{"id":"756d4694-c681-40a5-a1e0-488c803a8445","arxiv_id":"2507.13359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured review that divides aerial open-vocabulary detection methods into pseudo-labeling and CLIP-driven integration families and catalogs the missing benchmarks in the field.","lead":"This survey organizes open-vocabulary object detection for drone imagery into two method families: pseudo-labeling approaches and CLIP-driven integration models. A generalist reader gets a compact map of a young field whose goal is letting drones find objects described in plain language, such as 'overturned cars' or 'temporary shelters', without retraining.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The two-family taxonomy in Section 3 is internally inconsistent: two of the six surveyed methods are placed in the 'CLIP-driven integration' family despite being described as MLLM-decoupled or representation regularization, so the survey's central organization does not hold as written.","rationale":"The strongest claim of the paper is precisely the taxonomy: 'we divide current methods into two major types.' A survey whose organizing dichotomy is internally inconsistent cannot support the claim that the literature is accurately organized by that dichotomy. I therefore treat this as the load-bearing issue rather than Table 1's protocol mismatch, which is a real but secondary problem affecting one comparative paragraph. The membership test settles the issue without external benchmarks: it only requires applying the paper's own definitions to its own descriptions. If the test shows DescReg and LLaMA-Unidetector lack integrated CLIP encoders, the text and Figure 4 must be reconciled by either renaming the family or adding branches; either revision changes the central contribution. If their source papers do use CLIP encoders, the survey should document that explicitly. This does not warrant rejection: the dataset overview and challenge discussion remain useful, and the flaw is fixable. Hence the reader's CONDITIONAL verdict stands unchanged. I mark agreement as partial because the reader's weakest assumption identified Table 1 comparability, whereas the more fundamental problem is the two-family taxonomy itself; both are conditions that a revision must address.","tokens_in":25546,"tokens_out":8713,"duration_ms":96737,"concrete_test":"Build a membership table for the six methods in Figure 4 using only the paper's own descriptions. For each method, record (1) whether a CLIP image or text encoder is actually integrated into the detection architecture, and (2) whether the method's novelty is pseudo-label generation, MLLM-based decoupled recognition, or representation regularization. Under the definitions in Sections 3.1-3.2, LLaMA-Unidetector and DescReg should fail test (1). If they do, revise the survey to present at least three families and align Figure 4 with the text, then re-run the Section 3.3 family-level comparison with those methods removed from the CLIP-driven family. If their source papers instead show CLIP encoders in the pipeline, the survey must cite that evidence and explicitly define what 'CLIP-driven' means.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3's central claim is that aerial OVOD methods divide into exactly two families: Pseudo-Labeling and CLIP-driven Integration. The paper's own Figure 4, however, shows four sub-branches under 'OVOD UVA': End-to-End Vision Language Fusion, Semi-Supervised Pseudo-Labeling, Decoupled Recognition with MLLMs, and Representation Regularization. LLaMA-Unidetector and DescReg are then discussed inside the CLIP-driven integration family in Section 3.2, even though the text describes LLaMA-Unidetector as a two-stage pipeline whose recognition is done by TerraOV-LLM (an MLLM, not CLIP) and DescReg as a method that 'does not propose a new detector architecture' and instead regularizes the embedding space with visual descriptions. Either these methods are not CLIP-driven, in which case the stated dichotomy is not exhaustive and the taxonomy has at least four families, or 'CLIP-driven' is defined so broadly that it cannot distinguish any approach. This internal inconsistency undermines the survey's main organizational contribution and also confounds the family-level comparison in Section 3.3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of open-vocabulary object detection (OVOD) applied to UAV/aerial imagery. It motivates the topic, reviews background material on traditional detection and CLIP-based alignment, proposes a taxonomy that separates methods into pseudo-labeling and CLIP-driven integration, describes six representative