{"id":"2cddc035-24f1-44ef-9c9a-37ba1f8f9aa2","arxiv_id":"2504.20234","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A point-oriented SORT variant with camera-motion compensation, altitude-aware assignment, trajectory classification, and reused-feature correlation filters reduces trajectory counting error to 15% on UP-COUNT-TRACK and 23% on DroneCrowd.","lead":"This paper presents an online tracking algorithm for counting people from drone footage, combining a point-based SORT tracker with camera motion compensation, altitude-aware matching, a validation classifier, and correlation filters that reuse the detector's features. On two drone video datasets it roughly halves counting error and sharply cuts identity switches. Why read it: a practical method plus a new annotated dataset for a hard small-object tracking problem.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed ID-SW and counting gains hinge on DDCF, but the paper never verifies that its feature-based re-detections recover the same person; without such a check, the gains could stem from trajectory stitching rather than true continuity.","rationale":"The paper's central contribution is an online tracker whose most dramatic reported gains come from DDCF; every other component changes metrics only slightly. On UP-COUNT-TRACK, cumulative ID-SW before DDCF is 2943 and after DDCF it is 287, a reduction of about 90%. The reader's weakest assumption identifies exactly this dependency: the correlation filter's re-localisation is assumed to preserve identity. My reading of the method section confirms that this is not checked internally: DDCF is initialized on the last localization and visual features, and its output is accepted as the trajectory position (Section 4.5), but no experiment verifies that this output corresponds to the same pedestrian. The near-flat HOTA despite the large ID-SW drop is a red flag that merits explanation, since HOTA's association component should improve if fragmentation is genuinely reduced. The proposed test would settle the ambiguity by measuring per-object re-identification success against a cheap baseline. Until such a test is run, the paper's strongest claim should remain conditional. I would not reject the work because the dataset, ablations, and reproducible pipeline are valuable and the concern is addressable.","tokens_in":12084,"tokens_out":5324,"duration_ms":60723,"concrete_test":"Conduct an identity-preservation audit on both datasets. For each ground-truth trajectory, start from a correctly matched frame, artificially withhold detections for k = 5, 10, 20, and 30 frames, run only the DDCF module from the last true detection, and record its response location each frame. Label a response a success if it lies within the dynamic assignment radius Tr of the same GT person's annotated position and is farther from every other GT person's position than from the true one; otherwise classify it as drift to a neighbor, drift to background, or lost. Repeat the same simulation with a constant-velocity Kalman-only baseline that has no DDCF.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables 2 and 3 attribute the decisive improvements to DDCF: adding it cuts ID-SW from 2943 to 287 on UP-COUNT-TRACK and from 6645 to 388 on DroneCrowd, while Tr-nMAE drops from 0.30 to 0.15 and from 0.30 to 0.23. HOTA, which explicitly rewards correct association, stays nearly flat (0.63 to 0.63; 0.52 to 0.54). The mechanism behind this gain is asserted, not verified. Section 4.4.1 reuses the 16-channel second-to-last-layer feature map (544x940x16) of a people-localisation U-Net, a representation trained to detect persons, not to distinguish identities. For sub-ten-pixel objects in dense crowds, a per-object ECO filter built on these features has no demonstrated ability to re-localise the same individual after a missed-detection gap; it may lock onto a neighbor or background. The paper reports no per-track re-identification accuracy, no failure analysis, no comparison to cheaper gap-filling such as Kalman prediction, and no HOTA decomposition showing association gains. The headline reductions are therefore consistent with an alternative explanation: DDCF stitches unrelated detections or hallucinations into long trajectories, which suppresses ID-SW and lowers counting error without genuinely preserving identity. This assumption is load-bearing because without it, the central contribution reduces to unvalidated post-processing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online, point-oriented multi-object tracker for drone-based crowd monitoring, built on SORT with a point-distance assignment metric. It adds three minimal-cost techniques: camera motion compensation (CMC), altitude-aware dynamic assignment thresholds, and a classification-based trajectory validation step. It further integrates Deep Discriminative Correlation Filters (DDCF) that reuse spatial feature maps from the Dot Localisation network, aiming to maintain trajectory continuity during missed detections. The method is evaluated on the newly released UP-COUNT-TRACK dataset and on DroneCrowd, reporting reduced identity switches (ID-SW) and trajectory counting errors (Tr-nMAE) relative to a baseline SORT variant and to the offline GOG algorithm.","tokens_in":12421,"tokens_out":2555,"duration_ms":27590,"significance":"If the DDCF module genuinely preserves object identity during missed-detection gaps, the paper offers a practical, computationally efficient online recipe for tracking sub-ten-pixel people in drone footage, with a valuable new benchmark dataset (UP-COUNT-TRACK) and a zero-waste reuse of localisation features. The incremental ablation design is commendable: Tables 2 and 3 show that each added component either improves or holds the reported metrics. However, the central claimed breakthrough—the drastic ID-SW reduction attributed to DDCF—is not supported by direct evidence of per-object re-identification fidelity, and the comparison set lacks established online trackers. The