{"id":"e4288e4b-4702-4f22-99cd-d2e4a00f9c15","arxiv_id":"2506.11913","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"O2Former adds a multi-scale query generator and an orientation-aware module to Mask2Former and reports improved SAR ship instance segmentation on SSDD and HRSID.","lead":"Researchers adapt Mask2Former for ship segmentation in radar images by adding two modules that generate better query vectors and encode ship orientation. The model reports top scores on two public SAR ship datasets, though the gains over the strongest baseline are small and no code is released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA margin may be a transfer-initialization artifact: SSDD numbers use HRSID-pretrained weights while baseline initialization is unspecified.","rationale":"The reader's CONDITIONAL verdict is well-founded, but the strongest threat is more specific than generic 'single-run variance': Section IV-C reveals that O2Former's SSDD results were obtained after fine-tuning from weights trained on HRSID, while no comparable pretraining is described for the baselines. Since HRSID is the same sensor domain and contains 5,605 slices, initialization from HRSID can provide a substantial advantage on SSDD. The reported offshore margins over DiffSARShipInst are only 0.8-1.3 AP, comfortably within the range a transfer-initialization difference could produce. Thus the central SOTA claim is not yet established under controlled conditions. I therefore keep the reader's CONDITIONAL verdict; the condition should be release of code/configs and a matched transfer protocol, or the comparison remains confounded.","tokens_in":17549,"tokens_out":10834,"duration_ms":236688,"concrete_test":"Obtain the authors' training code and hyperparameter configurations and rerun the Table II comparison under a single matched protocol: every baseline, including DiffSARShipInst and Mask2Former, must use the same initialization schedule (ImageNet -> HRSID -> SSDD), same batch size, same epochs, and same learning rate schedule as O2Former. If either baseline's offshore APm rises by at least 1 point (i.e., to at least 71.6) or its inshore APm rises by at least 1 point relative to its reported Table II value, the marginal offshore lead is within transfer-initialization effects and the SOTA claim fails. If code is unavailable, request an explicit statement, for each baseline, of whether HRSID pretraining was used before SSDD fine-tuning; without an affirmative answer for all baselines, the comparison remains uncontrolled.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that O2Former's gains come from OQG+OAEM, not from training setup. Section IV-C specifies only: 'The proposed method was tested on SSDD as well as HRSID, initialized on HRSID using ResNet-50 pretrained weights on the ImageNet dataset and fine-tuned on SSDD using weights on the HRSID.' Thus the SSDD results in Table II, including the 71.9 vs 70.6 offshore and 63.2 vs 56.2 inshore margins over DiffSARShipInst, were produced by a model that was first trained on HRSID and then fine-tuned on SSDD. The paper never states that DiffSARShipInst, Mask2Former, or any other baseline received the same HRSID-to-SSDD transfer initialization. In small-data SAR tasks, such transfer initialization can be worth several AP points, which is comparable to or larger than the 0.8-1.3 point offshore margin over the strongest baseline. If the baselines were trained ImageNet-only while O2Former used HRSID pretraining, the comparison is not controlled and the claimed SOTA superiority is not established. The internal inconsistency of OQG's equations (Eq. (1) produces a per-scale vector after global pooling, yet the text claims a BxNxC tensor with N=HxW) further prevents independent reproduction of the method as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes O2Former, an extension of Mask2Former for SAR ship instance segmentation, with two new modules: an Optimized Query Generator (OQG) that initializes transformer queries from multi-scale backbone features, and an Orientation-Aware Embedding Module (OAEM) that injects direction-sensitive and polar-coordinate geometric priors. The method is evaluated on PSeg-SSDD and HRSID against several state-of-the-art baselines, reporting improved AP in most settings, with the largest gains on inshore scenarios of PSeg-SSDD (63.2% AP vs. 56.2% for the strongest baseline). The paper also includes an ablation study quantifying the contribution of each