{"id":"2c04fb9b-d04c-46cc-b14b-64a0ed1c4589","arxiv_id":"2411.13847","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-task SAR ship detector jointly trained with a Gaussian-mask segmentation task and a denoising module reports improved detection accuracy over five baselines on SSDD+ and HRSID.","lead":"This paper introduces MLDet, a multi-task deep learning system that detects ships in radar images by combining object detection, noise suppression, and target segmentation into one network. The authors report accuracy improvements over existing detectors on two standard SAR ship datasets, SSDD+ and HRSID.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim may rest on 137k HH/HV pretraining pairs whose source and split are undisclosed; neither dataset can supply them, so external data or test-set leakage could drive the gains.","rationale":"The reader's weakest assumption focused on registration quality of HH/HV pairs and on the missing description of DFF at test time. My concern is closely related but centers on data provenance and experimental fairness: the paper's internal numbers make it impossible for the 137,327 pairs to come from the described datasets, so either external data or test-set leakage is likely. This is more directly load-bearing for the central SOTA claim because the ablation study shows DFF contributes about five AP50 points, and if the DFF pretraining corpus is not available to baselines or overlaps with test images, the comparison in Table VIII is confounded. The correct response is to maintain the CONDITIONAL verdict: the empirical claims should be re-verified with a disclosed data-preparation pipeline and controlled ablations. I do not see grounds to reject outright, because the multitask framework itself is coherent and the code is promised, but the current evidence is insufficient to accept the SOTA claim as isolated.","tokens_in":21246,"tokens_out":5787,"duration_ms":56199,"concrete_test":"Obtain the released code and data-preparation scripts and identify the exact source, polarization channels, patchification, and train/test split of the 137,327 HH/HV pairs. Then run two controlled variants: (a) retrain MLDet with DFF pretrained only on the official training splits of SSDD+ and HRSID, ensuring no test images appear in the 137,327 pairs; (b) retrain MLDet from scratch without DFF pretraining while keeping all other multitask losses and the WRBF strategy. Compare AP50 on both datasets against Table VIII and against a YOLOX baseline retrained under identical data conditions. If either variant reduces the AP50 margin over YOLOX by more than a few points, the reported state-of-the-art claim is not attributable to the proposed architecture alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A states that 137,327 HH/HV pairs are used to train the denoised feature fusion (DFF) module, yet the paper describes SSDD+ as 1,160 images and HRSID as 5,604 images. Neither dataset can yield 137,327 pairs without extensive, unspecified patchification, and the paper never states the source, polarization channels, registration quality, or train/test split of these pairs. This creates two failure modes for the central SOTA claim in Section IV. First, if the pairs come from external SAR data (e.g., Sentinel-1), MLDet's DFF is pretrained on a large corpus unavailable to the YOLOX, RetinaNet, CenterNet, and Oriented R-CNN baselines in Table VIII, so the reported AP50 margins (+2.86 on SSDD+, +5.05 on HRSID) may reflect a data advantage rather than the proposed multitask architecture, ARW loss, or WRBF. Second, if the pairs are derived from the full SSDD+ or HRSID datasets including their test splits, the DFF has already seen test scenes during pretraining, leaking information into the frozen feature extractor and artificially inflating detection performance. The ablation tables attribute roughly five AP50 points to the DFF module, so the provenance of these pairs is directly load-bearing: without a controlled comparison in which baselines receive the same auxiliary data and no test-set contamination, the empirical evidence does not isolate the contribution of the proposed method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MLDet, a rotated bounding-box ship detector for SAR images that combines a CSPDarknet detection module with two auxiliary tasks: a denoised feature fusion (DFF) module trained on paired HH/HV SAR images, and a target segmentation (TS) module supervised by rotated Gaussian masks. It also introduces an angle classification loss with aspect-ratio weighting (ARW), a weighted rotated boxes fusion (WRBF) post-processing strategy, and reports experiments on SSDD+ and HRSID with ablations and comparisons against Faster R-CNN, RetinaNet, Oriented R-CNN, CenterNet, and YOLOX. The central claim is that the full MLDet model achieves state-of-the-art AP50 on both datasets.","tokens_in":21534,"tokens_out":4760,"duration_ms":49276,"significance":"If the central claim holds, the paper would provide a useful template for combining speckle suppression, segmentation, and rotated detection in SAR imagery. The strengths are that the experiments use two public benchmarks, standard detection metrics (AP50, AP75, precision, recall, F1), an ablation for each proposed module, and comparisons against five baselines. The paper also states that code is available. However, the load-bearing evidence depends on an undisclosed 137,327-pair HH/HV pretraining corpus, on the unclear test-time use of the denoising module, and on baseline re-implementations whose configurations are not given; these issues must be resolved before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The