{"id":"bb5e2368-2c07-47a5-84ff-de1db5db2bc0","arxiv_id":"2501.06053","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CASS-Det, a YOLOX-based SAR ship detector with a center-enhancement module, a neighbor attention module, and a cross-connected feature pyramid network, achieves state-of-the-art mAP on SSDD, HRSID, and LS-SSDD.","lead":"A new detector for synthetic aperture radar images uses three modifications: rotated convolutions to emphasize ship centers, an attention block for crowded scenes, and a modified feature pyramid for ships of very different sizes. It reports the best published accuracy on three SAR ship detection benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA margins are only 0.012–0.018 mAP, but the paper never shows competitor numbers were produced under the same split and evaluation code; a protocol mismatch could erase the claimed gains.","rationale":"The reader's conditional verdict is well-calibrated. I considered three candidate weak points: (i) the rotational-convolution identity in Eq. (2) is algebraically wrong as written—the equality Ci * Rθ f = Rθ Ci * f does not follow from Eq. (1) and is not a valid convolution property in general; (ii) the ablation baseline in Table IV is labeled 'FPN' but the numbers match the PAFPN row in Table V, so the textual claim that the baseline is CSPDarknet+FPN is inconsistent; (iii) the comparison tables lack any statement of how competitor numbers were obtained and no error bars are reported. Of these, (iii) bears most directly on the paper's headline: the claimed advantages are 0.012, 0.018, and 0.018 mAP, which are the same order of magnitude as typical cross-paper protocol differences on these benchmarks. If the baselines were transcribed under different splits or evaluation code, the SOTA claim is not established even though CASS-Det may be a perfectly good detector. The Eq. (2) error is a genuine correctness issue in the method's derivation, but it does not by itself falsify the empirical comparison; the ablation labeling similarly obscures but does not overturn the result. I therefore agree with the reader's weakest_assumption and recommend keeping the verdict CONDITIONAL, with the concrete check being a single re-run of the strongest baseline under the paper's stated protocol.","tokens_in":16240,"tokens_out":9149,"duration_ms":86112,"concrete_test":"Reproduce YOLO v7 on SSDD using exactly the protocol in §IV-A: suffix-1/9 test split, images resized to 512×512, 500 epochs, Adam, batch size 8, and the same mAP script used to produce CASS-Det's 0.992. If the reproduced YOLO v7 mAP differs from Table I's 0.980 by more than 0.012, or if re-running the remaining baselines changes the rankings, the claimed SOTA margins are not established until a single unified protocol is applied to all methods.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim is that CASS-Det exceeds the second-best detector by 0.012, 0.018, and 0.018 mAP on SSDD, HRSID, and LS-SSDD. This claim is only meaningful if every entry in Tables I–III was obtained under the same data split, resizing, training budget, and mAP evaluation code. Section IV-A describes the authors' protocol (suffix-based SSDD split, HRSID 0.65/0.35 split, LS-SSDD 6000/600 sub-images, 512×512 or 800×800 resizing, 500 epochs, Adam, batch size 8) but does not state whether competitor numbers were re-run under this protocol or copied from their original papers. The margins are small enough that routine protocol differences—e.g., SSDD random 70/30 instead of suffix split, HRSID official split instead of 0.65/0.35, VOC mAP@0.5 versus COCO-style mAP, or different NMS settings—could change a baseline's score by more than 0.012–0.018. Without a provenance statement, per-run details, or released evaluation code, the headline comparison cannot be independently validated from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CASS-Det, a one-stage SAR ship detector built on CSPDarknet and YOLOX, and attributes its performance to three modules: CEM, based on rotational convolution to strengthen ship centers; NAM, a neighbor cross-attention module fusing adjacent feature levels; and CC-FPN, a cross-connected feature pyramid with extra shallow/deep fusion. Experiments are reported on SSDD, HRSID, and LS-SSDD-v1.0 with mAP 0.992, 0.931, and 0.821, beating the second-best methods by 0.012, 0.018, and 0.018. The paper also provides per-module ablations, FPN/PAFPN/CC-FPN comparisons, NAM versus In-NAM analysis, and transfer experiments to YOLOv7 and YOLOv8.","tokens_in":16420,"tokens_out":8647,"duration_ms":80601,"significance":"If