{"id":"0efd91eb-3c2c-4422-8c80-e607b5497bbd","arxiv_id":"2509.06011","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper introduces UAVDE-2M and UAVCAP-15K datasets and a CAGE fusion module, claiming improvements in UAV open-vocabulary detection, but the module's own gains are modest.","lead":"A new dataset pipeline and a fusion module aim to improve open-vocabulary object detection from drone imagery. The headline gain of +5.3 mAP mostly comes from the new training data, not from the module itself.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"UAVDE-2M's LVLM-generated labels are never validated; the benchmark's core value and the reported gains depend on their accuracy.","rationale":"The paper's central contribution is a large UAV OVD benchmark plus a lightweight fusion module. The benchmark's scale is partly built from LVLM-generated labels with no validation; this is the least secure condition because the entire dataset value depends on it, and the available experiments cannot separate pseudo-label quality from the effect of domain-specific pre-training. The reader's weakest_assumption identified the same issue. I do not think this warrants rejection: the authors may be able to supply label-quality evidence or an ablation isolating COD/FOP contributions. But as written, the claim that UAVDE-2M provides 'clean yet exhaustive annotations' is unsupported, and the confidence in the central claim is therefore conditional. The CAGE attribution issue is secondary; the module's effect is directly measured in Table 2, so it is less load-bearing than the unvalidated pseudo-labels. A concrete validation study would settle the concern.","tokens_in":12232,"tokens_out":8964,"duration_ms":101023,"concrete_test":"Randomly sample ~1,000 instances from COD and FOP subsets of UAVDE-2M (e.g., ERA, FloodNet, SUES, University-1652, AeroScapes). Have two annotators with UAV/remote-sensing expertise judge whether the assigned class matches the cropped object and whether the SAM proposal is a real object; compute precision, recall, and inter-annotator agreement. If precision is high (e.g., >90% on a representative sample) and agreement is adequate, the label-quality concern is resolved; if precision is low or agreement is poor, the dataset's stated value and the derived zero-shot gains require re-evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central dataset contribution rests on pseudo-labels produced by Qwen2.5-VL for the COD and FOP subsets described in Sec. 3.1. No precision/recall measurement, human-agreement study, or comparison against the source datasets' original annotations is reported. The COD class counts in Table 1 (e.g., DenseUAV 748, ERA 838, FloodNet 772, SUES 563, University-1652 643) come entirely from the LVLM pipeline and are never checked; if these labels are noisy or inconsistent across datasets, the 'clean yet exhaustive' benchmark claim loses support. Ablation Table 4 only removes UAVDE-2M as a whole, so the contribution of the pseudo-labeled portion is not isolated. The +5.3 mAP headline is also an aggregate effect: Table 2 shows most of the gain comes from domain pre-training (8.59→12.2), with CAGE adding 12.2→13.9 on VisDrone-L and actually slightly decreasing SIMD L-scale mAP (11.2→11.0). The deepest unverified condition, however, is label quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a complete UAV-domain open-vocabulary object detection (OVD) solution, consisting of two new benchmarks (UAVDE-2M and UAVCAP-15K) built with a semi-automated 'UAV-Label Engine', and a lightweight Cross-Attention Gated Enhancement (CAGE) module that is inserted into the YOLO-World-v2 neck as a drop-in replacement for T-CSPLayer. The authors report that the combined method improves zero-shot VisDrone mAP from 8.59 to 13.9 for the L-scale model, while reducing parameters and GFLOPs, and shows cross-domain gains on the SIMD remote-sensing benchmark. They also present real-world deployment results on a Jetson Orin NX.","tokens_in":12580,"tokens_out":6079,"duration_ms":61717,"significance":"If the underlying data are reliable and the reported gains are correctly attributed, this is a useful engineering contribution to UAV-focused OVD. The scale of the proposed datasets (2.4M instances, 1,853 