recovering their boundaries and shapes under atmospheric degradation. The proposed SGTN combines a transformer encoder called LSwin with a shape guidance module; LSwin adds vertical and horizontal one-dimensional global self-attention to the local shifted-window Swin Transformer, and the shape module supervises class-agnostic foreground, edge, and corner maps and then uses the foreground map to refine predicted instance masks. The paper reports the best average precision scores on the WHU, BITCC, and NWPU VHR-10 datasets against all compared methods, with ablations giving consistent gains to both components. A sympathetic reader would care because the ingredients are inexpensive: the added attention is linear, and the shape maps come from simple supervised heads rather than a new segmentation paradigm.","feed_headline":"Shape-guided transformer network tops three public benchmarks","feed_subtitle":"New encoder adds global 1-D attention while a shape module sharpens masks, raising AP on WHU, BITCC, and NWPU VHR-10.","key_machinery":"The load-bearing objects are the LSwin block and the shape guidance module. A LSwin block interleaves two ordinary Swin Transformer blocks with two long-range correlation blocks; the correlation blocks perform self-attention along full rows (H-MSA) and full columns (V-MSA) instead of within local 2-D windows, with instance normalization inside each 1-D window. The two streams are fused as $\\alpha z_{sw} + \\beta z_{lr}$, with $\\alpha=1$ and $\\beta=0$ initially so that the pretrained Swin weights dominate at the start. The shape guidance module starts from a shallow detail-rich feature, predicts class-agnostic foreground, edge, and corner maps under three binary cross-entropy losses (edge pixels doubled, corner pixels quadrupled in weight), and then refines each resized instance mask by the Hadamard product with the cropped foreground map, $M_i = M_c \\odot M_s$. These mechanisms carry the argument: one supplies global context at linear cost, the other supplies boundary-preserving detail.","core_discovery":"The central claim is that local detail and global context are complementary and can be fused in one end-to-end network. LSwin is built on Swin-S by interleaving long-range correlation blocks with the original shifted-window blocks; each long-range block runs self-attention along rows (H-MSA) and along columns (V-MSA) with instance normalization, giving explicit cross-image interaction at linear cost, and the two block streams are combined with learnable coefficients $\\alpha$ and $\\beta$ initialized to $1$ and $0$. In parallel, the shape guidance module concatenates a shallow detail-rich feature with the encoder feature, supervises three binary maps (foreground, edges, corners) with weighted binary cross-entropy, and refines each RoI-aligned mask by element-wise multiplication with the predicted foreground map, $M_i = M_c \\odot M_s$. The paper claims this combination yields the highest mask AP on two single-class building datasets and one ten-class dataset, with LSwin improving over ResNet-50 and Swin-S at similar efficiency and with SGM adding 1.1 to 1.4 AP points in ablations.","pith_inferences":["One untested consequence of the linear attention design is that LSwin's relative advantage over Swin should grow as tile size grows, because 1-D global attention scales linearly while the number of windows grows; running the same comparison on larger crops would test this directly.","Because the shape supervision is category-agnostic, the foreground/edge/corner maps may transfer across datasets and even to new object classes without retraining; the paper does not report such cross-dataset experiments.","The Hadamard fusion rule is a hard gate: it can only remove pixels from the resized mask, never add them. A learned or soft fusion could reveal how much of SGM's gain comes from the gate versus the auxiliary supervision itself."],"forward_implications":["LSwin is a drop-in encoder: the paper shows it raises average precision for CenterMask and DANCE as well as for SGTN, so the global-attention design transfers to other instance segmentation frameworks.","The two attention paradigms are complementary: LSwin beats both Swin-S and a pure long-range-correlation encoder on all three datasets, which implies that local window attention and 1-D global attention are not substitutes.","SGM's gains concentrate at high IoU thresholds (AP75 up 0.8, 0.9, and 2.1 percentage points), so the shape module is improving boundary quality rather than merely adding detections.","With the same ResNet-50 backbone, SGTN still beats the second-best CenterMask by 0.2, 1.0, and 2.8 AP on the three datasets, so the