{"id":"763f6ea3-e6e3-4dab-ae59-f46ec6015154","arxiv_id":"2608.09270","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"GRASP combines region-focused alignment with semantic prototype based hard-negative synthesis to improve drone-view fine-grained image-text retrieval.","lead":"GRASP is a training-time framework that improves fine-grained image-text retrieval for drone-view aerial images by aligning object regions and synthesizing hard negatives from a semantic prototype codebook. It reports consistent gains over prior methods on the GeoText-1652 benchmark and competitive zero-shot transfer to the ERA dataset, which matters for applications like natural-language-guided drones and aerial search.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 zero-shot ERA comparison is confounded: fine-tuned ERA baselines (VCSR etc.) are not controlled for backbone or pretraining, so the claim that GRASP outperforms the best fine-tuned ERA method may reflect the stronger XVLM backbone, not the proposed losses.","rationale":"I focused on the zero-shot ERA claim rather than the VSP in-distribution assumption because the central claim has two parts: the fine-tuned GeoText-1652 gain and the zero-shot ERA result. The fine-tuned gain is internally consistent and supported by the ablation table. The zero-shot ERA result is the paper's headline generalization evidence, but Table 3 compares across heterogeneous backbones without reporting the confound. The paper's own GeoText-1652 baseline with the same XVLM backbone reaches 38.00 zero-shot on ERA, so the 'beats VCSR' result is largely explained by backbone strength. This is a protocol gap, not an internal inconsistency, but it is directly load-bearing for the claim 'GRASP outperforms all fine-tuned ERA methods.' If the controlled re-run still shows an advantage, the concern is resolved; if not, the conclusion must be weakened. The VSP in-distribution guarantee is also overstated, but it affects the theoretical framing rather than the empirical central claim; the ablation attributes part of the gain to VSP, and a flawed justification can be corrected without overturning the result. I do not object to the GeoText-1652 fine-tuned result or the training-only inference cost claim, and I do not see evidence of fabrication. The reader's CONDITIONAL verdict is appropriate; my concern adds a specific condition but does not move the verdict.","tokens_in":13963,"tokens_out":8666,"duration_ms":80958,"concrete_test":"Re-run VCSR (or the best fine-tuned ERA method) on ERA using the same XVLM-16M backbone and the same training budget as GRASP's fine-tuning, and report its mR. If VCSR-XVLM reaches or exceeds 39.47, the claimed zero-shot superiority is an artifact of backbone; if it remains below, the method-level generalization claim is supported. Additionally, report backbone and pretraining details for every row of Table 3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest generalization claim, in Section 4.5 and Table 3, is that zero-shot GRASP reaches mR 39.47 on ERA and thereby outperforms the best fine-tuned ERA method (VCSR, mR 38.96). This comparison is uncontrolled. Table 3 lists no parameter counts, pretraining data, or backbone for the fine-tuned ERA methods; the listed models (VSE++, PVSE, CLIP, PCME, AMFMN, GALR, VCSR) are standard retrieval architectures, not XVLM-family models. GRASP is built on XVLM pretrained on 16M image-text pairs. The same-backbone GeoText-1652 baseline already reaches 38.00 mR zero-shot on ERA, only 0.96 below VCSR's fine-tuned 38.96, so the backbone alone nearly closes the gap. Thus the headline comparison conflates method contribution with backbone/pretraining advantage. The within-backbone zero-shot comparison (GRASP 39.47 vs GeoText-1652 38.00) is the valid one and supports a positive but smaller claim. The paper should either re-run a strong ERA baseline with the same XVLM backbone or explicitly drop the 'outperforms all fine-tuned ERA methods' wording.