categories) and the inclusion of a caption dataset for pretraining are valuable resources for the community. The paper also contains a real-world embedded deployment and latency measurements, which are often missing from OVD papers. However, the central dataset contribution rests on unvalidated LVLM-generated pseudo-labels, and the headline +5.3 mAP conflates dataset and module effects. These issues substantially temper the significance unless the authors can provide additional validation and clearer attribution.","major_comments":[{"comment":"The COD and FOP subsets of UAVDE-2M are annotated using the Qwen2.5-VL LVLM pipeline, but no quantitative validation of these pseudo-labels is provided. There is no precision/recall measurement against source datasets' original annotations, no human-agreement study, and no sample-based error analysis. The class counts for COD datasets (e.g., DenseUAV 748, ERA 838, FloodNet 772) are entirely produced by the LVLM and are never checked. Since these pseudo-labels are a central contribution and are claimed to be 'clean yet exhaustive', the benchmark's reliability is unestablished. The authors should provide a label-quality study, e.g., comparing a random sample against human annotations or the original dataset labels, with per-dataset accuracy/agreement numbers.","section":"Section 3.1 / Table 1"},{"comment":"The abstract and conclusion attribute the +5.3 mAP VisDrone improvement to CAGE ('By embedding CAGE into the YOLO-World-v2 framework... improving zero-shot detection on VisDrone by +5.3 mAP'). However, Table 2 shows that this is the combined effect of domain-specific pretraining (8.59→12.2 for L-scale) plus CAGE (12.2→13.9). Controlled comparisons on the same UAV data give CAGE only +0.4 (S), +1.4 (M), and +1.7 (L) mAP. Moreover, on SIMD (Table 3) CAGE slightly decreases L-scale mAP (11.2→11.0). The isolated contribution of the module should be reported in the abstract and conclusion, and the current phrasing should be corrected.","section":"Abstract / Conclusion vs. Table 2"},{"comment":"The implementation details state that VisDrone images are excluded from pretraining, but Table 1 lists VisDrone as a component of UAVDE-2M with 8,629 images and 457,066 instances. If VisDrone is removed during training, the dataset composition used for the experiments in Table 2 differs from the published UAVDE-2M statistics. Please clarify whether the 'UAV' training data is UAVDE-2M \\ VisDrone, and provide the exact instance/image counts after exclusion. If VisDrone images are not fully excluded, the zero-shot evaluation on VisDrone is compromised.","section":"Section 5.1 / Table 1"},{"comment":"The CAGE module is composed of four proposed components: multi-head cross-attention, a gated context refinement head, a global FiLM layer, and a residual path. No ablation isolates the contribution of each component. Given that the module's controlled gains over the YOLO-World-v2 baseline are modest (+0.4 to +1.7 mAP), the necessity and individual benefit of each design choice are not demonstrated. Please add a component-wise ablation study.","section":"Section 4.1"}],"minor_comments":[{"comment":"Notation issues: 'D' is used both for depth of text embeddings and is described as 'denotes batch size'; the text embeddings dimension is defined inconsistently. Also, 'Ddonates' should be 'D denotes'. Please clean up the notation.","section":"Section 4.1"},{"comment":"Typo: 'V osDrone' should be 'VisDrone'.","section":"Figure 2 caption"},{"comment":"The text says 'our L-scale model achieves a highly competitive performance comparable to its baseline', but Table 3 shows YOLO-World-v2-L w/ CAGE (11.0) is slightly lower than YOLO-World-v2-L (11.2). Please rephrase to avoid implying a gain where there is none.","section":"Table 3 / Section 5.2"},{"comment":"The abbreviation 'UAV' is defined in the caption, but the caption does not state that the UAVDE-2M component excludes VisDrone. Please make this explicit for reproducibility.","section":"Table 2 caption"},{"comment":"The paper uses 'UA V' with a space in many places (e.g., 'UA VDE-2M', 'UA V platform'). Please standardize to 'UAV' throughout.