module.","tokens_in":17830,"tokens_out":2816,"duration_ms":36981,"significance":"If the reported results hold, the paper provides a practically useful improvement for a niche but application-relevant task: instance segmentation of ships in SAR imagery. The design is well motivated by SAR-specific challenges (scale variation, orientation diversity, boundary fuzziness), and the combination of query initialization with orientation-aware features is a reasonable extension of Mask2Former. The experimental scope is appropriate for the venue, with two public datasets and multiple baselines. However, the paper currently lacks the experimental rigor needed to substantiate the central claim of state-of-the-art performance: no repeated runs or error bars, unspecified baseline initialization, and incomplete method details that prevent independent reproduction. The theoretical novelty is limited, but the empirical contribution could be acceptable if the issues are resolved.","major_comments":[{"comment":"The SSDD results are obtained with a model that was pre-trained on HRSID and then fine-tuned on SSDD ('initialized on HRSID ... and fine-tuned on SSDD'), but the paper does not state whether any of the comparison baselines (Mask2Former, DiffSARShipInst, etc.) received the same transfer initialization. In small-data SAR settings, HRSID pre-training can yield gains of several AP points, which is larger than the reported 0.8–1.3 point offshore margins over DiffSARShipInst. To support the claim that OQG and OAEM are responsible for the improvement, the authors must either specify that all baselines used identical pre-training/initialization procedures or provide an additional controlled comparison, e.g., training every method from ImageNet-only weights and reporting both settings.","section":"§IV-C, Table II"},{"comment":"The OQG equations are internally inconsistent and prevent reproduction. Global average pooling in Eq. (1) reduces each feature map to a 1×1×C vector, so after flattening and stacking, F in Eq. (2) should be of shape B×4×C. The text, however, states F∈R^{B×N×C} with N=H×W, which cannot be obtained from the stated operations. Moreover, Eq. (4) sums over i=0 to 4, but there are only four scales (C2–C5), so the index range is incorrect (should be i=1..4 or i=0..3). These ambiguities must be corrected or the module described precisely.","section":"§III-B, Eqs. (1)–(4)"},{"comment":"In the OAEM description, Eq. (7) declares X_rot^i ∈ R^{(C/N)×H×W} after rotating the input X∈R^{C×H×W}, but rotating a feature map should preserve the channel dimension. It is unclear whether the channels are split into N groups before rotation, or whether the rotation is applied to the full tensor and the output dimension is a typo. This matters because Eq. (9) concatenates N branches to obtain C×H×W, which implies each branch has C/N channels. Please clarify the channel handling and specify the hyperparameter N used in the experiments, as it does not appear in Section IV-C.","section":"§III-C, Eq. (7)"},{"comment":"The paper reports single-run results for all methods and ablations, with no error bars, standard deviations, or repeated-seed experiments. Given that several reported differences are small (e.g., 0.8 AP on HRSID offshore in Table III), and the ablation gains vary substantially (e.g., OQG+OAEM inshore AP75 is 0.812 versus 0.631 for OAEM alone in Table IV), it is impossible to assess whether these differences are statistically meaningful. Please add at least three runs with different seeds and report mean±std, or justify why single runs are adequate for the claimed conclusions.","section":"§IV-C and Table IV"}],"minor_comments":[{"comment":"The implementation description states 'PyTorch 3.8,' which is not a valid PyTorch version; presumably this should be Python 3.8 with a specific PyTorch release.","section":"§IV-C"},{"comment":"There is a typo in Eq. (15): 'F_poalr' should be 'F_polar'.","section":"§III-C, Eq. (15)"},{"comment":"'presion' should be 'precision.'","section":"§IV-B, Eqs. (16)–(17)"},{"comment":"The AP integral in Eq. (18) is written without integration limits; consider adding 0 to 1 explicitly for clarity.","section":"§IV-B, Eq. (18)"},{"comment":"The introduction refers to 'traditional method [15], [16], [17], [18]' with inconsistent singular/plural agreement; minor editorial revision is needed.","section":"§I"}],"recommendation":"major_revision","confidential_remarks":"The paper heavily cites the authors' own prior work (e.g., [3], [4], [7], [18], [57]) and includes several references to recent arXiv preprint papers on garment synthesis and pose generation that appear unrelated to SAR segmentation. While this does not affect the technical correctness, the citation list seems padded and may not meet the journal's expectation of a focused related-work section. The editor may wish to request the authors to prune irrelevant references."