{"id":"7b310908-5710-444d-a7fc-3db9de23ae1a","arxiv_id":"2602.10710","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"FGAA-FPN combines box-supervised foreground masks with orientation-biased attention in a feature pyramid and reports +3.9 mAP over FPN on DOTA v1.5, while its claimed SOTA on DOTA v1.0 is contradicted by its own table.","lead":"This aerial-imaging detection paper adds two modules to a standard feature pyramid: a foreground mask that suppresses clutter in fine-scale features, and an attention mechanism biased by relative object orientation. It reports higher accuracy on DOTA v1.5 but its stated state-of-the-art result on DOTA v1.0 is contradicted by its own comparison table.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline SOTA claim is contradicted by the paper's own Table 1: AO2-DETR (77.7) exceeds FGAA-FPN (75.5) on the same ResNet-50 DOTA v1.0 setting, so 'best overall performance' cannot stand.","rationale":"In good faith, the paper proposes two concrete modules (FGFM and AAMHA) and provides ablations that show internal improvements over an FPN baseline. That is a legitimate contribution, and the reader correctly noted that the internally supported claim is narrower than the advertised SOTA. The formal weakest_assumption in the reader's verdict—that a learned 2D orientation vector per head adequately encodes relative orientation—is a plausible correctness risk in the mechanism's design, but it is not the most decisive problem. The most load-bearing issue is simpler and more direct: the paper's own Table 1 contains a method (AO2-DETR) with a higher mAP50 (77.7) under the same backbone and benchmark, making the 'state-of-the-art' and 'best overall performance' statements false on the paper's own evidence. This is an internal inconsistency, so no external benchmark or reimplementation is needed to establish that the headline overclaims. The reader's rationale did mention this Table 1 contradiction, but did not make it the formal weakest assumption. Thus my concern partially overlaps with the reader's written rationale but diverges from the formal weakest_assumption field. Since the reader's verdict is already REJECT and my concern reinforces that rejection without changing its basis, the appropriate verdict adjustment is UNCHANGED. A corrected paper could perhaps earn CONDITIONAL or ACCEPT if it narrows the claim to 'improves over FPN under Oriented R-CNN on DOTA v1.5' and releases reproducible code and error bars, but as submitted the central claim cannot be accepted.","tokens_in":17884,"tokens_out":2929,"duration_ms":28307,"concrete_test":"Check Table 1 directly: confirm AO2-DETR's 77.7 and FGAA-FPN's 75.5 are both mAP50 on DOTA v1.0 with ResNet-50 and the same evaluation protocol. If so, recompute the ranking over all rows: FGAA-FPN is not first. Additionally, run FGAA-FPN from the released code (if provided) on DOTA v1.0 with the stated training setup and evaluate the official test server or validation split; if the resulting mAP does not exceed 77.7, the SOTA claim is empirically refuted. This single comparison settles the central advertised claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'FGAA-FPN achieves state-of-the-art results, reaching 75.5% and 68.3% mAP' (Abstract) and 'the best overall performance on the DOTA v1.0 dataset' (§4.2.1)—is contradicted by the paper's own Table 1. That table lists AO2-DETR at 77.7 mAP50 with a ResNet-50 backbone, 2.2 points above the reported 75.5 for FGAA-FPN under the same benchmark and same backbone. The table header is explicitly mAP50, so this is an apples-to-apples comparison, not a difference in metric. This is an internal inconsistency, not a dispute about external consensus. The internally supported version of the claim is much narrower: replacing FPN with FGAA-FPN under Oriented R-CNN improves DOTA v1.5 mAP50 from 64.4 to 68.3 (Table 3). But that result comes from single runs with no error bars, no released code or weights, and the bias scales γ and β were selected on the same benchmark (§4.3.3), leaving open the possibility of tuning on the test set. The orientation-bias encoding in Eq. 15–16, while plausible, is secondary: even if AAMHA is a valid mechanism, the paper's advertised SOTA status is already falsified by its own table. The load-bearing condition for the headline claim is that FGAA-FPN beats all compared detectors on DOTA v1.0; Table 1 fails that condition. Therefore the abstract and §4.2.1 overclaim, and the verdict should remain REJECT.