provenance of the 137,327 HH/HV training pairs for the DFF module is undisclosed and appears inconsistent with the described datasets. Section III-A states that SSDD+ has 1160 images and HRSID has 5604 images, so neither dataset can alone yield 137,327 corresponding pairs without extensive, unspecified patchification. Section III-C additionally says the DFF module is 'independently trained by the dataset SSDD+.' Because the ablations attribute approximately five AP50 points to the DFF module, this pretraining corpus is load-bearing for the state-of-the-art claim. The authors must report the source, polarization channels, registration procedure, patchification, and exact train/test split of these pairs. They must also rule out test-set leakage from SSDD+/HRSID and perform a controlled comparison in which baselines receive the same auxiliary data, or show that DFF still helps without this pretraining.","section":"Section III-A, Section III-C, Tables I, II, VIII"},{"comment":"The test-time operation of the DFF module is unspecified. The denoising loss in Eq. (4) uses paired HH and HV images as noisy/clean label pairs, but the detection evaluation on SSDD+ and HRSID is reported without stating which polarization is used for a given test image and without explaining how DFF is applied when no HH/HV pair is available. If DFF at inference takes a single input image, the paper should state this explicitly and reconcile it with the paired-image training setup. Otherwise, the reported gains may not transfer to the evaluation protocol actually used.","section":"Section II-B, Section III-A"},{"comment":"The description of the training procedure contradicts the 'multitask learning' and 'end-to-end' claims in the abstract and Section I. In the second training stage, λ1 is set to 0 and the DFF parameters are frozen, so the speckle-suppression task is not jointly optimized with detection and segmentation. The paper should either present results with truly joint training of all three tasks or revise the claims to describe DFF as a pretrained, frozen feature-enhancement module rather than a jointly learned subtask.","section":"Section III-C, Eq. (9)"},{"comment":"The headline improvements over YOLOX (+2.86 on SSDD+ and +5.05 on HRSID) are offshore-subset values, not full-dataset values. The text in Section III-G-1 lists these numbers without specifying that they are offshore, which overstates the overall margin. In addition, the five baselines are re-implemented without shipping code or configuration files, so the exact training schedules, anchor settings, and test-time details used for the comparisons are not verifiable. The authors should report full-dataset AP separately, make baseline configurations available, or use official pre-trained models and published results.","section":"Section III-G, Table VIII"}],"minor_comments":[{"comment":"There are numerous typos and inconsistent abbreviations, e.g., 'supression', 'DDF' vs. 'DFF', 'ARWS' vs. 'ARW', 'simutanously', 'demenstrates', 'organzied', and 'Guassion'. A careful proofread is needed.","section":"Throughout"},{"comment":"The denoising comparison uses 'real SAR-1' and 'real SAR-2' images whose origin, resolution, and polarization are not described, and the ENL/EPD-ROA values are reported without error bars or a description of how many regions were selected. Please specify the data source and variability.","section":"Section III-E, Tables V and VI"},{"comment":"Equations (7) and (8) are placed inside Algorithm 1 but not referenced in the surrounding text; please integrate them or cite them explicitly so the WRBF update rule is clear.","section":"Section II-D, Algorithm 1"},{"comment":"The Gaussian-mask formula uses 'ih' and 'jh' for coordinates of the horizontal Gaussian mask but the notation is not defined in the text; please define all variables in Eq. (5) explicitly.","section":"Section II-C, Eq. (5)"},{"comment":"The sentence '137327 pairs of corresponding HH and HV images are selected as the training set for DDF module' has a typo ('DDF' should be 'DFF') and, more importantly, does not state whether these pairs come from SSDD+, HRSID, or an external corpus; this should be clarified.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The undisclosed 137,327-pair HH/HV corpus is the single most serious issue: it is load-bearing for the ablation gains and yet cannot be reproduced from the two public datasets as described. If the authors cannot release the exact source and split of these pairs, the state-of-the-art claim will remain unverifiable. I would ask the editor to require this information and a controlled comparison before considering the paper for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nQuick take on 2411.13847. The paper is a competent, incremental engineering contribution. It assembles known components—rotated Gaussian-mask segmentation, a denoising feature fusion module, an aspect-ratio-weighted angle loss, and weighted rotated box fusion—into a multitask SAR ship detector, and shows consistent AP50 gains over five baselines on SSDD+ and HRSID with ablations for each module. The individual pieces are not new, but the specific package is, and the experimental discipline (ablations, inshore/offshore splits, runtime/params) is decent.