{"id":"a5ce9398-40b3-47e5-b7fe-591ea3eec034","arxiv_id":"2607.08075","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A training-free pipeline that periodically redetects text-query objects, propagates masks in short segments, and stitches local tracks into global IDs outperforms transferred OV-VIS baselines on a new UAV benchmark.","lead":"This paper defines a new task: finding and segmenting objects in drone videos from free-form text queries, and proposes AeroTrack, a training-free system that combines existing vision models with periodic detection and short-segment mask propagation. It also introduces AeroVIS, a benchmark of 8,279 trajectories over 9 UAV categories, and reports that AeroTrack beats transferred open-vocabulary video instance segmentation methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported superiority may be inflated by per-category thresholds tuned on AeroVIS test videos; benchmark-mask circularity is a secondary concern.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test does not change that. I identify the per-category threshold tuning as the single most load-bearing concern because it directly biases the comparative claim (Table 1) independent of benchmark-mask quality. The paper says only that thresholds are configured at the category level; unless a validation split was used, this is test-set adaptation. The circularity concern is genuine but weaker: AeroTrack with SAM2 (not SAM3) also outperforms DEVA by ~13 HOTA, so sharing the GT mask generator is unlikely to explain the whole margin. A concrete re-evaluation with validation-tuned thresholds would settle the issue. If the margin persists, the central claim is solid; if it shrinks, the paper should be revised to report fixed or validation-tuned thresholds. Code/data release status is inconsistent (abstract says upon acceptance, full text says released with GitHub URL); this should be clarified but is secondary.","tokens_in":15180,"tokens_out":10716,"duration_ms":98550,"concrete_test":"Re-run all AeroTrack variants and the six baselines on a new 80/20 split of AeroVIS, choosing per-category thresholds only on the validation 20% (and also a single global threshold), then report HOTA on the held-out 80%. If the ordering and margins remain close to Table 1, the claim is robust; if the gap narrows materially (e.g., >5 HOTA) or reverses, the reported margins are artifacts of test-set threshold tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that AeroTrack's margin over transferred OV-VIS baselines (Table 1: best 55.39 HOTA vs 34.98 DEVA, 25.61 SAM3) is achieved under a fair protocol. The paper states (Implementation Details/Setup): 'we adopt unified inference settings, configuring only the single thresholds of the Recognizer and Segmenter at the category level while keeping all other hyperparameters unchanged.' No AeroVIS validation split is described, and no evidence is given that these per-category thresholds were selected without using the evaluation videos. If thresholds were set to maximize performance on the AeroVIS test set, the comparison is optimistic: the baselines are evaluated with 'recommended configurations' and do not receive analogous per-category adaptation. Because HOTA is sensitive to detection precision/recall, this tuning could close a substantial part of the reported gap. The reader's primary concern—SAM3-generated GT masks partly circular with SAM3-based AeroTrack—is real but partially mitigated: AeroTrack variants using SAM2 as Segmenter (Table 1: 47.55/48.42 HOTA) also outperform the baselines by large margins, so the advantage is not solely explained by sharing the mask generator. The threshold-tuning issue, by contrast, affects every variant and the headline comparison directly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces UAV-OVVIS, a new task for open-vocabulary video instance segmentation in UAV footage, together with AeroTrack, a training-free framework that composes existing visual foundation models. AeroTrack periodically runs an open-vocabulary detector (YOLO-World, Grounding DINO, or SAM3) on key frames, uses a promptable segmenter (SAM2 or SAM3) to propagate masks within short segments, and employs a Lifecycle-aware ID Association (LIA) module to merge segment-level local IDs into globally consistent video identities. The authors construct AeroVIS, a benchmark with 9 categories and 8,279 trajectories derived from UAV tracking datasets, and report that the best AeroTrack variant (Grounding DINO + SAM3) achieves 55.39 Overall HOTA versus 25.61 for Native SAM3 and 34.98 for DEVA. They also evaluate on YouTube-VIS 2019/2021 and LV-VIS and provide memory/speed analyses showing that segmented propagation scales better than running SAM3 over whole videos.","tokens_in":15558,"tokens_out":4194,"duration_ms":43587,"significance":"If