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: O2Former is a plausible engineering modification of Mask2Former for SAR ship instance segmentation, with two sensible modules and a decent ablation, but the headline SOTA claim is not established. The SSDD comparison appears to use transfer initialization from HRSID for their model and unspecified (likely ImageNet-only) initialization for the baselines. That alone can explain much of the reported margin.\n\nWhat's actually new: OQG initializes queries from pooled multi-scale features with a learnable prototype and similarity weighting; OAEM applies rotated convolutions plus polar-coordinate geometric priors. Both are reasonable adaptations of existing ideas (DAB-DETR-style query refinement, orientation-aware convolutions in other domains), and the ablation shows each module adds a few AP points on PSeg-SSDD. That part is fine as an engineering story.\n\nThe soft spots:\n1. The transfer-initialization problem. Section IV-C says their SSDD model was initialized with HRSID-trained weights. No baseline is reported to receive the same treatment. On small SAR datasets, HRSID pretraining can easily be worth more than the 0.8–1.3 AP offshore margins over DiffSARShipInst. Without matched initialization, the state-of-the-art claim is unsupported.\n2. Single runs, no error bars. Several claimed wins are within typical run-to-run noise.\n3. Missing hyperparameters and internal inconsistencies. Query count and OAEM angle count N are not reported. Eq. (1) flattens pooled features to vectors but the text says N=H*W; Eq. (4) sums i=0..4 for four scales; 'PyTorch 3.8' is not a thing. These make independent reproduction harder than it should be.\n4. Some references (IMAGGarment-1, etc.) are tangential.\n\nNone of this makes the modules worthless; the ablation is internally consistent and the inshore gains on SSDD are larger. But the central empirical claim is not proven as written. Who should read it: groups working on SAR instance segmentation who want to try query-based methods and see the baseline-matching pitfalls. It deserves peer review, but a referee should demand matched initialization, error bars, and full hyperparameters.","headline":"O2Former's modules are reasonable, but the reported SOTA margin is not credible because the SSDD runs use HRSID transfer initialization while baselines likely do not.","tokens_in":18336,"tokens_out":2528,"would_cite":false,"duration_ms":30161,"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":"O2Former claims that giving Mask2Former a multi-scale query generator and an orientation-aware embedding module produces state-of-the-art SAR ship instance segmentation, reaching 71.9% AP offshore and 63.2% AP inshore on PSeg-SSDD.","keywords":["SAR ship instance segmentation","Mask2Former","query-based learning","orientation-aware embedding","polar coordinate encoding","multi-scale feature fusion","PSeg-SSDD","HRSID"],"falsifier":"Re-train O2Former and every baseline on PSeg-SSDD and HRSID with identical data splits, augmentation, learning-rate schedule, and inference settings, running at least five seeds per method, and check whether the 1.3-point offshore and 7.0-point inshore AP margins over DiffSARShipInst persist; a separate check is to sweep the OAEM angle count N (for example N=4, 8, 12) and measure whether performance is flat, which would indicate the orientation module is not the source of the gain.","tokens_in":17335,"feed_emoji":"🛰️","tokens_out":6008,"duration_ms":66280,"temperature":0.7,"pith_summary":"O2Former claims that the standard query-based instance segmentation architecture Mask2Former can be made substantially more accurate on SAR ship imagery by intervening at two points: how the decoder queries are initialized and how directional information is encoded into features. The paper introduces the Optimized Query Generator (OQG), which builds queries from globally pooled multi-scale backbone features plus a learned ship prototype, and the Orientation-Aware Embedding Module (OAEM), which extracts features at several rotated angles