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The dataset construction and deployment effort are potentially valuable, but the lack of validation for the LVLM-generated labels and the misleading attribution of the +5.3 mAP gain to CAGE alone are substantial issues. I would like to see a revised version that adds label-quality validation, corrects the attribution, and clarifies the VisDrone exclusion before considering it for publication. The paper is within the scope of the journal and the authors state they intend to release the datasets/code, which would be a positive contribution if the quality concerns are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful dataset contribution wrapped in an overstated method claim. The UAVDE-2M and UAVCAP-15K datasets are real work — aggregating 27 existing UAV datasets, standardizing formats, adding a partially-annotated category, and generating captions. The CAGE module is a combination of known components (cross-attention, gating, FiLM) that shows small consistent gains on VisDrone at S/M/L scales (0.4–1.7 mAP) and on SIMD-M, though it slightly hurts on SIMD-L. The deployment on Orin NX with speed numbers is a plus.\n\nThe abstract's +5.3 mAP is the main thing to fix. That number is against the OGC baseline and includes both the UAV dataset pretraining and CAGE. Table 4 shows most of the gain comes from the dataset alone (4.58→9.22), and CAGE adds 0.4–1.7. The conclusion repeats the misleading claim. Not fatal, but it undersells the real contribution.\n\nThe bigger soft spot is the label quality for the COD and FOP subsets. The paper uses Qwen2.5-VL to generate class labels for ~660K instances in COD and to re-classify ambiguous labels in FOP, but no precision/recall, no human agreement, no comparison against source annotations. The word clouds show richer classes, but richer isn't automatically correct. If those labels are noisy, the benchmark's value drops. The paper claims open-sourcing but gives no link, which makes it hard to check.\n\nThere's also a count error: the intro says \"over 3 million instances\" but Table 1 sums to ~2.4M. Minor but sloppy.\n\nThe VisDrone exclusion is clearly stated, so the leakage concern is not strong unless you suspect UAVDT/VisDrone scene overlap. Not a decisive issue.\n\nWho this is for: researchers working on UAV object detection who want a large pretraining corpus and a fast OVD model for edge deployment. As a dataset paper, it deserves referee time. The referees should ask for label validation, an artifact release, and a revised headline separating dataset and module gains. I'd accept it for review.","headline":"Useful UAV dataset benchmark, but the +5.3 mAP headline conflates dataset and module gains and the LVLM-generated labels need validation.","tokens_in":12969,"tokens_out":4578,"would_cite":false,"duration_ms":43613,"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":"The paper argues that open-vocabulary detection on drone imagery can be made practical by combining domain-specific UAV pretraining data with a lightweight cross-modal fusion module, and supports this with two new benchmarks and a CAGE modu","keywords":["open-vocabulary object detection","UAV imagery","dataset construction","vision-language pretraining","cross-modal fusion","zero-shot detection","pseudo-labeling","real-time detection"],"falsifier":"Re-annotate a random sample of about 1,000 images from the COD and FOP subsets of UAVDE-2M by human experts and measure label agreement with the Qwen2.5-VL annotations. Then train identical YOLO-World-v2 models on the full auto-labeled set and on the human-audited subset: if VisDrone mAP is unchanged while agreement is low, the auto-labeled data are not actually supplying the claimed supervision.","tokens_in":12225,"feed_emoji":"🚁","tokens_out":7315,"duration_ms":68548,"temperature":0.7,"pith_summary":"The paper argues that open-vocabulary object detection fails on drone imagery not because the models are weak but because they are trained on ground-level data. To close that gap, it builds two large UAV-view datasets—UAVDE-2M with 2.4M instances across 1,853 categories and UAVCAP-15K with richly captioned image-text pairs—using an automated label engine, and adds a lightweight cross-attention gated enhancement (CAGE) module to YOLO-World-v2. On VisDrone zero-shot evaluation, the combined contribution raises