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FGAA-FPN, a feature-pyramid neck for oriented object detection that combines Foreground-Guided Feature Modulation (FGFM) on lower pyramid levels (P3–P5) with Angle-Aware Multi-Head Attention (AAMHA) on higher levels (P5–P7), integrated into the Oriented R-CNN detector. The authors report 75.5 mAP on DOTA v1.0 and 68.3 mAP on DOTA v1.5, and claim state-of-the-art results. The core evidence is the controlled neck-replacement comparison in Table 2, the component ablations in Tables 3–5, and a generalization experiment across detectors in Figure 4.","tokens_in":18289,"tokens_out":6374,"duration_ms":60075,"significance":"If the controlled neck-replacement results were statistically robust, the two modules would constitute a modest but potentially useful contribution: FGFM adds weakly supervised foreground saliency to low-level features, and AAMHA injects an explicit directional bias into self-attention. The paper's strength is the controlled neck-only comparison on DOTA v1.5 and the placement/cost analysis in Tables 3–5. However, the headline SOTA claim is contradicted by the paper's own Table 1, the controlled experiments lack error bars and multiple runs, and the bias hyperparameters are selected and evaluated on the same benchmark. As submitted, the evidence does not support the advertised claims; the defensible contribution is narrower and needs stronger validation.","major_comments":[{"comment":"The abstract's claim of 'state-of-the-art results, reaching 75.5%' and §4.2.1's claim of 'the best overall performance on the DOTA v1.0 dataset' are directly contradicted by the paper's own Table 1. AO2-DETR, with the same ResNet-50 backbone, reports 77.7 mAP50, 2.2 points above FGAA-FPN's 75.5. Since the table header is explicitly mAP50, this is an apples-to-apples comparison. The central empirical claim is false as stated; at minimum it must be narrowed to 'improves over the FPN baseline under Oriented R-CNN' or 'competitive among two-stage FPN-based detectors.'","section":"Abstract; §4.2.1, Table 1"},{"comment":"The orientation-bias and mask-bias scales γ=0.7 and β=0.6 are selected through ablations on DOTA v1.5 (§4.3.3, Table 5) and then used in the final reported DOTA v1.5 result (68.3 mAP50, Table 2/Table 3). No separate validation split, multiple seeds, or error bars are reported. If model selection and final evaluation use the same split/benchmark, the gains may be selection artifacts. This is load-bearing because the claimed DOTA v1.5 improvement over BVAMFPN is only 0.9 points, smaller than typical run-to-run variation in this setting. Please report mean±std over at least three runs, specify the exact train/val/test protocol, and fix γ, β a priori or select them on a disjoint validation set.","section":"§4.3.3, Table 5; §4.3.1; Table 2"},{"comment":"The key premise that normalized coordinate differences between arbitrary feature-grid tokens encode 'relative orientation relationships' is not established. Because u_pq is the pixel-space direction between tokens, w_h^T u_pq is effectively a linear positional-encoding term; it does not depend on the oriented object's angle. To support the 'angle-aware' attribution of the gains in Tables 3 and 5, the paper should include controls such as (a) fixed random w_h, (b) a per-head learnable scalar with no directional structure, or (c) a standard relative-position bias. Without such controls, the mechanism claim is not strongly supported, although this concern is secondary to the SOTA contradiction.","section":"§3.3.2, Eqs. (15)–(16)"},{"comment":"The generalization experiment is difficult to evaluate as presented. The text states improvements of 1.5%, 0.3%, 1.2%, 0.7%, and 2.8% mAP for Rotated Faster R-CNN, RoI Transformer, Gliding Vertex, Rotated RetinaNet, and Oriented R-CNN, respectively, but the bar values in Fig. 4 imply different gaps (e.g., Rotated Faster R-CNN 64.58→65.58 is 1.0, not 1.5). No details are given about training schedules, epochs, or whether the