{"id":"6b548990-f466-4b9f-8591-1ba9890ac3b7","arxiv_id":"2601.12500","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"GD3A and DVTrack, driven by optimal-transport descriptor matching with an adaptive dustbin score, set state-of-the-art results on a new moving-drone dense-crowd counting and tracking benchmark.","lead":"Researchers introduce MovingDroneCrowd++, a large video dataset of dense crowds filmed from fast-moving drones, plus two methods that count unique people across a video and track them. On this benchmark, the methods roughly halve counting error and improve tracking accuracy by about 40–65% versus prior approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-label supervision in Eq. 17 assumes each head patch translates rigidly between frames; under moving-drone scale/rotation this may systematically mislabel the pixel matches that GD3A/DVTrack depend on, and the paper never validates it.","rationale":"The reader's weakest assumption identifies the same mechanism I would stress: the only source of pixel-level supervision for the descriptor matching is the rigid-translation heuristic in Eq. 17. This is not a disagreement with an external consensus but an internal correctness risk: if the drone's motion induces scale or rotation, the pseudo-labels are biased, and because OT has hard marginal constraints, the bias can propagate to cross-identity matches in dense crowds. I considered whether the evaluation-number inconsistencies (e.g., 39.2% vs 64.6% HOTA improvement) are more load-bearing; they are real and should be corrected, but they affect the reporting of the result rather than the mechanism that produces it. The central methodological claim—that pixel-level OT matching and voting, rather than explicit localization, are responsible for the gains—stands or falls on the validity of the training correspondences. Since the paper provides no check of Eq. 17, the conditional verdict is appropriate. The concrete test I propose would settle it directly. Therefore the reader's CONDITIONAL verdict should remain unchanged.","tokens_in":22483,"tokens_out":10155,"duration_ms":108246,"concrete_test":"Use the dataset's head-center/identity annotations: for each training frame pair, estimate a homography H from the corresponding head centers (RANSAC), or a per-head affine from the head boxes. For every matched head pair (p_t, p_{t+δ}) and every offset Δ with ||Δ||∞<r (using the feature-map pixel radius), compute residual e = ||(p_{t+δ}+Δ) - H(p_t+Δ)||. Report the mean and the fraction of pairs with e > 1 feature-map pixel (or > 0.5 pixel). If the fraction is large, Eq. 17 is systematically wrong. Then retrain GD3A with pseudo-labels generated by the estimated homography/affine instead of Eq. 17 and compare test MAE (Table II) and HOTA (Table III); a material change settles whether the rigid-translation supervision is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The training signal for the descriptor association that drives both GD3A and DVTrack is generated by Eq. 17, which assigns pixel-level correspondences p_t+Δ ↔ p_{t+δ}+Δ for all Δ in a local head patch. This is a pure-translation model of apparent motion. On a moving drone with changing altitude, camera angle, and perspective, the true mapping of a head patch is closer to a per-pedestrian homography/affine transform; the error of Eq. 17 grows with |Δ| and with the amount of scale/rotation. Because the OT problem in Eqs. 9–12 enforces one-to-one marginals, a systematic label bias can force descriptors of neighboring heads into wrong cross-identity matches in dense crowds, which would directly corrupt the shared/outflow/inflow decomposition used for counting and the descriptor votes used for tracking. Training samples δ randomly from 3 to 8 frames, while test-time intervals in Fig. 7(a) extend to 6 s, so the regime in which the assumption is most likely to break is exactly the regime the paper claims robustness in. No experiment or error analysis is provided for Eq. 17; the paper's central claim that pixel-level matching drives the gains therefore rests on an untested approximation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MovingDroneCrowd++, a dataset of 120 video clips captured by moving drones in dense, large-scale crowd scenes with full head-bounding-box and identity annotations, and two methods: GD3A for video individual counting and DVTrack for multi-object tracking. GD3A decomposes a predicted global density map into shared, inflow, and outflow density maps using pixel-level pedestrian descriptor matching via optimal transport with an adaptive dustbin score. DVTrack converts the same descriptor matches into instance-level associations through descriptor voting. On MovingDroneCrowd++, the authors report a 47.4% reduction in counting MAE and a 39.2% (or, per the abstract, 64.6%) improvement in tracking HOTA over prior methods; experiments on the fixed-camera VSCrowd benchmark also show state-of-the-art counting results. The paper includes ablations on the matching components, the adaptive dustbin score, frame sampling interval, and sensitivity to global density