the reported results hold, CASS-Det would be a useful modular detector for dense, multi-scale SAR ship detection, with practical engineering value: the modules are simple enough to be inserted into several YOLO-style frameworks, and the evaluation covers three widely used datasets with scale-aware AP metrics. The ablation design is a strength, and the transfer tests give some evidence of generality. However, the headline state-of-the-art margins are small, and the manuscript currently lacks the protocol-uniformity evidence needed to validate them; in addition, the CEM derivation contains a mathematical error and the ablation baseline is inconsistent between the text and Table V. These are load-bearing issues for the central claims, so the paper needs major revision before it can be accepted.","major_comments":[{"comment":"Equation (2) is not a consequence of Eq. (1). Equation (1) states a * [R b] = R^{-1}[(R a) * b], so the correct identity for the CEM branches is Concat(C_i * R^{k pi/2} f) = Concat(R^{-k pi/2}(R^{k pi/2} C_i) * f), with the inverse rotation applied to each branch. As written, the manuscript drops the inverse rotation and equates the left-hand side to Concat(R^{k pi/2} C_i * f), which is wrong except for k=2. If the implementation rotates kernels rather than feature maps, the left-hand side of Eq. (2) should be presented as the definition and the equality to the right-hand side should be removed. As written, the derivation does not support the rotation-equivariance claim for CEM.","section":"III-A, Eq. (2)"},{"comment":"The dimensions in the NAM equations are not reconciled. Q and K are computed from F_j in R^{C x H x W}, while F_{j+1} is in R^{2C x H/2 x W/2}. Equation (7) adds H to F_{j+1} through a 3x3 convolution, but no stride or resampling is specified for that convolution; a stride-1 3x3 convolution would preserve the H x W spatial size of H and cannot be added to an H/2 x W/2 tensor. The channel dimensions are also unspecified for the addition to F_{j+1}. Please specify the stride, padding, and output channels for the Conv3x3(H) term and for the DeConv3x3 operator in Eq. (4); otherwise the NAM equations are not implementable as written.","section":"III-B, Eqs. (4)-(7)"},{"comment":"The paper does not state whether all competitor results in Tables I-III were obtained by re-running the comparison methods under CASS-Det's protocol, or were transcribed from the original papers. Section IV-A describes the authors' protocol (SSDD suffix-based split, HRSID 0.65/0.35 split, LS-SSDD 6000/600 sub-images, 512x512 or 800x800 resizing, 500 epochs, Adam, batch size 8), but no equivalent provenance is given for the baselines. The reported margins over the second-best methods are only 0.012-0.018 mAP, so differences in data split, mAP definition, NMS settings, input resolution, or training budget could change baseline scores by more than these margins. The manuscript needs a clear provenance statement for every comparison entry, and ideally re-run baselines or variance estimates, before the state-of-the-art claim can be validated.","section":"IV-A, Tables I-III"},{"comment":"The ablation baseline is inconsistent between the text and the tables. Section IV-D states that the ablation baseline is the combination of CSPDarknet and FPN, but the baseline row of Table IV (mAP 0.977, F1 0.946, recall 0.928, precision 0.965, APs 0.543, APm 0.663, APl 0.634) is identical to the PAFPN row of Table V, not to the FPN row of Table V (mAP 0.970, F1 0.925, recall 0.947, precision 0.904, APs 0.570, APm 0.690, APl 0.657). Consequently, the ablation increments attributed to CEM, NAM, and CC-FPN are computed against PAFPN while being described as FPN improvements. This mismatch must be corrected, and the textual claims about FPN baselines should be re-derived from the correct table entries.","section":"IV-D, Tables IV-V"}],"minor_comments":[{"comment":"Equation (10) writes DConv1x1, while the text immediately after the equation says that DConv 3x3 represents a 3x3 dilated convolution; a 1x1 dilated convolution is not standard, so the notation should be aligned.","section":"III-C, Eq. (10)"},{"comment":"Section IV-A says 'Experiments in Section III are all implemented' but the experiments are described in Section IV; similarly, the last paragraph of Section IV-C says the comparison is shown in 'Section III-C', which should be Section IV-C.","section":"IV-A and IV-C"},{"comment":"The first sentence of Section IV-D refers to the proposed modules as 'CEM, NAM, and CC-Net'; this should be CC-FPN for consistency with the rest of the