mAP from 8.59 to 13.9 for the large model while cutting parameters by about 29% and GFLOPs by about 30%, and it transfers to a different remote-sensing benchmark. The authors claim this makes practical, real-time open-vocabulary detection feasible on resource-limited UAV platforms.","feed_headline":"UAV datasets and fusion module add 5.3 mAP to zero-shot detection","feed_subtitle":"A 2.4M-instance drone dataset plus a gated fusion module lift zero-shot detection in aerial views.","key_machinery":"The load-bearing mechanism is the CAGE module: a dual-path fusion block whose cross-attention path grounds textual tokens into spatial image features, a learned per-pixel gate decides where text context is relevant so irrelevant or occluded regions are not polluted, a global FiLM path shifts and scales the whole feature map from a pooled text vector, and a residual connection preserves the original visual stream. Its role is to make a real-time detector align text and vision more cheaply than the T-CSPLayer blocks it replaces, while the datasets supply the aerial-domain priors the module needs.","core_discovery":"The central claim is that the domain gap between ground-level and UAV imagery can be overcome by domain-specific pretraining data plus a cheap fusion module, without enlarging the detector. The paper introduces a refined UAV-Label Engine that turns existing drone and remote-sensing datasets into standardized detection annotations, treating partially-annotated and ambiguously-labeled data as a distinct category and using Qwen2.5-VL for fine-grained re-classification; this yields UAVDE-2M. A second dataset, UAVCAP-15K, supplies long-form captions for vision-language pretraining. The CAGE module replaces the text-vision fusion blocks in YOLO-World-v2's neck with a dual-path design: multi-head c","pith_inferences":["Beyond the paper: because the ablation attributes most of the gain to the dataset rather than the module, a lighter path to similar results would be to pretrain an existing open-vocabulary detector on UAVDE-2M alone and skip architectural changes when deployment simplicity matters more than peak accuracy.","The gate's per-pixel relevance map could be repurposed as a signal for occlusion or domain shift, since it learns where text context is trusted; one testable extension is to check whether its activations correlate with small-object density in aerial scenes.","The same UAV-Label Engine pipeline could be pointed at other under-annotated video corpora, such as satellite or underwater footage, to bootstrap open-vocabulary detectors for those domains; this is an extension the paper does not test.","A direct test of pseudo-label quality would be to train a model on a human-audited subset of UAVDE-2M and compare its VisDrone mAP to the model trained on the full auto-labeled set; if the numbers match despite low human-model label agreement, the LVLM annotations are not actually carrying the claimed benefit."],"forward_implications":["Domain-specific pretraining on UAV-view data is the dominant source of zero-shot gain: adding UAVDE-2M alone lifts VisDrone mAP from 4.58 to 9.22, and UAVCAP-15K adds another 0.38.","CAGE improves accuracy at lower cost: at the L scale, mAP rises from 12.2 to 13.9 while parameters drop from 48M to 34M and GFLOPs from 204.5 to 144.0.","The approach transfers across domains: models pretrained on UAV data reach 11.2 mAP on SIMD versus 8.93 for general-purpose pretraining.","The trained model runs in real time on the tested embedded platform, achieving 22.9 ms per frame compared with 25.38 ms for the baseline.","Because CAGE is a drop-in replacement for the original fusion blocks, it can be transplanted to other vision-language detection frameworks without architectural changes."],"supporting_citations":[{"why":"Supplies the YOLO-World real-time open-vocabulary detection framework and the OGC pretraining baseline that the paper's datasets and CAGE module are compared against.","marker":"[7]"},{"why":"Provides the original LAE-Label Engine that the proposed UAV-Label Engine refines by introducing the partially-annotated dataset (FOP) category.","marker":"[28]"},{"why":"Qwen2.5-VL is the LVLM used for class prediction, fine-grained re-classification of ambiguous labels, and generation