same hyperparameters were used across detectors. Please correct the numbers and provide a table with exact values and standard deviations.","section":"§4.3.4, Fig. 4"}],"minor_comments":[{"comment":"The mAP notation is inconsistent. Section 4.1.3 defines mAP as the mean over multiple IoU thresholds, but all tables report 'mAP50'. The abstract also reports 75.5% and 68.3% as 'mAP'. Please use consistent notation and state explicitly that the headline numbers are mAP50 unless a multi-IoU mAP is actually computed.","section":"§4.1.3; Tables 1–5; Abstract"},{"comment":"The description of the mask-bias placement is inconsistent. Section 3.3.3 broadcasts B_fg to all heads, while §4.3.1 states 'the mask bias is additionally enabled only at P5' in the full FGAA-FPN setting. Please clarify whether β applies at P5 only or at all AAMHA levels, and update the architecture description and Figure 3 accordingly.","section":"§3.3.3 vs §4.3.1"},{"comment":"Several implementation details needed for reproducibility are missing: the values of the foreground-loss balancing factor λ_fg and Dice weight λ_d, the initialization/ranges for the learnable scalars k, b, λ in Eq. (6), and the scaling factor α in Eq. (11). Please report these hyperparameters explicitly.","section":"§3.4; Eq. (6); Eq. (11)"},{"comment":"In DOTA v1.5, the 'container crane' (CC) category has very low AP (12.3–17.3) across all methods, including FGAA-FPN. A brief discussion of whether the proposed foreground or angle-aware modules particularly help or harm rare categories would improve the analysis.","section":"§4.2.2, Table 2"},{"comment":"The figure is visually crowded and the bar labels are difficult to read. Also, 'RoiTransformer' and 'RoITransformer' are used inconsistently across text, figures, and references. Please standardize.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The paper has a usable empirical setup (neck-only comparison on DOTA v1.5) and a plausible architectural idea, but the headline SOTA claim is internally contradicted by Table 1, and the controlled gains are not statistically validated. The issues are not merely presentational: the central advertised result is false as stated, and the mechanism attribution requires additional controls. I would recommend rejection, though a substantially revised and re-scoped submission—with corrected claims, validation protocols, error bars, and code release—could be considered on its own merits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you want a clean example of a paper whose advertised headline is falsified by its own results table. The method itself is not absurd: FGFM estimates a coarse foreground probability map from low-level pyramid features and uses it to reweight features; AAMHA adds a per-head orientation bias plus a foreground bias to attention at high levels. Placing FGFM on P3-P5 and AAMHA on P5-P7 is a sensible hierarchical decomposition, and the ablations are internally consistent: each module helps individually, and the combination gives +3.9 mAP50 over FPN on DOTA v1.5 under Oriented R-CNN. That is a plausible incremental contribution.\n\nBut the paper does not present it that way. The abstract claims state-of-the-art 75.5% on DOTA v1.0, and §4.2.1 repeats \"best overall performance.\" Table 1 in the same paper lists AO2-DETR at 77.7% with the same ResNet-50 backbone, 2.2 points higher. That is an internal contradiction, not a matter of benchmark methodology. The DOTA v1.5 story is weaker: the claimed advantage over BVAMFPN is 0.9 mAP, with no error bars, no multiple seeds, and no code or weights released. The strongest internally supported result is the +3.9-point ablation over the FPN baseline, but γ and β were selected by ablating on DOTA v1.5 and then reported on that same benchmark, which is effectively tuning on the test set.\n\nThe orientation mechanism itself is plausible but the weakest conceptual link: one learned 2D vector per attention head over normalized coordinate differences is not clearly tied to object orientation, and the gains could come from generic spatial bias. The loss and detach tricks are clean, and the placement ablation is actually thoughtful. But none of that survives verification, because the numbers are not independently reproducible.