\n\nThe soft spot is load-bearing. The DFF module is trained on 137,327 pairs of HH and HV images, but the paper never states where these pairs come from, how they were split, or how they were registered. Section III-A says they are 'selected as the training set' with no source; Section III-C says the DFF is 'independently trained by the dataset SSDD+,' which has 1,160 images. That is inconsistent—neither SSDD+ nor HRSID can yield that many pairs without substantial patchification that is not described. The ablation tables attribute roughly five AP50 points to DFF, so the provenance of this data directly determines whether the SOTA claim reflects the architecture or a private pretraining corpus the baselines never saw. If the pairs are external (e.g., Sentinel-1), the comparison in Table VIII is unfair; if they leak test scenes, the numbers are inflated. This needs to be documented and ideally controlled by giving the same auxiliary data to the baselines.\n\nMinor issues: the five baselines are re-implemented with no code or configs shipped, no error bars are reported, and the GitHub link is not pinned to a revision. None of these are fatal, but they raise the verification burden.\n\nOverall, the central argument is plausible and the engineering is coherent, but the empirical claim is conditional on resolving the data-provenance question. This deserves peer review rather than a desk reject; a serious referee should ask for the DFF training data documentation and a controlled baseline comparison before accepting the SOTA claim.","headline":"Plausible engineering result with a load-bearing missing detail: the 137k HH/HV pairs behind the denoising module have no documented source or split, so the SOTA claim is conditional.","tokens_in":22077,"tokens_out":3211,"would_cite":false,"duration_ms":27887,"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":"Jointly training rotated SAR ship detection with speckle suppression and Gaussian-mask segmentation yields top AP50 scores on SSDD+ and HRSID.","keywords":["SAR ship detection","multitask learning","rotated bounding box","speckle suppression","Gaussian mask segmentation","angle classification loss","weighted rotated boxes fusion","denoised feature fusion"],"falsifier":"Run MLDet's full pipeline on a test set of single-polarization SAR images with the denoised feature fusion module removed, and compare AP50 against the full model; if the gap disappears or reverses, the paired-polarization denoising pretraining, not the shared architecture, is the source of the reported gains.","tokens_in":21010,"feed_emoji":"🛰️","tokens_out":10476,"duration_ms":88509,"temperature":0.7,"pith_summary":"This paper claims that SAR ship detection improves when detection is trained jointly with two auxiliary tasks instead of alone: rotated bounding-box detection, speckle suppression, and pixel-level target segmentation. The proposed MLDet framework shares one backbone across these tasks, fuses denoised and shallow features through a dual-feature fusion attention mechanism, and supervises segmentation with a rotated Gaussian mask that puts the highest probability at the ship center. An angle-classification loss with aspect-ratio weighting handles angular periodicity, and a weighted rotated boxes fusion step combines detection and segmentation outputs. On SSDD+ and HRSID, the paper reports AP50 gains over five comparison detectors, including 2.86 and 5.05 points over YOLOX, respectively. A reader should care because the result suggests that auxiliary tasks can make SAR detectors more resistant to speckle and clutter without sacrificing real-time speed.","feed_headline":"Multitask SAR detector beats YOLOX by 5 points","feed_subtitle":"Joint speckle suppression and Gaussian-mask segmentation lift detection accuracy on SSDD+ and HRSID.","key_machinery":"The machinery has four cooperating parts. The rotated Gaussian mask is a soft segmentation label: a ship's center carries the highest probability and surrounding pixels decay under a two-dimensional Gaussian whose covariance depends on the ship's aspect ratio and rotation, so the segmentation head learns to concentrate on the ship body rather than its enclosing background. The denoised feature fusion module (DFF) is pretrained on 137,327 HH–HV polarization image pairs treated as noisy/clean label pairs, then uses a dual-feature fusion attention mechanism to concatenate channel-refined denoised features with shallow features, preserving small targets while suppressing speckle. The angle-classification loss with aspect-ratio weighting multiplies smooth L1 angle error by $|\\sin(\\alpha(\\theta-\\hat{\\theta}))|$ with $\\alpha=2$ for square-like boxes, making the angle loss periodic where angular periodicity otherwise inflates training loss. Weighted rotated boxes fusion (WRBF) clusters detection and segmentation boxes and averages their coordinates weighted by confidence, so localization, not just classification score, determines the final box. The load-bearing idea is that sharing one backbone across detection, denoising, and segmentation regularizes the detector.","core_discovery":"The paper's central claim is that the three tasks interact positively: the denoised feature fusion module suppresses speckle while preserving small-target features, the target segmentation module with the rotated Gaussian mask directs the network's attention to ship interiors, and the weighted rotated boxes fusion yields more accurate boxes than confidence-only NMS. On SSDD+ and HRSID the paper reports that MLDet outperforms Faster R-CNN, RetinaNet, Oriented R-CNN, CenterNet, and YOLOX on AP50 for both inshore and offshore subsets, with the full model reaching 95.3% AP50 on SSDD+ and 92.8% on HRSID at the 0.5 IoU threshold. The paper also reports larger AP75 gains over the baseline, which it attributes to more accurate localization from joint segmentation and fusion.","pith_inferences":["Editorial inference: the same three-task design could be tested on other noisy rotating-object domains, such as optical satellite images of ships with wakes or aerial