the empirical claims hold, the paper makes a useful contribution: it defines a practically relevant task, provides a modular and training-free strong baseline, and constructs the first UAV-specific benchmark for open-vocabulary video instance segmentation. The five variants are internally consistent, the LIA ablation is large, and the memory/speed analysis in Figure 5 is concrete and informative. The method is not a derivational circularity: no parameters are fitted to the reported numbers, and the underlying VFMs are external. The main value, however, rests on the trustworthiness of the AeroVIS benchmark and the fairness of the comparison protocol, and those are currently not established with sufficient rigor.","major_comments":[{"comment":"The paper states that 'unified inference settings' are used, 'configuring only the single thresholds of the Recognizer and Segmenter at the category level while keeping all other hyperparameters unchanged.' No AeroVIS validation split is described, and no evidence is given that these per-category thresholds were selected without tuning on the evaluation videos. Baselines are evaluated with 'recommended configurations' and do not receive analogous per-category adaptation. Since HOTA is sensitive to detection precision/recall through DetA, this asymmetry could explain a substantial part of the reported margin (55.39 vs 34.98 DEVA and 25.61 SAM3). Please clarify how thresholds were chosen, add a held-out validation protocol, report threshold values and sensitivity, and/or apply the same category-level threshold adaptation to the baselines.","section":"Implementation Details/Setup; Table 1"},{"comment":"AeroVIS ground-truth masks are generated by SAM3 from box/identity priors with 'manual reinspection,' while the strongest AeroTrack variants use SAM3 as the Segmenter (Table 1). This is a partial evaluative circularity. No inter-annotator agreement, number of corrected masks, or per-category audit statistics are reported, so it is impossible to quantify whether the benchmark systematically over-rewards SAM3-style mask predictions. The SAM2-based variants also outperform baselines, which mitigates the concern, but the absolute HOTA values and the comparison with Native SAM3 are not fully interpretable without independent annotation-quality analysis. Please report mask-quality statistics and a subset of masks generated/verified by an independent pipeline or annotators.","section":"Figure 3; Dataset"},{"comment":"The +27 to +32 HOTA gain from LIA is large, but the 'w/o LIA' condition leaves local IDs unassociated across segments, so under the global-ID HOTA protocol the low baseline is partly by construction. To substantiate the LIA design as opposed to merely having any cross-segment association, please compare LIA with a simpler association method (e.g., pure IoU/center-distance greedy matching without lifecycle decay) and ablate the lifecycle parameters (rho_min, rho_max, gamma, tau_life, T_lost) individually. This is not central to the headline comparison, but it is load-bearing for the paper's component-level claim.","section":"Table 3; Ablation Study"}],"minor_comments":[{"comment":"The SAM3 result on LV-VIS is from the test split, while all other results are from the val split. Direct comparison across splits is misleading; use the same split or clearly separate the footnote.","section":"Table 2"},{"comment":"The text says 'which we further analyze in §.' with a missing section number. Please fill in the reference.","section":"Comparison on AeroVIS"},{"comment":"There are widespread OCR-like spacing artifacts ('UA V', 'traffic', 'difficult', 'official'). These should be cleaned before publication.","section":"Throughout"},{"comment":"The caption lists queries on the same line ('bicycle car road ...'), which is hard to parse. It would be clearer to format each query as a separate item or use a table.","section":"Figure 4"},{"comment":"AeroVIS is described as containing 117 videos and 8,279 trajectories, but no explicit train/validation/test split is given. Since the paper tunes thresholds on the benchmark, a formal split should be defined and released.","section":"Dataset"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within scope for a CV venue and the proposed framework is sensible. The central issue is empirical: the headline comparison may be inflated by threshold tuning on the evaluation videos, and the benchmark masks share a generator with the strongest method. Both issues are addressable with a validation-based protocol and annotation audit, so I would not reject. If the authors can show that the thresholds are stable across settings or that the baselines also benefit from tuning, the paper could become a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about