\n\nNet: this is not a paper I would send to peer review as-is. It needs the SOTA language removed, the test-set selection addressed with a held-out validation split or fixed hyperparameters, and ideally code or at least multiple seeds. If those were fixed, it would be a reasonable incremental neck paper. I would not cite it in its current form, and I would not bring it to reading group except as a cautionary example.","headline":"A plausible incremental neck design whose headline SOTA claim is contradicted by its own Table 1 and whose remaining evidence is neither multi-seed nor released.","tokens_in":18807,"tokens_out":3773,"would_cite":false,"duration_ms":37895,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FGAA-FPN claims to reach 75.5% mAP on DOTA v1.0 and 68.3% on DOTA v1.5 by adding foreground-guided modulation and angle-aware attention to a feature pyramid for oriented object detection.","keywords":["oriented object detection","feature pyramid network","foreground-guided modulation","angle-aware attention","DOTA dataset","remote sensing imagery","multi-scale features","weak supervision"],"falsifier":"An ablation that replaces the orientation vector w_h with fixed random directions that are independent of coordinates, while keeping the same parameter count and tuning γ on the validation set; if DOTA v1.5 mAP50 stays near 68.0–68.3, the directional encoding is not the operative factor. Alternatively, evaluate FGAA-FPN on images rotated by 90° at test time: a genuine orientation-aware mechanism should lose accuracy relative to the FPN baseline when the learned orientation priors no longer match the data.","tokens_in":17711,"feed_emoji":"🛰️","tokens_out":5138,"duration_ms":47472,"temperature":0.7,"pith_summary":"The paper proposes a redesign of the feature pyramid network for oriented object detection in aerial imagery, arguing that pyramid levels have complementary functional roles: fine levels need background suppression, coarse levels need orientation-preserving interaction. It introduces two modules — a Foreground-Guided Feature Modulation (FGFM) applied at high-resolution levels, and an Angle-Aware Multi-Head Attention (AAMHA) applied at low-resolution levels — and combines them in a single FGAA-FPN. On DOTA benchmarks with a ResNet-50 backbone and Oriented R-CNN detector, this raises mAP50 on DOTA v1.5 from 64.4 (baseline FPN) to 68.3, and reaches 75.5 on DOTA v1.0. The ablations show each module contributes independently, and generalization experiments show consistent gains across five different detection heads, suggesting the design principle is not tied to one architecture.","feed_headline":"Angle-aware pyramid lifts DOTA v1.5 mAP to 68.3%","feed_subtitle":"Foreground suppression at fine scales plus orientation attention at coarse scales: consistent gains across five detectors.","key_machinery":"Two complementary modules. FGFM predicts a per-pixel foreground probability map directly from the feature map, calibrates it via learnable sharpness/bias/strength parameters, generates channel-wise modulation weights by concatenating the calibrated map with the input, and applies residual scaling Fi ⊙ (1 + α M′). AAMHA applies multi-head self-attention where each head h has a learnable 2D orientation vector w_h; for every token pair (p,q), it computes the normalized relative direction u_pq = (c_p − c_q)/(‖c_p − c_q‖ + ε) and adds an orientation bias B_ori_h(p,q) = w_hᵀ u_pq to the attention logits with scale γ. A foreground bias B_fg(p,q) = m_p(2m_q − 1) using the FGFM map is added with scal","core_discovery":"The central discovery is that explicitly allocating different mechanisms to different pyramid levels—foreground suppression at fine, high-resolution levels and angle-aware global interaction at coarse, semantic levels—improves oriented object detection compared with the uniform fusion of a standard FPN. Internally supported, the paper shows that under the same Oriented R-CNN (ResNet-50) setup, replacing FPN with FGAA-FPN raises DOTA v1.5 mAP50 from 64.4% to 68.3% and mAP75 from 41.7% to 44.1%, with FGFM alone contributing +2.7 mAP50 and AAMHA alone +1.9 mAP50. The same neck also produces consistent gains when plugged into Rotated Faster R-CNN, RoI Transformer, Gliding