estimation.","tokens_in":22888,"tokens_out":6352,"duration_ms":57108,"significance":"If the reported results are validated, the paper makes a strong contribution: MovingDroneCrowd++ is the largest moving-drone dense-crowd video dataset with full trajectory annotations, and the descriptor-association design is a plausible route to avoiding hard localization in dense crowds. Strengths include the public release of code, dataset, and pretrained models; evaluation on an external fixed-camera benchmark (VSCrowd); and a well-structured ablation program (Tables V–VI, Fig. 7). However, the current evaluation tables contain serious internal inconsistencies, and the pseudo-label model that supervises the descriptor matching (Eq. 17) is unvalidated for the moving-drone regime. These issues affect the central quantitative claims and must be resolved before the results can be considered credible.","major_comments":[{"comment":"The headline numbers do not agree across tables. GD3A MAE is 40.11 in Table II, 32.25 in Table VI (row 'Ours' with ADS), and 66.61 in Table V (row with all components). Since Table VI is an ablation of the dustbin score and Table V an ablation of the matching process, the fully configured model should appear with consistent MAE; the paper does not specify which subset (test vs validation) or which frame-interval/global-density setting each table uses. Additionally, Table III reports improvement arrows that do not match the tabulated numbers: HOTA 20.9→34.4 is +64.6%, not +39.2%, and AssA 17.6→41.2 is +134%, not +40%. The abstract states a tracking improvement of 64.6% while the introduction and conclusion state 39.2%. These inconsistencies must be reconciled and the reporting settings for each table stated explicitly.","section":"§V.B, Tables II, III, V, VI"},{"comment":"The pseudo-label supervision assumes a rigid translation of head patches: p_t+Δ ↔ p_{t+δ}+Δ for all Δ with ||Δ||_∞ < r. For a moving drone with changing altitude and camera angle, the apparent motion of a head includes scale, rotation, and perspective distortion; the error grows with |Δ| and with the frame interval δ. Training samples δ randomly from 3 to 8 frames, while the test-time evaluation in Fig. 7(a) extends to 6 s, so the low-validity regime is exactly the one the paper claims robustness in. Since Eq. 17 generates the supervision for the descriptor association that drives both GD3A and DVTrack, a systematic label bias could corrupt the density decomposition and the voting-based tracking. The paper provides no validation of this assumption (e.g., an analysis of pseudo-label correctness or a comparison with an affine/homography-based pseudo-label). This is load-bearing and needs a","section":"§IV.D, Eq. (17)"},{"comment":"DRNet is described as an 'OT-based instance-level matching method' in the ablation, but in §V.B.1 and Table II it is grouped under 'Localization-based VIC Methods'. This inconsistency makes it impossible to interpret the adaptive-dustbin comparison: if DRNet is not an OT-based method, the controlled comparison is not what the text claims. Please clarify the categorization and describe the exact DRNet configuration used in the ablation.","section":"§V.C.2, Table VI"},{"comment":"The robustness claim with respect to frame sampling interval is based only on a qualitative plot. No quantitative numbers, error bars, or a description of the subset/scoring protocol are given. Provide the MAE/RMSE values at representative intervals (e.g., 1s, 3s, 6s) for GD3A and the compared methods so the reader can verify the claimed stable performance up to 6 s.","section":"§V.C.3, Fig. 7(a)"}],"minor_comments":[{"comment":"The column header 'MIOE' should be 'MOAE' per the metric definitions in §V.A.2 (Eq. 19).","section":"Table II"},{"comment":"There are missing spaces in two cells: '8.525.0' and '24.947.9' should likely be '8.5 25.0' and '24.9 47.9'.","section":"Table IV"},{"comment":"The tracking improvement is reported as 64.6% in the abstract but 39.2% in the introduction and conclusion. After reconciling Table III, please align all occurrences.","section":"Abstract / Introduction / Conclusion"},{"comment":"The first metric column is labeled 'MSE', but the paper defines and uses MAE as the counting metric; MSE is not defined. This appears to be a typo and should be corrected.","section":"Table V"},{"comment":"The method is referred to as 'DVTracker' in the text, but the method name used elsewhere is 'DVTrack'.","section":"§V.B.3"}],"recommendation":"major_revision","confidential_remarks":"The dataset contribution is potentially valuable and the method is interesting, but the inconsistent evaluation tables (Tables II/III/V/VI) and the unvalidated pseudo-label assumption in Eq. 17 make the current version unsuitable for publication. I recommend a major revision with a careful re-check of all reported numbers and a clear specification of the evaluation protocol for each table. The revised version should be re-reviewed by a referee with access to the code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a genuinely useful new dataset and a plausible way to do video individual counting and tracking from moving drones. The adaptive dustbin in the OT matching and the descriptor-voting tracking are new relative to the prior work, including the authors' own SDNet. The dataset is larger and more diverse than existing moving-drone crowd sets, and the scene-level split and external VSCrowd evaluation are the right instincts. So there is a solid core here.