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GRASP, a training-time framework for fine-grained image-text retrieval in drone views, built on an XVLM backbone. GRASP combines Region-Focused Alignment (RFA), which adds region-level contrastive and ranking losses, with Semantic Perturbation Enhanced Matching (SPEM), which constructs textual and visual hard negatives via a Semantic Prototype Codebook (SPC). The authors report fine-tuned mR 38.04 on GeoText-1652 versus 36.05 for the GeoText-1652 baseline, and zero-shot mR 39.47 on ERA, which they claim outperforms the best fine-tuned ERA method (VCSR, mR 38.96). The method adds no inference cost because all new losses are used only during training.","tokens_in":14233,"tokens_out":4523,"duration_ms":43500,"significance":"If the results hold, GRASP would offer a practical contribution to drone-view retrieval by improving attribute-level discrimination without changing inference-time model complexity. The strengths include the use of standard external benchmarks (GeoText-1652 and ERA), an internal ablation study, and a public code release. The two proposed mechanisms are clearly motivated by the drone-view characteristics of background clutter and visual isomorphism. However, the headline zero-shot generalization claim currently rests on an uncontrolled comparison, and the main numbers lack statistical grounding, so the significance of the empirical contribution is not yet fully established.","major_comments":[{"comment":"The central claim that zero-shot GRASP (mR 39.47) outperforms the best fine-tuned ERA method (VCSR, mR 38.96) is not supported by the evidence because the comparison is confounded by backbone and pretraining differences. GRASP uses an XVLM backbone pretrained on 16M image-text pairs, whereas the fine-tuned ERA baselines listed (VSE++, PVSE, CLIP, PCME, AMFMN, GALR, VCSR) use different, generally smaller backbones with unknown or different pretraining. This is not an apples-to-apples comparison. The same-backbone zero-shot GeoText-1652 baseline already reaches mR 38.00 on ERA, only 0.96 below VCSR's fine-tuned score, showing that the backbone alone nearly closes the gap. The valid within-backbone comparison is GRASP (39.47) versus GeoText-1652 (38.00), which supports a positive but much smaller claim. The authors should either run a same-backbone fine-tuned ERA baseline or explicitly retract the claim of outperforming all fine-tuned ERA methods.","section":"Section 4.5, Table 3"},{"comment":"The main GeoText-1652 results are reported as single point estimates with no error bars, multiple seeds, or significance tests, yet the paper uses the word 'significant' for the +3.3% R@1 improvement. Given that the mR differences between GRASP and the GeoText-1652 baseline are relatively small (38.04 versus 36.05), measurement noise could affect the conclusions. The authors should report at least three runs with standard deviations or conduct a paired significance test to substantiate the improvement before using 'significant' in the text.","section":"Section 4.3, Table 1"},{"comment":"The VSP method relies on a load-bearing assumption that is asserted rather than demonstrated: that attention-derived foreground patch selection (Eq. 12) yields clean object features, that the SPC prototypes span the relevant attribute manifold, and that the convex combination in Eq. (15) stays within the in-distribution manifold. The paper states that this 'guarantees' safety against out-of-distribution noise, but no empirical evidence is provided. If attention weights are contaminated by background or prototypes do not align with semantic attributes, the synthesized negatives could be mislabeled and degrade the fine-grained matching loss (Eq. 18). The authors should provide evidence for the manifold-constrained property, for example by visualizing prototypes, measuring the distance of synthesized features to the training distribution, or ablating with random prototypes instead of SPC-based ones.","section":"Section 3.3, Eqs. (12)-(15)"}],"minor_comments":[{"comment":"The TSP method depends on 'predefined dictionaries' of colors, orientations, and antonyms, but the paper does not describe how these dictionaries were constructed, their size, or their coverage. This information is necessary for reproducibility and should be included in the supplementary material or described in the main text.","section":"Section 3.3, Eq. (8)"},{"comment":"The two-stage training schedule initializes the SPC using features from the stage-1 model and then fine-tunes with VSP using that same codebook. This self-referential training loop is not discussed; the authors should comment on whether results are sensitive to the quality of the stage-1 model and whether the SPC could be iteratively updated or re-initialized for stability.","section":"Section 4.2"},{"comment":"The ablation table is difficult to read because the checkmark columns do not unambiguously map to the configurations described in the text. For example, the sentence 'applying SPEM with semantic perturbations (TSP+VSP) improves performance to 37.01 %' does not clearly match any single row in the rendered table. Please use explicit configuration labels or separate rows for each component combination.","section":"Section 4.4, Table 2"},{"comment":"The description of the ERA dataset states only that it contains 2,864 drone-view images; the paper should specify the number of text queries and confirm that the official evaluation splits and retrieval protocol are used, so that the zero-shot numbers are interpretable.