paper.","section":"IV-D"},{"comment":"For LS-SSDD, only mAP and precision are reported, while Tables I-II also include F1, recall, and scale-wise AP; reporting the same metric set would make the cross-dataset comparison complete.","section":"Table III"},{"comment":"Equation (5) computes an HW x HW affinity matrix, which can be prohibitively large at the higher input resolutions used in this paper; please report where NAM is inserted, the spatial size at which the attention is computed, and the resulting memory/computation overhead.","section":"III-B"},{"comment":"The caption contains a typo: 'our medules' should be 'our modules'.","section":"Fig. 10 caption"}],"recommendation":"major_revision","confidential_remarks":"The conditional verdict from the reader is appropriate. The protocol-uniformity question is the most serious threat to the state-of-the-art claim, and the ablation baseline mismatch is a clear correctness issue that the authors can fix by re-deriving the affected numbers. The mathematical error in Eq. (2) is localized and fixable. I see no reason to suspect misconduct, but the manuscript needs substantial technical clarification before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper reports state-of-the-art results on three SAR ship detection benchmarks, and the integrated system (CASS-Det) is new. The work is a solid engineering effort, but the central claim rests on a comparison protocol that the manuscript never describes, and the math behind one module is wrong as written.\n\nWhat is genuinely new: the combination of the three modules—CEM, NAM, and CC-FPN—in a YOLOX-style detector, evaluated on SSDD, HRSID, and LS-SSDD-v1.0. The ablations are fairly complete, including a transfer test to YOLOv7 and YOLOv8. The modules are motivated by real SAR challenges, and the writing is mostly readable. I have no reason to think the experiments were faked; the numbers are internally consistent.\n\nThe soft spots, in order of severity. First, the baseline comparison is the load-bearing issue. The authors describe their own training splits and evaluation, but never state whether the numbers for YOLOv7, YOLOv8, FCOS, etc. were re-run under the same protocol or transcribed from previous papers. Since the reported margins are only 0.012–0.018 mAP, a difference in data split or mAP computation could erase them entirely. This needs a clear provenance statement, ideally a unified re-run plus code release. Second, Eq. (2) drops the inverse rotation that Eq. (1) says is required. As written, the CEM derivation is mathematically incorrect; the authors should fix it or explain why alignment is preserved. Third, the NAM's query and key come from the same feature map, even though the text promises cross-layer dependencies. The cross-layer part is only in the value path. That's overstating the mechanism. Fourth, the ablation baseline in Table IV is labeled FPN but its numbers match PAFPN in Table V. The HRSID text also says 'perfect' for non-perfect metrics and mixes up the AP_s/AP_m/AP_l order. These are fixable but embarrassing. Fifth, no error bars or repeated runs, which matters more because the gains are small.\n\nRecommendation: this deserves a serious referee. The topic is useful, the system is new, and the empirical claim is testable. But it needs major revision before acceptance: fix the math, clarify the baselines, correct the labeling and overstatements, and release code/weights. If the authors cannot establish that their comparisons are apples-to-apples, the SOTA claim should be softened to 'competitive with published numbers.'","headline":"Useful SAR ship detector with strong benchmark numbers, but the comparison protocol is unverifiable and Eq. (2) has a clear math error; worth a serious referee after major revision.","tokens_in":17016,"tokens_out":6795,"would_cite":false,"duration_ms":59557,"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":"CASS-Det, a one-stage SAR ship detector, reports state-of-the-art mAP of 0.992 on SSDD, 0.931 on HRSID, and 0.821 on LS-SSDD by combining center enhancement, neighbor attention, and cross-connected feature fusion.","keywords":["synthetic aperture radar","SAR ship detection","center enhancement module","rotational convolution","neighbor attention","cross-connected feature pyramid","multi-scale detection","dense target detection"],"falsifier":"Re-run CASS-Det and the three strongest competing detectors on SSDD, HRSID, and LS-SSDD using one shared evaluation script, the same train/test splits, the same image resizing, and the same COCO-style mAP code; if the reported 0.012–0.018 mAP margins shrink to