methods, compiles relevant datasets and evaluation metrics, and concludes with challenges and future directions. The paper is positioned as the first comprehensive survey in this specific niche.","tokens_in":25568,"tokens_out":3603,"duration_ms":46164,"significance":"If the manuscript were revised to fix its internal inconsistencies, it would be a useful entry point for researchers entering the intersection of OVOD and UAV perception: it collects recent methods, datasets, and open problems in one place, and it correctly identifies the absence of a standardized aerial OVOD benchmark as a key community gap. The descriptive portions of the survey are broadly faithful to the cited works, and the dataset table in Section 4.2 is a useful reference. However, the paper's two main intellectual contributions—the taxonomy in Section 3 and the comparative conclusions in Section 3.3—are currently undercut by an internal inconsistency in the taxonomy and by a comparison table that the text itself says is not protocol-comparable. These issues are central to the survey's organization and claims, so they require substantive revision rather than copyediting.","major_comments":[{"comment":"The central taxonomy is internally inconsistent as written. The text says that existing methods are divided into two major types, Pseudo-Labeling and CLIP-driven integration, but Figure 4 displays four sub-branches: End-to-End Vision Language Fusion, Semi-Supervised Pseudo-Labeling, Decoupled Recognition with MLLMs, and Representation Regularization. LLaMA-Unidetector is discussed in Section 3.2 as a CLIP-driven integration method even though the text explicitly describes its recognition stage as TerraOV-LLM, an MLLM, not CLIP; DescReg is also placed in Section 3.2 even though the text says it 'does not propose a new detector architecture' and regularizes the embedding space with visual descriptions. Either the taxonomy should be expanded to the four families shown in Figure 4, or the definition of 'CLIP-driven integration' should be made precise enough that an MLLM-decoupled pipeline and a representation-regularization method are genuinely instances of it. As presented, the two-family claim does not organize the surveyed methods, and this weakens the family-level comparison in Section 3.3.","section":"Section 3, Figure 4"},{"comment":"The comparative conclusion that 'pseudo-labeling-based algorithms generally achieve superior performance on novel classes' is not supported by the evidence in Table 1. The table mixes mAP-based evaluations with AP50-based evaluations, and the text itself states that 'these scores are not directly comparable to the mAP-based results due to the differing evaluation criteria.' The mAP rows do not state the base/novel split definitions used by each source paper, and the novel-class comparison driving the narrative is based on two contemporaneous methods, DescReg and CastDet, plus one later method, OVA-DETR. In addition, the AP50 rows are incomplete for the dataset columns, with dashes for LAE-DINO on DOTA and OPEN-RSD on DIOR, making those rows difficult to read. The family-level claim should either be removed or restricted to a direct comparison under a single, explicitly described evaluation protocol; at minimum, the table should report the split definition and metric for each method rather than mixing protocols.","section":"Section 3.3, Table 1"},{"comment":"The manuscript does not disclose that two of the works it assesses are co-authored by the survey's authors: OVA-DETR [49] includes H. Zhang, and MI-OAD [81] includes X. Xue. Because the survey makes favorable comparative statements about OVA-DETR in Section 3.2 and about MI-OAD in Section 4.2, this overlap is material to a reader's ability to calibrate the evaluation. I am not suggesting misconduct, but a survey with comparative conclusions about the authors' own methods should include a conflict-of-interest or author-contribution statement, and ideally should separate the description of those works from the comparison section.","section":"Section 3.2 and Section 4.2"}],"minor_comments":[{"comment":"There are numerous language and typographical errors, including 'an indispensable tools,' 'application scenarios have extend from,' 'UA V' spacing, 'Figure 3)(a),' and 'OVOD UVA' in Figure 4; a thorough proofreading pass is needed.","section":"Throughout"},{"comment":"The text discusses the STAR dataset [82] but STAR is omitted from the dataset statistics in Table 2; adding it would make the table consistent with the narrative.","section":"Section 4.2, Table 2"},{"comment":"The claim that this is 'the first comprehensive survey' would be easier to evaluate if the manuscript described its search and inclusion criteria; without such a statement, the completeness claim is hard to verify.","section":"Section 1"},{"comment":"The illustration text in Figure 2 repeats 'An image of Parking Lot' four times; this is presumably a placeholder and should be cleaned up.