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is plausible and the GeoText-1652 internal comparison appears sound, but the headline zero-shot claim is currently overstated because the ERA comparison does not control for backbone and pretraining. The lack of error bars also weakens the quantitative claims. These issues are fixable within the scope of a revision by adding a same-backbone fine-tuned baseline, running multiple seeds, and providing a more careful analysis of the VSP manifold assumption. I would not recommend rejection, but the revision needs to address these points before the paper is ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core result here is credible and useful. GRASP is a competent engineering contribution: it takes a standard XVLM backbone and adds region-level contrastive learning, a margin-based ranking loss, and synthetic hard negatives, with the main new twist being an attention-purified Semantic Prototype Codebook (SPC) that produces visual perturbations in feature space. The ablations are internally consistent, and the reported gain on GeoText-1652 (mR 36.05 to 38.04) is plausible for this kind of fine-grained hard-negative work. I also give credit for releasing code and for the training-only design, which keeps inference cost unchanged.\n\nThe real soft spot is the zero-shot ERA claim in Table 3. The paper says GRASP reaches 39.47 mR zero-shot and outperforms the best fine-tuned ERA method (VCSR, 38.96). That comparison is uncontrolled: VCSR and the other ERA baselines are different architectures, and their backbones and pretraining data are not listed. GRASP uses XVLM pretrained on 16M image-text pairs, and the same-backbone GeoText-1652 baseline already scores 38.00 zero-shot on ERA. So the backbone alone nearly closes the gap to VCSR. The valid comparison is within the XVLM family: GRASP gains +1.47 mR over the GeoText-1652 baseline under the same zero-shot protocol. That is a positive result, but the 'outperforms all fine-tuned ERA methods' wording should be dropped unless the authors re-run a strong ERA method with the same XVLM backbone.\n\nOther issues are more minor. There are no error bars or significance tests, and the hyperparameters (loss weights, margin, codebook size, mixing coefficient, thresholds) are numerous and hand-set without sensitivity analysis. The claim that VSP stays within the in-distribution manifold is asserted rather than demonstrated, though the construction is sensible enough that I would not call it a fatal gap. The self-referential SPC loop is not a circularity problem because the final metrics come from held-out test splits.\n\nWho is this for? Researchers working on aerial image-text retrieval, or on hard-negative synthesis for fine-grained vision-language fine-tuning generally. It deserves a serious referee, but the referee should push for a controlled ERA baseline and some robustness checks. My recommendation: send it to peer review with major revision requests focused on the ERA comparison and experimental reporting.","headline":"Solid incremental contribution to drone-view retrieval, with an overclaimed zero-shot ERA comparison that needs a controlled baseline before the generalization claim can stand.","tokens_in":14855,"tokens_out":1715,"would_cite":true,"duration_ms":17850,"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":"GRASP claims that object-centric region alignment plus synthesized attribute-level hard negatives, applied only during training, improve fine-grained image-text retrieval in drone views and transfer to unseen aerial scenes.","keywords":["natural language-guided drones","cross-modal retrieval","fine-grained understanding","region-focused alignment","semantic perturbation enhanced matching","visual isomorphism","zero-shot transfer","drone-view benchmarks"],"falsifier":"Replace the K-Means Semantic Prototype Codebook with the same number of randomly sampled foreground patches (or random in-distribution patches) and