Vertex, Rotated RetinaN","pith_inferences":["The orientation bias uses raw pixel-grid coordinates, so it implicitly encodes image-frame direction rather than object-relative direction; if the dataset has a canonical orientation (e.g., ships aligned to coastlines), part of the gain could come from learning dataset-specific directional priors rather than true orientational generalization. A test: apply random 90°/180° rotations to test images ","The hierarchy-aware principle could be extended to a learned router that assigns per-level whether to apply foreground suppression, orientation attention, both, or neither; the fixed placement (FGFM on P3–P5, AAMHA on P5–P7) may not be optimal for datasets with different scale distributions.","Because foreground supervision is generated by projecting oriented box interiors onto feature maps, the masks are coarse; combining FGFM with pixel-wise segmentation labels where available might sharpen the foreground prior and further suppress false positives in heavily cluttered scenes."],"forward_implications":["Replacing a standard FPN with FGAA-FPN under Oriented R-CNN raises DOTA v1.5 mAP50 from 64.4% to 68.3% and mAP75 from 41.7% to 44.1%, with the two modules providing complementary gains.","The design is not head-specific: using FGAA-FPN as the neck in Rotated Faster R-CNN, RoI Transformer, Gliding Vertex, and Rotated RetinaNet improves mAP by 1.5, 0.3, 1.2, and 0.7 points respectively (Oriented R-CNN: +2.8).","High-level placement of AAMHA (P5–P7) is cost-effective: it achieves 66.3 mAP with 3.61M parameters and 60.88 GFLOPs, versus 66.5 mAP with 4.37M parameters and 80.94 GFLOPs when applied to all levels.","Within AAMHA, the orientation bias is the primary driver: enabling it alone adds 1.8 mAP50 on DOTA v1.5, while the foreground mask bias alone adds 1.0; combining them reaches 68.3.","On DOTA v1.0, FGAA-FPN reaches 75.5 mAP50, outperforming compared necks with the same ResNet-50 backbone, with notable gains on Small vehicle (77.2) and Harbor (65.1)."],"fun_headline_variants":["Foreground suppression + angle attention: FPN rework lifts DOTA v1.5","FGAA-FPN: fine-scale foreground, coarse-scale angle attention improves DOTA","Pyramid split: foreground at fine, angle at coarse lifts DOTA v1.5 to 68.3%","Fine foreground, coarse angle: FGAA-FPN neck gains 3.9 mAP on DOTA v1.5","FGAA-FPN: hierarchical FPN with angle-aware attention hits 68.3% on DOTA"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the normalized coordinate difference between two feature-grid locations is a valid encoding of their object-orientation relationship, and that a single learned 2D vector per attention head is enough to exploit it; if raw pixel direction does not correspond to the object-relative orientation needed for regression, the angle-aware gains in the ablations could be artifacts of tuning γ and β on the same benchmark.","fun_headline_variants_meta":{"raw":{"variants":["Foreground suppression + angle attention: FPN rework lifts DOTA v1.5","FGAA-FPN: fine-scale foreground, coarse-scale angle attention improves DOTA","Pyramid split: foreground at fine, angle at coarse lifts DOTA v1.5 to 68.3%","Fine foreground, coarse angle: FGAA-FPN neck gains 3.9 mAP on DOTA v1.5","FGAA-FPN: hierarchical FPN with angle-aware attention hits 68.3% on DOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001156,"raw_usage":{"total_tokens":4656,"prompt_tokens":801,"completion_tokens":3855,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":3735}},"tokens_in":545,"tokens_out":3855,"duration_ms":26631,"temperature":1.0,"reasoning_tokens":3735,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:59:52.650306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation that replaces the orientation vector w_h with fixed random directions that are independent of coordinates, while keeping the same parameter count and tuning γ on the validation set; if DOTA v1.5 mAP50 stays near 68.0–68.3, the directional encoding is not the operative factor. Alternatively, evaluate FGAA-FPN on images rotated by 90° at test time: a genuine orientation-aware mechanism should lose accuracy relative to the FPN baseline when the learned orientation priors no longer match the data.","supporting_citations":[],"review_version":1}