and fuses them with polar-coordinate geometric priors. On the PSeg-SSDD benchmark the combined model reaches 71.9% AP in offshore scenes and 63.2% AP inshore, versus 70.6% and 56.2% for the previous best model DiffSARShipInst and 67.2% and 52.5% for Mask2Former. On HRSID the reported gains are smaller but consistent, with O2Former leading on high-IoU (AP75) inshore segmentation, which the paper reads as evidence of sharper mask boundaries. If these numbers hold under matched training conditions, the work shows that query initialization and orientation priors are decisive for dense, directionally diverse SAR scenes.","feed_headline":"Direction-aware queries lift SAR ship segmentation AP to 71.9%","feed_subtitle":"Two lightweight modules add multi-scale query cues and polar orientation priors to Mask2Former for sharper ship masks.","key_machinery":"The load-bearing machinery is the pair of modules inserted between the backbone and the transformer decoder. The Optimized Query Generator (OQG) performs a softmax-weighted fusion of global-average-pooled multi-scale features with learnable scale embeddings, then updates a randomly initialized ship prototype through a cosine-similarity signal, so each query carries multi-scale semantic information before it enters the decoder. The Orientation-Aware Embedding Module (OAEM) samples the feature map at N angles θ_i = iπ/N via affine-grid rotation, applies an independent convolution to each rotated copy, encodes spatial position as normalized polar coordinates (r, θ), and merges these two streams with an input-dependent weight. The polar-coordinate radius is meant to separate central and edge regions, while the angle channel lets convolution kernels learn rotation patterns explicitly, and the dynamic fusion weight lets the network choose how much to trust semantic versus geometric structure per location. Together they are intended to fix the feature-alignment and directional-diversity failures the paper attributes to zero-initialized queries and orientation-blind attention in Mask2Former.","core_discovery":"The paper's central claim is that O2Former, by adding OQG and OAEM to Mask2Former, outperforms state-of-the-art instance segmentation methods on the PSeg-SSDD and HRSID SAR ship datasets. The OQG replaces zero-initialized queries with a context-aware initialization: the four backbone feature maps C2 through C5 are globally pooled, flattened, stacked with per-scale embeddings, and passed through a linear layer with softmax attention; the resulting fused feature is matched against a learnable ship prototype by cosine similarity, and the updated prototype becomes the query set. The OAEM rotates the input feature map by N preset angles using affine-grid sampling, runs a dedicated convolution branch per angle, converts Cartesian coordinates into normalized polar coordinates (radius and angle), and dynamically fuses the orientation-sensitive and polar features with softmax-derived weights. The paper reports that the two modules together lift Mask2Former from 67.2% to 71.9% AP in PSeg-SSDD offshore scenes and from 52.5% to 63.2% AP inshore, with the biggest absolute gains at the high IoU threshold AP75.","pith_inferences":["A natural stress test the paper leaves implicit is to rotate ships in the test set by arbitrary angles and measure AP; OAEM's polar encoding is angle-sensitive, but the fixed set of N angles may not deliver full rotation invariance, and the paper does not report an angle sweep.","OQG's design is not SAR-specific and could be plugged into other query-based detectors and segmenters where query initialization is a known convergence bottleneck, possibly transferring the multi-scale initialization benefit to optical imagery as well.","The reported AP_L numbers on PSeg-SSDD rest on very few large targets (19 inshore, 52 offshore), so the large-object gains are statistically fragile and should be read as indicative rather than conclusive.","The single-run evaluation protocol leaves open the possibility that part of the reported lead comes from training stochasticity; reporting seeded variance would let the field separate architectural gains from run-to-run noise."],"forward_implications":["On PSeg-SSDD, adding both modules raises AP over Mask2Former by 4.7 points offshore and 10.7 points inshore, with the largest gains at the high IoU threshold AP75 (up 9.7 points offshore and 21.3 points inshore).","OQG alone improves multi-scale and small-object perception, while OAEM alone