retrain; if GeoText-1652 mean recall stays at 38.04 rather than dropping toward the 36.05 baseline, the claim that foreground-purified semantic prototypes drive the visual-perturbation gain is falsified.","tokens_in":13674,"feed_emoji":"🚁","tokens_out":6959,"duration_ms":62770,"temperature":0.7,"pith_summary":"The paper tries to establish a specific diagnosis and cure for fine-grained image-text retrieval from drones. It argues that overhead views create two failure modes: global matching is dominated by background clutter, while top-down geometry makes different buildings look alike, and existing contrastive training does not expose the model to the hard attribute-level cases. It proposes GRASP, which adds Region-Focused Alignment (object-centric contrastive and ranking losses) and Semantic Perturbation Enhanced Matching (synthesized textual and visual hard negatives built from a foreground-purified prototype codebook). On the GeoText-1652 benchmark the framework raises mean recall from 36.05 to 38.04 over the fine-tuned baseline, and when transferred to the unseen ERA dataset it reports mean recall 39.47, above the best fine-tuned ERA method (38.96). Because both additions act only on training losses, the inference pipeline is unchanged.","feed_headline":"Zero-shot drone retrieval beats fine-tuned rivals at 39.47 mR","feed_subtitle":"Region alignment and semantic perturbations sharpen fine-grained attribute matching at zero added inference cost.","key_machinery":"The load-bearing machinery is a pair of training-only losses. Region-Focused Alignment computes a region embedding from the attention-weighted patches inside each target bounding box and optimizes a region-level image-text contrastive loss together with a margin-based ranking loss that pushes positives apart from the hardest in-batch negatives. Semantic Perturbation Enhanced Matching synthesizes hard negatives: Textual Semantic Perturbation replaces attribute words (colors, orientations, then general antonyms) through deterministic dictionaries, and Visual Semantic Perturbation builds a Semantic Prototype Codebook, defined as K-Means centroids of attention-selected foreground patches updated by exponential moving average, then interpolates a region feature with a prototype sampled from the top-$k$ nearest neighbors to make a counterfactual visual negative. A binary cross-entropy fine-grained matching loss supervises the fusion encoder over positive, in-batch hard, and synthesized pairs.","core_discovery":"The central claim is that fine-grained attribute understanding in drone imagery can be improved without any retrieval-time cost by reshaping only the training objectives. Starting from the same XVLM backbone used by the GeoText-1652 baseline, GRASP reports mean recall 38.04 on GeoText-1652 versus 36.05 for the baseline and 36.15 for a prior negative-augmentation method, with the largest gains in image-query R@1 (+3.3). In zero-shot transfer to ERA after training only on GeoText-1652, it reports mean recall 39.47, exceeding the best fine-tuned ERA method (VCSR, 38.96) and improving text-query R@1 by +2.37 over the GeoText-1652 zero-shot baseline. Ablations attribute the gain to both mechanisms, with visual perturbation contributing more than textual perturbation once region-level alignment is present, which the authors read as evidence that visual isomorphism is the harder bottleneck. The paper's stated mechanism is that object-centric region alignment suppresses background, while semantic perturbation forces the model to discriminate subtle attributes such as roof color and materials.","pith_inferences":["Implicit extension: if the reported gains are real, the same training-time recipe could transfer to other overhead-imagery retrieval tasks such as satellite-to-text matching or cross-view geolocalization, where background clutter and viewpoint flattening are equally severe; the paper does not test this.","Implicit extension: the codebook of foreground prototypes could be inspected as an attribute lexicon (colors, roof materials), offering a built-in explanation channel for why a retrieval succeeded; this interpretability use is not developed in the paper.","Implicit extension: because textual perturbation uses deterministic dictionaries, its gains may depend on vocabulary coverage; a direct test would compare dictionary substitution against a language-model-based substitution while controlling for syntactic drift.","Implicit extension: the zero-shot result hints that synthesizing