images of parked aircraft, where the Gaussian-mask segmentation would supply similar center-biased supervision.","Editorial inference: because the DFF module is pretrained on HH–HV pairs and the paper does not specify its test-time input on single-polarization datasets, the reported gains may depend on paired-polarization pretraining; an evaluation with DFF ablated at test time would separate denoising pretraining from architectural benefit.","Editorial inference: the rotated Gaussian mask could be used as a soft target for weakly supervised instance segmentation, generating pixel-level pseudo-labels from box annotations without extra human labeling."],"forward_implications":["The framework outputs detection, denoised features, and segmentation from one shared backbone, which the paper presents as an end-to-end alternative to separate denoising and detection pipelines.","The reported AP75 gains over all comparison detectors imply that the segmentation and fusion components improve localization, not only recall, which matters for densely docked ship scenes.","The ARW loss should reduce angle-mismatch training losses for near-square ship boxes, making rotated detectors more stable on small or broadside ships.","Because the model is trained from scratch without pretrained weights, the multitask recipe offers a path for SAR domains where pretrained features transfer poorly."],"supporting_citations":[{"why":"Provides the SSDD+ benchmark with 1,160 multi-polarization images used for training and evaluation.","marker":"[49]"},{"why":"Provides the HRSID benchmark with 5,604 high-resolution images used for training and evaluation.","marker":"[52]"},{"why":"YOLOX is the strongest one-stage comparison baseline; the paper's claimed AP50 advantage over it is the headline quantitative result.","marker":"[51]"},{"why":"Supplies the rotated-bounding-box detection formulation that the ARW angle loss modifies.","marker":"[50]"},{"why":"Supplies the self-supervised denoising training idea of using corresponding same-scene SAR polarization images as noisy/clean pairs.","marker":"[56]"},{"why":"Supplies the paired-polarization denoising idea and also serves as a comparison method in the despeckling evaluation.","marker":"[57]"},{"why":"Supplies the focal classification loss used in the object detection loss and the RetinaNet baseline that is compared.","marker":"[24]"},{"why":"Provides the topological structural analysis used to convert a segmentation mask into a minimum enclosing oriented rectangle for fusion.","marker":"[53]"}],"fun_headline_variants":["MLDet: Multitask SAR detector beats YOLOX by 5 points","Gaussian-mask segmentation boosts SAR ship detection accuracy","Multitask learning with speckle suppression lifts SAR ship AP50","MLDet combines detection, denoising, segmentation for SAR ships","SAR ship detector uses rotated Gaussian masks to beat baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 137,327 HH–HV image pairs used to train the denoising module are genuine same-scene registrations, so the network learns to remove speckle rather than legitimate polarization-dependent scattering; the paper also leaves unspecified how the DFF module is fed at test time when only single-polarization images are available.","fun_headline_variants_meta":{"raw":{"variants":["MLDet: Multitask SAR detector beats YOLOX by 5 points","Gaussian-mask segmentation boosts SAR ship detection accuracy","Multitask learning with speckle suppression lifts SAR ship AP50","MLDet combines detection, denoising, segmentation for SAR ships","SAR ship detector uses rotated Gaussian masks to beat baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000856,"raw_usage":{"total_tokens":3693,"prompt_tokens":895,"completion_tokens":2798,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2723}},"tokens_in":511,"tokens_out":2798,"duration_ms":15633,"temperature":1.0,"reasoning_tokens":2723,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:26.607584+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MLDet's full pipeline on a test set of single-polarization SAR images with the denoised feature fusion module removed, and compare AP50 against the full model; if the gap disappears or reverses, the paired-polarization denoising pretraining, not the shared architecture, is the source of the reported gains.","supporting_citations":[{"cited_title":"Ship detection in SAR images based on an improved faster r-cnn,","cited_arxiv_id":null,"evidence_quote":"Provides the SSDD+ benchmark with 1,160 multi-polarization images used for training and evaluation."},{"cited_title":"Hrsid: A high-resolution SAR images dataset for ship detection and instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the HRSID benchmark with 5,604 high-resolution images used for training and evaluation."},{"cited_title":"Oriented r-cnn for object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the rotated-bounding-box detection formulation that the ARW angle loss modifies."},{"cited_title":"A CNN- based self-supervised synthetic aperture radar image denoising approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised denoising training idea of using corresponding same-scene SAR polarization images as noisy/clean pairs."},{"cited_title":"Transformer-based SAR image despeckling,","cited_arxiv_id":null,"evidence_quote":"Supplies the paired-polarization denoising idea and also serves as a comparison method in the despeckling evaluation."},{"cited_title":"Focal loss for dense object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the focal classification loss used in the object detection loss and the RetinaNet baseline that is compared."}],"review_version":1}