improves boundary detail, so the two modules are complementary rather than redundant.","On HRSID, which has more small ships and complex background clutter, O2Former still leads on most metrics, including AP75 inshore (45.9%), indicating the orientation priors help in dense and cluttered scenes.","The model's reduced false alarms, fewer missed small targets, and clearer separation of adjacent ships in the reported visualizations suggest it can be applied to maritime monitoring tasks that demand reliable instance-level masks.","Because both modules attach to the Mask2Former backbone-decoder interface, the reported improvements are obtained without changing the training data or the base query-based segmentation paradigm."],"supporting_citations":[{"why":"Mask2Former is the base architecture that O2Former extends; all reported gains are measured as improvements over it.","marker":"[14]"},{"why":"PSeg-SSDD (SSDD) is the primary benchmark dataset, and its official train/test split defines the offshore and inshore evaluation scenes.","marker":"[19]"},{"why":"HRSID is the second benchmark dataset, providing the high-resolution, small-ship-heavy test case used to validate generalization.","marker":"[20]"},{"why":"DiffSARShipInst is the strongest prior method that O2Former must beat, supplying the key SOTA comparison numbers in both tables.","marker":"[59]"},{"why":"Mask R-CNN is the classic region-proposal baseline whose performance frames the difficulty of SAR ship instance segmentation.","marker":"[23]"},{"why":"HQ-ISNet represents the remote-sensing instance segmentation baseline that the paper argues ignores SAR-specific directional structure.","marker":"[29]"}],"fun_headline_variants":["Direction-aware queries sharpen SAR ship mask predictions","Multi-scale query and orientation modules improve SAR segmentation","O2Former outperforms SOTA on SAR ship instance segmentation","Orientation-aware queries lift SAR ship AP to 71.9%","Polar encoding and direction-aware conv refine SAR ship masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that all comparison methods were trained and evaluated under genuinely matched conditions (same data splits, preprocessing, and hyperparameter care) and that a single run of each method represents its true performance; the paper gives only coarse training details and does not state the number of queries, the angle count N in OAEM, or multiple seeds.","fun_headline_variants_meta":{"raw":{"variants":["Direction-aware queries sharpen SAR ship mask predictions","Multi-scale query and orientation modules improve SAR segmentation","O2Former outperforms SOTA on SAR ship instance segmentation","Orientation-aware queries lift SAR ship AP to 71.9%","Polar encoding and direction-aware conv refine SAR ship masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000668,"raw_usage":{"total_tokens":3065,"prompt_tokens":981,"completion_tokens":2084,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2004}},"tokens_in":597,"tokens_out":2084,"duration_ms":19434,"temperature":1.0,"reasoning_tokens":2004,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T01:00:15.954811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-train O2Former and every baseline on PSeg-SSDD and HRSID with identical data splits, augmentation, learning-rate schedule, and inference settings, running at least five seeds per method, and check whether the 1.3-point offshore and 7.0-point inshore AP margins over DiffSARShipInst persist; a separate check is to sweep the OAEM angle count N (for example N=4, 8, 12) and measure whether performance is flat, which would indicate the orientation module is not the source of the gain.","supporting_citations":[{"cited_title":"Sar ship detection dataset (ssdd): Official release and comprehensive data analysis,","cited_arxiv_id":null,"evidence_quote":"PSeg-SSDD (SSDD) is the primary benchmark dataset, and its official train/test split defines the offshore and inshore evaluation scenes."},{"cited_title":"Ship instance segmentation based on rotated bounding boxes for sar images,","cited_arxiv_id":null,"evidence_quote":"DiffSARShipInst is the strongest prior method that O2Former must beat, supplying the key SOTA comparison numbers in both tables."},{"cited_title":"Yolact: Real-time instance segmentation,","cited_arxiv_id":null,"evidence_quote":"HQ-ISNet represents the remote-sensing instance segmentation baseline that the paper argues ignores SAR-specific directional structure."}],"review_version":1}