hard negatives during training can substitute for target-domain fine-tuning; testing on a third aerial benchmark would show whether that substitution generalizes."],"forward_implications":["On the paper's numbers, the same backbone reaches higher retrieval accuracy on GeoText-1652 without any change at inference time, so the recipe can be added to existing drone-view retrieval pipelines for free at deployment.","The larger improvement on image-query R@1 implies GRASP specifically attacks the visual-isomorphism bottleneck, suppressing geometrically similar but attribute-mismatched buildings.","Zero-shot results on ERA suggest the learned attribute discrimination transfers across aerial domains, meaning a model trained on one region can be deployed on a new one without fine-tuning.","Because visual perturbation contributes more than textual perturbation, the paper's analysis favors investing in visual attribute-manifold modeling when geometric cues are flattened in overhead views.","The method retains a grounding loss for bounding-box supervision, so retrieval quality still depends on the quality of the box annotations used during training."],"supporting_citations":[{"why":"Supplies the GeoText-1652 benchmark, the fine-tuning setup, and the 36.05 mR baseline that GRASP claims to improve.","marker":"[8]"},{"why":"Provides the XVLM backbone and multi-grained pre-training that GRASP fine-tunes.","marker":"[37]"},{"why":"The feature-space negative-construction baseline (NAS) that GRASP compares against on GeoText-1652.","marker":"[32]"},{"why":"Supplies the ERA dataset and the VCSR method whose fine-tuned mR of 38.96 GRASP claims to beat in zero-shot transfer.","marker":"[15]"},{"why":"Provides the image-text contrastive loss design that RFA's global and region-level contrastive losses build on.","marker":"[19]"},{"why":"Establishes the CLIP-style global contrastive baseline and zero-shot evaluation context for drone-view retrieval.","marker":"[26]"}],"fun_headline_variants":["GRASP: region alignment and semantic perturbation sharpen drone retrieval","Zero-shot drone retrieval beats fine-tuned on unseen ERA data","Training-only tweaks give drone retrieval a zero-cost boost","Region alignment plus perturbation learning sharpens drone retrieval","GRASP: zero-shot drone retrieval beats fine-tuned at zero inference cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the attention-weighted foreground patches are clean object regions and that interpolating a region feature with a nearby prototype yields valid in-distribution hard negatives; if attention is contaminated by background or prototypes do not correspond to real attributes, the synthesized negatives could mislabel examples and erase the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["GRASP: region alignment and semantic perturbation sharpen drone retrieval","Zero-shot drone retrieval beats fine-tuned on unseen ERA data","Training-only tweaks give drone retrieval a zero-cost boost","Region alignment plus perturbation learning sharpens drone retrieval","GRASP: zero-shot drone retrieval beats fine-tuned at zero inference cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001496,"raw_usage":{"total_tokens":6040,"prompt_tokens":1019,"completion_tokens":5021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":4937}},"tokens_in":635,"tokens_out":5021,"duration_ms":33775,"temperature":1.0,"reasoning_tokens":4937,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:24:36.407737+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the K-Means Semantic Prototype Codebook with the same number of randomly sampled foreground patches (or random in-distribution patches) and retrain; if GeoText-1652 mean recall stays at 38.04 rather than dropping toward the 36.05 baseline, the claim that foreground-purified semantic prototypes drive the visual-perturbation gain is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GeoText-1652 benchmark, the fine-tuning setup, and the 36.05 mR baseline that GRASP claims to improve."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The feature-space negative-construction baseline (NAS) that GRASP compares against on GeoText-1652."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ERA dataset and the VCSR method whose fine-tuned mR of 38.96 GRASP claims to beat in zero-shot transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the image-text contrastive loss design that RFA's global and region-level contrastive losses build on."}],"review_version":1}