\n\nThe math itself is not circular: OT matching with adaptive dustbin is a clean formulation, and the decomposition of the global density map into shared/inflow/outflow follows naturally from the matching. The paper also says code/data and pretrained models are available, which is real evidence.\n\nNow the soft spots, and I think they are in proportion to the paper's central claims. The evaluation tables are not self-consistent. GD3A's MAE in Table II is 40.11, but Table V's full-configuration row is 66.61 and Table VI's 'Ours with ADS' is 32.25. Table V labels a column 'MSE' where the surrounding metrics and text say MAE. Table VI's DRNet baseline (91.79 without ADS) doesn't match Table II's DRNet (83.04). The tracking arrows in Table III don't line up with any single baseline — the abstract says a 64.6% HOTA improvement, the introduction and Table III say 39.2%, and the actual gap to DiffMOT is 65.6%. These are the kinds of numbers a reader will instantly trip over, and they're the exact numbers used to sell the paper.\n\nThe other substantive concern is Eq. 17. The pixel-level supervision is generated by assuming each head patch translates rigidly between frames — p_t + Delta maps to p_{t+delta} + Delta for all small Delta. Under a moving drone with altitude changes and perspective, that's an approximation that breaks down as scale and rotation appear, and the error grows with |Delta| and with temporal interval. The paper never examines this. Worse, the method is advertised as robust at test intervals up to 6 seconds, while training only samples delta from 3 to 8 frames. I'd like to see an ablation with affine or homography-based pseudo-labels, or at least an analysis of how label noise degrades the learned descriptors.\n\nThere are also missing hyperparameters (tau, top-K, theta, r, AGNN depth, Sinkhorn iterations) and no error bars. These are individually minor but compound the impression that the numbers aren't pinned down.\n\nWho this is for: anyone working on drone-based crowd analysis, video individual counting, or MOT in dense scenes. The dataset alone is valuable. The methods could become the reference point if the numbers are corrected and the pseudo-label issue is addressed. I'd send it to a competent referee and ask for those fixes, not desk-reject it. But I would not currently trust the exact percentages.","headline":"Worth a serious referee: the dataset is a real contribution and the method ideas are plausible, but the evaluation tables are internally inconsistent enough that the headline gains can't be trusted until the numbers are corrected.","tokens_in":23337,"tokens_out":4449,"would_cite":true,"duration_ms":43511,"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":"Moving-drone video counting and tracking works by matching pixel-level pedestrian descriptors instead of locating each person.","keywords":["video individual counting","dense crowd counting","drone crowd tracking","optimal transport descriptor matching","adaptive dustbin score","density map decomposition","moving drone dataset","multi-object tracking"],"falsifier":"A direct test would be to annotate true pixel-level correspondences for a few dozen head pairs in the MovingDroneCrowd++ test set (e.g., by manual keypoint matching) and compare them against the matches predicted by GD3A; if the descriptor-matching accuracy on those verified pairs is near chance for heads that visibly change scale or rotate, the training-signal assumption is violated and the claimed mechanism would be suspect.","tokens_in":22430,"feed_emoji":"🎯","tokens_out":1214,"duration_ms":16993,"temperature":0.7,"pith_summary":"The paper claims that counting and tracking unique pedestrians in dense crowds filmed from moving drones should be done by matching pixel-level visual descriptors of heads across frames, not by detecting and localizing each individual. It introduces MovingDroneCrowd++, a large drone-video dataset with dense crowds, varied altitudes, angles, and lighting, and two methods built on this idea: GD3A counts by decomposing a global density map into shared, inflow, and outflow maps, while DVTrack turns descriptor matches into instance tracks by voting. On this dataset, the methods cut counting error by 47.4% and raise tracking HOTA by 39.2% over previous approaches. If correct, the core insight is that explicit pedestrian localization, which is fragile in dense and fast-moving aerial scenes, can be replaced by correspondence between many small features, making counting and tracking more robust.","feed_headline":"Matching head pixels, not locating people, tames drone-crowd counting","feed_subtitle":"Pixel-level descriptor association cuts counting error 47% and lifts tracking HOTA 39% on a new moving-drone benchmark.","key_machinery":"The central mechanism is optimal transport on pixel-level descriptors with an