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The taxonomy inconsistency and the non-comparable performance table are the two issues that, together, require a major revision. If the authors resolve those and add a disclosure for the self-authored methods, the survey could become a useful reference for the community. I would also encourage the editor to ask the authors to confirm that the GitHub link and figure labels are corrected before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first survey I've seen specifically on OVOD in UAV imagery, and it does fill a gap. The best material is the dataset roundup and the explicit argument that the field's biggest bottleneck is the absence of a standardized open-vocabulary benchmark with agreed base/novel splits. That point is made clearly, and it's the most actionable thing in the paper. The challenge list (domain gap, small objects, fine-grained, efficiency, prompt ambiguity, missing benchmark) is sensible.\n\nThe real problem is the taxonomy. Section 3 says methods split into two families, pseudo-labeling and CLIP-driven integration. But Figure 4 shows four branches: end-to-end vision-language fusion, semi-supervised pseudo-labeling, decoupled recognition with MLLMs, and representation regularization. LLaMA-Unidetector and DescReg are then discussed in the CLIP-driven section even though the text describes the first as a two-stage MLLM pipeline and the second as a representation-regularization method, not a CLIP-integrated detector. Either the survey has four families or 'CLIP-driven' is so broad it loses meaning. This is not a cosmetic point: the family-level comparison in Section 3.3 leans on CastDet vs DescReg to conclude that pseudo-labeling methods do better on novel classes. With the classification in question, that conclusion rests on a label, not on the methods' actual design.\n\nThe other issues are fixable but real. Table 1 mixes mAP and AP50 results, and the paper admits they're not directly comparable, yet still draws comparative conclusions. The base/novel splits are not stated. Two favorably assessed works (OVA-DETR, MI-OAD) have one of this survey's authors on the author list; that should have been disclosed. And the mechanical debt is heavy: VisDrone is cited to the wrong reference, [56,57] and [83,84] are off-topic, STAR appears in the text but not in Table 2, and the prose needs copyediting.\n\nI'd guess the survey is broadly faithful to the cited papers, and the descriptive content is probably usable. But the central organizing claim needs to be reworked before this can serve as the reference for the subfield. For a newcomer it's still a reasonable entry point; for people in the area, the benchmark argument is the takeaway.\n\nRecommendation: send it to peer review, with a major-revision request. The taxonomy inconsistency and the mixing of evaluation protocols are both fixable, and the topic deserves a good survey. I would not desk reject it.","headline":"A useful but sloppy survey of UAV open-vocabulary detection; the dataset overview and benchmark diagnosis are worth having, but the two-family taxonomy is undercut by the paper's own figure, and the comparison table mixes incompatible metrics.","tokens_in":26318,"tokens_out":4561,"would_cite":true,"duration_ms":49069,"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":"Open-vocabulary object detection in drone imagery can be organized into pseudo-labeling and CLIP-driven integration families, and the field's main obstacle is the lack of a standard benchmark.","keywords":["Open-vocabulary object detection","UAV imagery","Aerial object detection","Vision-language models","CLIP","Pseudo-labeling","Remote sensing","Survey"],"falsifier":"Running CastDet, OVA-DETR, and DescReg under one shared base/novel split on DOTA v1.0 with the same evaluation code would settle the family comparison; if a CLIP-driven method then matches or exceeds CastDet on novel mAP, the paper's central comparative claim collapses. Alternatively, identifying an aerial OVOD method from the same period that fits neither the pseudo-labeling nor the CLIP-driven family would falsify the taxonomy's exhaustiveness.","tokens_in":25151,"feed_emoji":"🚁","tokens_out":8788,"duration_ms":84394,"temperature":0.7,"pith_summary":"This review argues that open-vocabulary object detection (OVOD) is the next step for drone perception, because it lets an aerial system search for objects described in natural language without retraining on new categories. The paper's central contribution is a taxonomy that sorts existing aerial OVOD methods into two families: pseudo-labeling methods, which use a teacher-student loop to label unlabeled drone images, and CLIP-driven integration methods, which fuse text-image alignment directly into the detector. It reviews representative systems in each family, surveys the datasets used to train and test them, and lays out a research agenda of challenges from the ground-to-aerial domain gap and small objects to prompt ambiguity and missing benchmarks. The survey matters because it maps a rapidly growing field and identifies the missing standardized benchmark as the main obstacle to fair progress.","feed_headline":"Two families emerge in open-vocabulary drone detection","feed_subtitle":"Pseudo-labeling and CLIP-driven methods are the two families; the missing benchmark blocks fair comparison.","key_machinery":"The organizing device is the taxonomy itself: the two-way split between pseudo-labeling methods and CLIP-driven integration methods, anchored in the contrastive image-text alignment learned by CLIP. Pseudo-labeling methods rely on a teacher-student flywheel in which an external vision-language model labels proposals that a student detector learns from, while CLIP-driven methods reframe detection as region-text grounding inside a single network. The comparison rests on the standard OVOD evaluation setup of separate mAP on base and novel classes combined into a harmonic mean, which is the metric that exposes the claimed novel-class advantage of pseudo-labeling.","core_discovery":"On its own terms, the paper establishes that the literature on open-vocabulary object detection in UAV imagery can be organized by a single dichotomy: methods that generate pseudo-labels for unlabeled aerial data (CastDet and the MLLM-based discovery pipeline) versus methods that integrate CLIP-style vision-language alignment into the detector itself (OVA-Det, LAE-DINO, OpenRSD, LLaMA-Unidetector, and DescReg). It further claims that under the mAP-based protocol on DIOR and DOTA v1.0, pseudo-labeling methods such as CastDet generalize better to novel classes, while CLIP-driven methods such as DescReg retain higher base-class accuracy. The paper presents this as a trade-off, and it closes by arguing that the field's most urgent need is a community-standard UAV-OVOD benchmark with agreed base/novel splits and evaluation metrics.","pith_inferences":["The family-level comparison in Section 3.3 is weaker than it looks: the mAP rows mix different base/novel splits, and the AP50 rows use a different metric, so the novel-class ordering rests on very few contemporaneous methods and a unified protocol could reverse it.","The taxonomy may not stay exhaustive: methods that combine pseudo-labeling with end-to-end fusion are foreseeable, and the dichotomy may need to become a spectrum as the field grows.","The datasets built by automatic label engines (LAE-1M, ORSD+, MI-OAD) suggest that the next bottleneck will be label quality and vocabulary coverage, which the survey's challenge list does not yet foreground."],"forward_implications":["If the taxonomy is right, future aerial OVOD work can be positioned by which family it extends, making method comparison and combination more systematic.","The claimed novel-class advantage of pseudo-labeling would motivate hybrid designs that add CLIP-driven alignment on top of teacher-student labeling.","The paper's challenge list implies that solving the domain gap and small-object alignment are prerequisites for real deployment, not optional refinements.","The absence of a standard benchmark means that reported numbers across papers cannot yet be compared fairly, so a shared base/novel split on DOTA or DIOR would be a concrete first step."],"supporting_citations":[{"why":"Supplies the cross-modal image-text alignment that OVOD builds on, the foundation for both method families.","marker":"[20]"},{"why":"CastDet, the archetypal pseudo-labeling method whose teacher-student flywheel anchors the first taxonomy branch.","marker":"[44]"},{"why":"The MLLM-based open-set discovery pipeline, the second pseudo-labeling example that extends labels to unknown objects.","marker":"[46]"},{"why":"OVA-Det, the efficient CLIP-driven integration method that anchors the second branch and contributes the real-time comparison point.","marker":"[49]"},{"why":"LAE-DINO, the performance-first CLIP-driven method whose LAE-1M dataset supplies the million-instance