of UAVCAP-15K captions.","marker":"[1]"},{"why":"VisDrone is the main zero-shot evaluation benchmark whose test set supports the headline +5.3 mAP improvement claim.","marker":"[43]"},{"why":"YOLOE is the competing real-time open-vocabulary detector used in the speed comparison and baseline context.","marker":"[36]"},{"why":"Objects365 is part of the general-purpose OGC pretraining data that the paper's UAV datasets must outperform on aerial imagery.","marker":"[32]"},{"why":"GoldG is part of the OGC pretraining mixture and represents the grounding data used by the general-purpose baselines.","marker":"[18]"},{"why":"Provides the precedent that detailed image captions improve open-vocabulary detection training, motivating the construction of UAVCAP-15K.","marker":"[11]"}],"fun_headline_variants":["CAGE fusion plus 2.4M-instance drone dataset add 5.3 mAP","UAV open-vocab detection: new dataset and gated module gain 5.3 mAP","VisDrone zero-shot detection gains 5.3 mAP with UAV dataset and CAGE","Drone data and lightweight fusion boost zero-shot detection by 5.3 mAP","Aerial open-vocabulary detection improves 5.3 mAP with UAV dataset and CAGE"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole pipeline depends on the assumption that the labels generated by Qwen2.5-VL for the unlabeled and partially labeled subsets of UAVDE-2M are accurate and consistent enough to train a detector that generalizes to real UAV scenes.","fun_headline_variants_meta":{"raw":{"variants":["CAGE fusion plus 2.4M-instance drone dataset add 5.3 mAP","UAV open-vocab detection: new dataset and gated module gain 5.3 mAP","VisDrone zero-shot detection gains 5.3 mAP with UAV dataset and CAGE","Drone data and lightweight fusion boost zero-shot detection by 5.3 mAP","Aerial open-vocabulary detection improves 5.3 mAP with UAV dataset and CAGE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001027,"raw_usage":{"total_tokens":4181,"prompt_tokens":778,"completion_tokens":3403,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":3294}},"tokens_in":522,"tokens_out":3403,"duration_ms":24997,"temperature":1.0,"reasoning_tokens":3294,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:37:39.770109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a random sample of about 1,000 images from the COD and FOP subsets of UAVDE-2M by human experts and measure label agreement with the Qwen2.5-VL annotations. Then train identical YOLO-World-v2 models on the full auto-labeled set and on the human-audited subset: if VisDrone mAP is unchanged while agreement is low, the auto-labeled data are not actually supplying the claimed supervision.","supporting_citations":[{"cited_title":"Yolo-world: Real-time open- vocabulary object detection","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLO-World real-time open-vocabulary detection framework and the OGC pretraining baseline that the paper's datasets and CAGE module are compared against."},{"cited_title":"Detection and tracking meet drones challenge.IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, 44(11):7380–7399, 2021","cited_arxiv_id":null,"evidence_quote":"VisDrone is the main zero-shot evaluation benchmark whose test set supports the headline +5.3 mAP improvement claim."},{"cited_title":"Yoloe: Real-time seeing anything, 2025","cited_arxiv_id":null,"evidence_quote":"YOLOE is the competing real-time open-vocabulary detector used in the speed comparison and baseline context."},{"cited_title":"Objects365: A large-scale, high-quality dataset for object detection","cited_arxiv_id":null,"evidence_quote":"Objects365 is part of the general-purpose OGC pretraining data that the paper's UAV datasets must outperform on aerial imagery."},{"cited_title":"Mdetr - mod- ulated detection for end-to-end multi-modal understanding","cited_arxiv_id":null,"evidence_quote":"GoldG is part of the OGC pretraining mixture and represents the grounding data used by the general-purpose baselines."},{"cited_title":"Llmdet: Learning strong open-vocabulary object detectors under the supervision of large language models","cited_arxiv_id":null,"evidence_quote":"Provides the precedent that detailed image captions improve open-vocabulary detection training, motivating the construction of UAVCAP-15K."}],"review_version":1}