input-adaptive dustbin score. A backbone plus pre-trained density estimator filters feature maps to head regions; each head yields multiple descriptors. These are position-enhanced and refined by an attentional graph neural network, then matched across frames by solving a Sinkhorn-regularized transport problem whose cost matrix includes a learnable dustbin score predicted from the current descriptor pair (rather than a fixed dataset-level constant). The matching matrix decomposes global density maps into shared/inflow/outflow components, and a reverse top-k strategy plus Hungarian assignment over descriptor votes y","core_discovery":"The authors propose that pixel-level optimal-transport matching of pedestrian head descriptors, with a per-frame-pair adaptive dustbin score, is enough to drive accurate video-level individual counting and tracking, without ever explicitly localizing or detecting pedestrians. GD3A uses these correspondences to split each frame's global density map into shared, inflow, and outflow density maps, summing inflow to get the unique-person count; DVTrack reuses the same matches to build a voting matrix that assigns pedestrians across frames. The central claim is that intra-group matching errors among descriptors of the same pedestrian do not hurt the final result, and that pixel-level association a","pith_inferences":["The rigid-translation pseudo-label (Eq. 17) used to supervise pixel-level matches is only an approximation: under a moving drone, head motion between frames includes scale, rotation, and perspective changes, so the training signal may inject systematic errors that the paper does not measure.","A natural extension is to test whether the same pipeline works on other aerial platforms (e.g., fixed-wing or high-altitude) or on crowd sub-tasks like density-based flow estimation, where the inflow/outflow decomposition could be reused.","Because the method avoids explicit localization, it may also be transferable to dense counting of other small objects (e.g., animals or vehicles) in aerial video, if a descriptor-level matching signal can be constructed.","The claimed robustness to density-map error suggests a possible soft-supervision variant that trains the counter jointly with the matcher, rather than using a frozen pre-trained counter, potentially improving performance further."],"forward_implications":["If the descriptor-matching approach generalizes, video individual counting in dense crowds no longer depends on accurate head localization, which typically degrades in crowded low-altitude aerial views.","A single learned association module can serve both counting and tracking, since DVTrack derives trajectories from the same pixel-level correspondences without extra training.","The adaptive dustbin score implies performance should hold across a wide range of frame intervals and drone speeds, as the paper's interval-sweep experiment shows.","Because the method is density-map based and avoids cross-frame cross-attention, it should be computationally cheaper than prior VIC methods while being more interpretable, since inflow/outflow maps are explicit.","The dataset, being the largest moving-drone dense-crowd video benchmark with full trajectory annotations, should set a new evaluation standard for both counting and tracking under aerial motion."],"fun_headline_variants":["Pixel-matching, not detection, tames drone-crowd counting","Ditch detection: pixel matches count crowds from drones","Drone crowds tamed by pixel-level head matching, not detection","Counting drone crowds without locating a single person","New benchmark: pixel matches beat person detection for drone crowds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The pixel-level supervision assumes that a pedestrian head translates rigidly between nearby frames (p + Δ maps to p' + Δ for all small offsets), but a moving drone also changes scale, rotation, and perspective, so those pseudo-labels are not exactly correct.","fun_headline_variants_meta":{"raw":{"variants":["Pixel-matching, not detection, tames drone-crowd counting","Ditch detection: pixel matches count crowds from drones","Drone crowds tamed by pixel-level head matching, not detection","Counting drone crowds without locating a single person","New benchmark: pixel matches beat person detection for drone crowds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3037,"prompt_tokens":826,"completion_tokens":2211,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":2129}},"tokens_in":570,"tokens_out":2211,"duration_ms":14955,"temperature":1.0,"reasoning_tokens":2129,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T09:44:25.950513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to annotate true pixel-level correspondences for a few dozen head pairs in the MovingDroneCrowd++ test set (e.g., by manual keypoint matching) and compare them against the matches predicted by GD3A; if the descriptor-matching accuracy on those verified pairs is near chance for heads that visibly change scale or rotate, the training-signal assumption is violated and the claimed mechanism would be suspect.","supporting_citations":[],"review_version":1}