in-domain data claim.","marker":"[50]"},{"why":"OpenRSD, the universal CLIP-driven framework whose ORSD+ dataset supports the cross-domain generalization discussion.","marker":"[51]"},{"why":"LLaMA-Unidetector, the decoupled localization-plus-MLLM recognition method that defines the semantic-analysis end of the spectrum.","marker":"[52]"},{"why":"DescReg, the regularization method whose weak novel-class numbers set up the paper's central base/novel comparison.","marker":"[53]"},{"why":"DOTA-v1.0, one of the two benchmark datasets used in the comparative table that supports the family-level claims.","marker":"[67]"},{"why":"DIOR, the other benchmark dataset in the comparison table, with its large-scale HBB annotations.","marker":"[70]"}],"fun_headline_variants":["Drone detection: pseudo-labels vs CLIP methods","Open-vocabulary drone detection: two families, one trade-off","UAV detection survey: which method wins? Depends on metric","Missing benchmark stalls open-vocabulary drone detection","Pseudo-labels beat CLIP for novel classes in drone images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's claim that pseudo-labeling methods generally beat CLIP-driven methods on novel classes depends on numbers collected from different papers, datasets, base/novel splits, and evaluation criteria being comparable, something the paper itself concedes is not yet the case.","fun_headline_variants_meta":{"raw":{"variants":["Drone detection: pseudo-labels vs CLIP methods","Open-vocabulary drone detection: two families, one trade-off","UAV detection survey: which method wins? Depends on metric","Missing benchmark stalls open-vocabulary drone detection","Pseudo-labels beat CLIP for novel classes in drone images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1250,"prompt_tokens":996,"completion_tokens":254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":169}},"tokens_in":612,"tokens_out":254,"duration_ms":3355,"temperature":1.0,"reasoning_tokens":169,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:14:48.968031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Running CastDet, OVA-DETR, and DescReg under one shared base/novel split on DOTA v1.0 with the same evaluation code would settle the family comparison; if a CLIP-driven method then matches or exceeds CastDet on novel mAP, the paper's central comparative claim collapses. Alternatively, identifying an aerial OVOD method from the same period that fits neither the pseudo-labeling nor the CLIP-driven family would falsify the taxonomy's exhaustiveness.","supporting_citations":[{"cited_title":"Toward open vocabulary aerial object detection with clip-activated student-teacher learning","cited_arxiv_id":null,"evidence_quote":"CastDet, the archetypal pseudo-labeling method whose teacher-student flywheel anchors the first taxonomy branch."},{"cited_title":"Advancing open-set object detection in remote sensing using multimodal large language model","cited_arxiv_id":null,"evidence_quote":"The MLLM-based open-set discovery pipeline, the second pseudo-labeling example that extends labels to unknown objects."},{"cited_title":"Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community","cited_arxiv_id":null,"evidence_quote":"LAE-DINO, the performance-first CLIP-driven method whose LAE-1M dataset supplies the million-instance in-domain data claim."},{"cited_title":"OpenRSD: Towards Open-prompts for Object Detection in Remote Sensing Images","cited_arxiv_id":"2503.06146","evidence_quote":"OpenRSD, the universal CLIP-driven framework whose ORSD+ dataset supports the cross-domain generalization discussion."},{"cited_title":"LLaMA-Unidetector: An LLaMA-Based Universal Framework for Open-Vocabulary Object Detection in Remote Sensing Imagery","cited_arxiv_id":null,"evidence_quote":"LLaMA-Unidetector, the decoupled localization-plus-MLLM recognition method that defines the semantic-analysis end of the spectrum."},{"cited_title":"Zero-shot aerial object detection with visual description regularization","cited_arxiv_id":null,"evidence_quote":"DescReg, the regularization method whose weak novel-class numbers set up the paper's central base/novel comparison."},{"cited_title":"DOTA: A large-scale dataset for object detection in aerial images","cited_arxiv_id":null,"evidence_quote":"DOTA-v1.0, one of the two benchmark datasets used in the comparative table that supports the family-level claims."},{"cited_title":"Object detection in optical remote sensing images: A survey and a new benchmark","cited_arxiv_id":null,"evidence_quote":"DIOR, the other benchmark dataset in the comparison table, with its large-scale HBB annotations."}],"review_version":1}