UAV video perception or open-vocabulary VIS. The new piece is the task definition and the AeroVIS benchmark: 117 videos, 9 categories, 8,279 trajectories built from VisDrone/UAVDT/SeaDronesSee with SAM3-generated masks plus manual reinspection. That is a real resource. AeroTrack itself is not architecturally novel — it is a staged composition of published detectors and segmenters with periodic key-frame detection and an output-level association module (LIA) using hand-set geometric gates and matching weights. But the LIA ablation is large and consistent across five variants (+27 to +32 HOTA), which suggests segment-level propagation without identity association genuinely fails on long dense UAV videos. The memory analysis is also useful: Native SAM3 runs out of memory at budget 7 while AeroTrack variants handle budgets more than 10 times larger.\n\nThe soft spots are in the evaluation protocol, not in the framework's internal logic. First, category-level Recognizer and Segmenter thresholds are tuned on AeroVIS, with no validation split described, while baselines are run with recommended configurations. The headline comparison (55.39 vs 34.98 DEVA, 25.61 Native SAM3) could shrink appreciably if thresholds were selected with test-set knowledge; HOTA is sensitive to detection precision/recall. Second, AeroVIS GT masks are generated by SAM3, the same model family used as Segmenter in the best AeroTrack variants. That is partly circular, though mitigated by the SAM2-based variants also beating baselines by large margins (48.42 vs 34.98), so the advantage is not entirely explained by sharing the mask generator. Still, no audit statistics or inter-annotator agreement are reported. Third, no error bars, and dataset and code are not yet released despite the promise.\n\nNone of this is fatal. The benchmark and framework are useful, and the concerns are addressable with a validation split, a threshold-tuning analysis, an independent mask audit, and released artifacts. As it stands, I trust the qualitative claims and resource analysis more than the exact numbers in Table 1. Send to review, but insist on these fixes.","headline":"Useful new UAV benchmark and training-free baseline, but the headline margins are vulnerable to threshold tuning on the test set and SAM3-generated GT masks.","tokens_in":15996,"tokens_out":2522,"would_cite":true,"duration_ms":22753,"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 claims that a training-free pipeline—periodic open-vocabulary detection, short-segment mask propagation, and lifecycle-aware identity association—can outperform existing open-vocabulary video instance segmentation methods on UAV v","keywords":["UAV video perception","open-vocabulary video instance segmentation","training-free framework","visual foundation models","instance segmentation trajectories","dense small objects","video instance segmentation","AeroTrack"],"falsifier":"Re-annotate a random subset of AeroVIS (e.g., 20 videos) with independent human annotators who do not use SAM3, then recompute overall HOTA for the best AeroTrack variant and for native SAM3; if the 55.39-vs-25.61 gap shrinks substantially, the central comparison depends on the benchmark's SAM3-generated masks.","tokens_in":15106,"feed_emoji":"🚁","tokens_out":6834,"duration_ms":58816,"temperature":0.7,"pith_summary":"The paper introduces a new task—open-vocabulary video instance segmentation for UAV footage—and a training-free pipeline, AeroTrack, that reuses off-the-shelf visual foundation models. AeroTrack periodically detects targets in key frames, propagates masks within short segments, and uses a Lifecycle-aware ID Association (LIA) module to stitch segment-level identities into globally consistent trajectories. The authors build AeroVIS, a 117-video, 9-category benchmark with 8,279 trajectories, and report that AeroTrack variants outperform general open-vocabulary VIS methods transferred to this benchmark, with the best variant reaching 55.39 overall HOTA versus 25.61 for native SAM3. They also show that LIA contributes +27 to +32 HOTA, and that segmented propagation expands the target budget more than tenfold without slowing inference. A sympathetic reader would take this as evidence that UAV open-vocabulary instance segmentation does not require domain-specific training.","feed_headline":"Training-free pipeline beats all transferred baselines on drone video","feed_subtitle":"No drone-specific training: one detector, one segmenter, and an ID-association module beat prior methods on UAV video.","key_machinery":"AeroTrack itself is the central mechanism: an Open-Vocabulary Recognizer runs text-conditioned detection every Δ frames, a Video Instance Segmenter propagates masks within the resulting segment and resets its memory at each boundary, and LIA (Lifecycle-aware ID Association) maps segment-local IDs to video-global IDs via a lifecycle bank of active and lost trajectories, hard geometric gating (scale ratio and center distance), and a matching score combining IoU, center similarity, and area similarity. This three-way split converts a long, dense, open-set video problem into bounded sub-problems: periodic detection covers newly entering/reappearing targets, short-segment propagation keeps memory","core_discovery":"The central claim is that UAV open-vocabulary video instance segmentation can be decomposed into three coordinated, training-free stages—periodic open-vocabulary detection on key frames, short-segment mask propagation with internal state resets, and output-level identity association—and that this decomposition is both effective and resource-efficient. On the authors' AeroVIS benchmark, all five AeroTrack variants beat the best transferred general-scene method; the top variant (Grounding DINO for detection, SAM3 for segmentation) reports 55.39 overall HOTA and higher DetA than AssA, indicating that detection coverage of text-specified targets, not identity association for already-found target","pith_inferences":["Because AeroVIS ground-truth masks are generated by SAM3 with manual reinspection, and the best AeroTrack variants also use SAM3 as segmenter, part of the performance gap may reflect shared model bias; an independent human-annotated subset would settle this.","The recognize-segment-associate decomposition is not UAV-specific; the same pattern could apply to other long-video, dense-scene, open-set tasks (surveillance, wildlife monitoring) where per-frame detection is too expensive.","LIA operates only on output-level boxes, so it is a plug-and-play identity-repair module that could be paired with any segmenter that resets state, independent of the specific detector and segmenter choices.","The refresh interval Δ is a manually fixed trade-off; adaptively choosing Δ based on scene motion or target density is a natural next step that the paper does not explore."],"forward_implications":["If the reported results hold, a training-free composition of existing visual foundation models is a strong and immediately usable baseline for UAV open-vocabulary video instance segmentation, removing the need for domain-specific annotations.","Segmented propagation with state resets cuts peak memory growth so that pipelines can handle more than ten times the target budget of whole-video propagation (at least 80 dense targets in the stress test), without sacrificing speed.","LIA's consistent +27 to +32 HOTA gain shows that identity fragmentation across segment boundaries—not within-segment mask quality—is the main obstacle to global trajectories in this setting.","AeroVIS, with 9 UAV categories and 8,279 trajectories, provides a reproducible evaluation protocol for future open-vocabulary instance segmentation research on UAV video.","Generalization results on YouTube-VIS 2019/2021 and LV-VIS indicate that the UAV-specific design does not degrade open-vocabulary capability on ordinary video."],"fun_headline_variants":["Training-free drone video segmenter beats all baselines","Zero-training open-vocab UAV segmentation outperforms prior","AeroTrack: no extra training, tops transferred methods","UAV video instance segmentation via foundation models only","Drone video: coordinate 3 big models, no retraining needed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The AeroVIS benchmark's ground-truth masks are generated by SAM3, the same model family used inside the best AeroTrack variants, with only manual reinspection and no reported inter-annotator agreement; if those masks contain systematic errors, the reported HOTA margins are inflated.","fun_headline_variants_meta":{"raw":{"variants":["Training-free drone video segmenter beats all baselines","Zero-training open-vocab UAV segmentation outperforms prior","AeroTrack: no extra training, tops transferred methods","UAV video instance segmentation via foundation models only","Drone video: coordinate 3 big models, no retraining needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1433,"prompt_tokens":794,"completion_tokens":639,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":558}},"tokens_in":538,"tokens_out":639,"duration_ms":6833,"temperature":1.0,"reasoning_tokens":558,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:56:07.816200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a random subset of AeroVIS (e.g., 20 videos) with independent human annotators who do not use SAM3, then recompute overall HOTA for the best AeroTrack variant and for native SAM3; if the 55.39-vs-25.61 gap shrinks substantially, the central comparison depends on the benchmark's SAM3-generated masks.","supporting_citations":[],"review_version":2}