reported advantage is not solely an effect of the transformer backbone."],"supporting_citations":[{"why":"Defines the Swin Transformer whose Swin-S blocks, patch partitioning, and pretrained weights LSwin extends and inherits.","marker":"[10]"},{"why":"Supplies the Mask R-CNN two-stage framework, RoI Align, and the instance mask and box heads that SGTN reuses.","marker":"[18]"},{"why":"Introduces the long-range correlation paradigm of decomposing 2-D self-attention into 1-D vertical and horizontal correlations, and provides the ARFEM module used by SGM.","marker":"[8]"},{"why":"Provides the anchor-free center-based candidate box generation module used in the first stage.","marker":"[49]"},{"why":"Inspires the use of a global-scale class-agnostic foreground classification map to refine instance masks.","marker":"[5]"},{"why":"Supplies the NWPU VHR-10 multi-class dataset and its evaluation protocol.","marker":"[28]"},{"why":"Supplies the WHU building dataset and its train/validation/test split.","marker":"[53]"},{"why":"Supplies the BITCC building dataset and its partition rule.","marker":"[54]"}],"fun_headline_variants":["Global attention plus shape module lifts remote sensing masks","LSwin and shape guidance set new AP records on three datasets","Global 1-D attention plus shape module improves remote instance masks","Shape-guided transformer beats three public datasets","Instance segmentation in remote sensing gets global-local boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline ranking rests on the authors' re-implementations of every baseline being at least as carefully tuned as their own method, since several winning margins are under half a percentage point of average precision and no repeated runs or error bars are reported.","fun_headline_variants_meta":{"raw":{"variants":["Global attention plus shape module lifts remote sensing masks","LSwin and shape guidance set new AP records on three datasets","Global 1-D attention plus shape module improves remote instance masks","Shape-guided transformer beats three public datasets","Instance segmentation in remote sensing gets global-local boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3092,"prompt_tokens":1050,"completion_tokens":2042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":1967}},"tokens_in":666,"tokens_out":2042,"duration_ms":15137,"temperature":1.0,"reasoning_tokens":1967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:52:43.457182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same baselines with the authors' training setup and multiple random seeds, or with their published official code and checkpoints; if CenterMask or DANCE matches or exceeds SGTN's average precision on WHU, BITCC, or NWPU VHR-10 once noise is accounted for, the claimed superiority is not established. A narrower test is to ablate Swin-S versus LSwin inside SGTN over several seeds and check whether the reported 0.5 to 2.2 average-precision gaps exceed the run-to-run spread.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Swin Transformer whose Swin-S blocks, patch partitioning, and pretrained weights LSwin extends and inherits."},{"cited_title":"Accurate Instance Segmentation for Remote Sensing Images via Adaptive and Dynamic Feature Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Mask R-CNN two-stage framework, RoI Align, and the instance mask and box heads that SGTN reuses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the long-range correlation paradigm of decomposing 2-D self-attention into 1-D vertical and horizontal correlations, and provides the ARFEM module used by SGM."},{"cited_title":"OEC-RNN: Object-oriented delineation of rooftops with edges and corners using the recurrent neural network from the aerial images,","cited_arxiv_id":null,"evidence_quote":"Provides the anchor-free center-based candidate box generation module used in the first stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the use of a global-scale class-agnostic foreground classification map to refine instance masks."},{"cited_title":"Contour Loss: Boundary-Aware Learning for Salient Object Segmentation","cited_arxiv_id":"1908.01975","evidence_quote":"Supplies the NWPU VHR-10 multi-class dataset and its evaluation protocol."},{"cited_title":"Learning to aggregate multi-scale context for instance segmentation in remote sensing images,","cited_arxiv_id":null,"evidence_quote":"Supplies the WHU building dataset and its train/validation/test split."},{"cited_title":"GLSANet: Global-local self-attention network for remote sensing image semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the BITCC building dataset and its partition rule."}],"review_version":1}