noise or reverse, the central claim would be falsified.","tokens_in":15976,"feed_emoji":"🛰️","tokens_out":10398,"duration_ms":87727,"temperature":0.7,"pith_summary":"The paper proposes CASS-Det, a one-stage object detector for synthetic aperture radar (SAR) ship images, and claims it outperforms all compared detectors on three public benchmarks. A one-stage detector predicts ship boxes directly from feature maps, without a separate proposal step. CASS-Det's reported gains come from three added modules: a center enhancement module that highlights ship centers through rotated convolution, a neighbor attention module that refines boundaries between touching ships using cross-level long-range dependencies, and a cross-connected feature pyramid that fuses shallow and deep features across more scales. The reported margins over the second-best method are 0.012, 0.018, and 0.018 mAP on SSDD, HRSID, and LS-SSDD, respectively. A sympathetic reading treats these results as evidence that explicitly targeting center localization, dense-target separation, and scale variation in one architecture pushes SAR detection accuracy forward.","feed_headline":"SAR ship detector beats previous best on three benchmarks","feed_subtitle":"CASS-Det's three modules separate tightly packed ships and span small to large scales in radar imagery.","key_machinery":"The central mechanism is the rotational-convolution identity $[a \\ast [R b]](x) = R^{-1}[(R a) \\ast b](x)$, which lets CEM obtain multi-orientation responses by rotating the input feature map by $k\\pi/2$ for $k=0,1,2,3$, convolving once, and concatenating the four rotated views before a $1\\times1$ convolution. NAM builds queries and keys from a shallower feature level and values from a deconvolved deeper level, applies a softmax affinity matrix, and adds the aggregated result back to both inputs, propagating global context across adjacent pyramid levels to refine boundaries between touching ships. CC-FPN adds a direct shallow connection, dilated convolutions, concatenations with higher-level maps, and an extra $P_6$ level so small and large ships both get strong features.","core_discovery":"CASS-Det's central claim is that the three main failure modes of SAR ship detection—cluttered backgrounds, densely packed inshore ships, and extreme scale variation—can be addressed jointly by modular additions to a standard CSPDarknet backbone with a YOLO-series detection head. The center enhancement module uses the rotational-convolution property to reinforce ship centers while suppressing clutter. The neighbor attention module computes long-range affinities between adjacent pyramid levels so that touching ships are separated. The cross-connected feature pyramid extends the standard fusion structure with a direct shallow-level connection, dilated convolutions, and an extra top level $P_6$, giving both small and large ships strong features. On SSDD, HRSID, and LS-SSDD-v1.0, the full model reaches 0.992, 0.931, and 0.821 mAP, exceeding the second-best compared detector by 0.012, 0.018, and 0.018.","pith_inferences":["The center-versus-boundary decomposition is not SAR-specific; a natural test is applying CEM and NAM to optical aerial imagery where objects touch, such as parked vehicles or roof segments, and measuring whether recall on dense scenes rises similarly.","The paper does not isolate whether CEM's gain comes from rotational convolution itself or simply from multi-view concatenation; an ablated variant that feeds the same four rotated inputs through standard convolutions would separate those effects.","NAM's full affinity matrix scales as $HW \\times HW$ per level pair, so the module may need approximation for very large images; a testable extension is replacing full affinities with criss-cross or local windows and comparing the mAP-versus-compute trade-off.","The reported margins are conditional on a shared evaluation protocol; re-implementing the baselines under one codebase with identical splits and resizing would reveal whether the 0.012–0.018 gaps persist."],"forward_implications":["CASS-Det reports the highest mAP on all three datasets, with the largest absolute gains on HRSID and LS-SSDD, the two benchmarks dominated by small ships.","The ablation study supports a cumulative design: CEM adds recall, NAM adds recall and large-ship average precision, and CC-FPN adds precision and large-ship average precision, together moving the SSDD baseline from 0.977 to 0.992 mAP.","The module transfer experiment shows recall rising by 4.3 to 5 percentage points when CEM, NAM, and CC-FPN are added to existing YOLO-style detectors on SSDD.","CC-FPN outperforms both FPN and PAFPN under the same backbone, improving SSDD mAP from 0.970 and 0.977 to 0.980 and raising large-ship average precision by 5.8 points over PAFPN."],"supporting_citations":[{"why":"Supplies the SSDD benchmark and its official suffix-based train/validation/test split used for the main SSDD results.","marker":"[31]"},{"why":"Supplies the HRSID benchmark whose dense small-ship scenes are used for the HRSID results and the 0.65:0.35 split.","marker":"[32]"},{"why":"Supplies the LS-SSDD-v1.0 large-scale Sentinel-1 benchmark and the 6000/600 sub-image protocol used for the LS-SSDD results.","marker":"[33]"},{"why":"YOLOX is the detection head and training framework that CASS-Det is built on and the baseline it improves.","marker":"[58]"},{"why":"FPN is the standard multi-scale fusion design that the paper uses as a baseline and extends with CC-FPN.","marker":"[29]"},{"why":"PAFPN is the neck used in the baseline and in competing YOLO detectors, and is the direct comparison for CC-FPN.","marker":"[30]"},{"why":"YOLO v7 is the strongest competing detector on SSDD and one of the frameworks used to test module transferability.","marker":"[60]"},{"why":"YOLO v8 is the strongest competing detector on HRSID and the other framework used in the module-transfer experiment.","marker":"[57]"}],"fun_headline_variants":["CASS-Det: one model, three modules, best SAR ship detection","New SAR detector tops three benchmarks for ship spotting","Center-aware modules sharpen SAR ship detection across scales","Rotational convolution helps CASS-Det master dense ship scenes","CASS-Det fuses features to catch ships from small to large"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central SOTA claim depends on the assumption that all comparison methods in Tables I–III were evaluated under the same data splits, image resizing, and mAP computation as CASS-Det, since the paper does not state whether competitor numbers were re-run locally or transcribed from their original papers.","fun_headline_variants_meta":{"raw":{"variants":["CASS-Det: one model, three modules, best SAR ship detection","New SAR detector tops three benchmarks for ship spotting","Center-aware modules sharpen SAR ship detection across scales","Rotational convolution helps CASS-Det master dense ship scenes","CASS-Det fuses features to catch ships from small to large"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000519,"raw_usage":{"total_tokens":2510,"prompt_tokens":936,"completion_tokens":1574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1489}},"tokens_in":552,"tokens_out":1574,"duration_ms":11623,"temperature":1.0,"reasoning_tokens":1489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:15.349100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run CASS-Det and the three strongest competing detectors on SSDD, HRSID, and LS-SSDD using one shared evaluation script, the same train/test splits, the same image resizing, and the same COCO-style mAP code; if the reported 0.012–0.018 mAP margins shrink to noise or reverse, the central claim would be falsified.","supporting_citations":[{"cited_title":"Ship detection in sar images based on an improved faster r-cnn,","cited_arxiv_id":null,"evidence_quote":"Supplies the SSDD benchmark and its official suffix-based train/validation/test split used for the main SSDD results."},{"cited_title":"Hrsid: A high-resolution sar images dataset for ship detection and instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the HRSID benchmark whose dense small-ship scenes are used for the HRSID results and the 0.65:0.35 split."},{"cited_title":"Ls-ssdd-v1.0: A deep learning dataset dedicated to small ship detection from large-scale sentinel-1 sar images,","cited_arxiv_id":null,"evidence_quote":"Supplies the LS-SSDD-v1.0 large-scale Sentinel-1 benchmark and the 6000/600 sub-image protocol used for the LS-SSDD results."},{"cited_title":"Feature pyramid networks for object detection,","cited_arxiv_id":null,"evidence_quote":"FPN is the standard multi-scale fusion design that the paper uses as a baseline and extends with CC-FPN."},{"cited_title":"Path aggregation network for instance segmentation,","cited_arxiv_id":null,"evidence_quote":"PAFPN is the neck used in the baseline and in competing YOLO detectors, and is the direct comparison for CC-FPN."},{"cited_title":"Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,","cited_arxiv_id":null,"evidence_quote":"YOLO v7 is the strongest competing detector on SSDD and one of the frameworks used to test module transferability."}],"review_version":1}