dataset contribution alone is significant for the community.","major_comments":[{"comment":"The DDCF module is credited with the dominant improvement: ID-SW drops from 2943 to 287 on UP-COUNT-TRACK and from 6645 to 388 on DroneCrowd, while HOTA remains essentially flat (0.63 to 0.63; 0.52 to 0.54). The paper does not demonstrate that the correlation-filter-based re-detections recover the same physical person; for objects of only a few pixels, the 16-channel feature map from the localisation head may not be identity-discriminative, and the filter could lock onto a neighbor or background. Without a per-track re-identification accuracy, a failure analysis, a comparison against cheaper gap-filling (e.g., Kalman prediction alone), or a HOTA decomposition showing association gains, the ID-SW reduction is equally consistent with trajectory stitching that artificially prolongs tracks. This is load-bearing: the central claim of improved trajectory continuity hinges on the DDCF's identity-preserving behaviour.","section":"Section 4.4-4.5 and Tables 2-3"},{"comment":"The comparison includes only the paper's own baseline with incremental components and the offline GOG algorithm. The abstract and Section 1 claim the method 'outperform[s] baseline online trackers', but no established online multi-object trackers (e.g., DeepSORT, ByteTrack, BoT-SORT, or a standard point-based SORT variant) are evaluated. Since SORT is acknowledged as the base, a comparison against a well-tuned SORT/DeepSORT would be necessary to substantiate the practical improvement claim.","section":"Section 5.3"},{"comment":"The altitude-aware threshold for DroneCrowd relies on assuming 100 m for 'small' objects and 50 m for 'large' objects because the dataset lacks altitude metadata. The paper does not justify these specific values or report sensitivity to them. Given that the threshold directly controls assignment radii and thus ID-SW, a sensitivity analysis (e.g., varying the assumed altitudes or the constant in Eq. (1)) is needed to show the reported gains are not artifacts of tuned assumptions.","section":"Eq. (1) and Section 4.3.2"},{"comment":"The classification-based validation uses an 80% average-probability threshold stated as 'determined experimentally', but no ablation on this threshold is reported, and the training/evaluation protocol for the classifier is underspecified: is the classifier trained on UP-COUNT-TRACK training split and then evaluated on DroneCrowd, or trained separately per dataset? The accuracy figures in Section 5.5 (0.987 and 0.939) do not indicate whether the classifier generalises across domains or whether the threshold was tuned on the test sets.","section":"Section 4.3.3 and Section 5.5"}],"minor_comments":[{"comment":"There is a stray spacing issue: 'drone -based' should be 'drone-based'.","section":"Abstract"},{"comment":"The numeric formatting of '49 .13± 117 .22' and similar entries inserts spaces before decimal points; these should be corrected throughout.","section":"Tables 2 and 3"},{"comment":"The spatial feature map dimensions 544×940×16 are mentioned without stating the input image resolution to which they correspond; adding this context would help readers judge the feature stride relative to object size.","section":"Section 4.4.1"},{"comment":"The definitions of Tr-MAE and Tr-nMAE use n for the number of sequences, but it is not explicitly stated that y_i and ŷ_i are the ground-truth and estimated numbers of unique trajectories per sequence; please make this explicit, as the metric is central to the counting claims.","section":"Section 5.2"},{"comment":"The BoT-SORT reference is incomplete: it lists 'Aharon, N., Orfaig, R., Bobrovsky, B., .' with no year; the arXiv identifier is given but the citation entry should be completed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The dataset contribution is a clear positive, and the incremental ablation is useful. However, the DDCF identity-preservation concern is central: the paper should either provide direct evidence of re-identification accuracy or temper the claims. The lack of online tracker baselines also weakens the practical significance. I would encourage the editors to request code and detailed per-sequence results, as the reported averaging across sequences may hide failure modes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives you a clean, honest engineering story: a point-based SORT variant with camera motion compensation, altitude-aware assignment, trajectory validation, and an ECO-style correlation filter that reuses the localisation network's features. The new UP-COUNT-TRACK dataset, with 33,751 frames and 1.36M labelled instances, is a real public resource and probably the most durable part of the work. The ablations are incremental and the tables support the central trend: each added component holds or improves the metrics, and DDCF produces a dramatic drop in ID-SW on both datasets. The authors also state their computational limitations plainly and release code and data, which I respect.\n\nThe soft spot is exactly what the stress-test note flags. The ID-SW reduction from DDCF is enormous—from 2943 to 287 on UP-COUNT-TRACK, from 6645 to 388 on DroneCrowd—yet HOTA, which directly rewards correct association, stays essentially flat (0.63 to 0.63 and 0.52 to 0.54). That flatness is a red flag. If the tracker were genuinely preserving identities, HOTA should move. The likely explanation is that the correlation filter prolongs trajectories without re-identifying the same person, suppressing ID-SW because the tracker never switches IDs on its own hallucinations. The paper never verifies per-track re-identification accuracy, never compares against a cheap gap-filler like pure Kalman prediction, and never reports a HOTA association breakdown. That is a load-bearing gap in the central claim.\n\nOther issues are minor by comparison. The baseline comparison is narrow—only the offline GOG method, not other online trackers like DeepSORT or SORT with IoU. The DroneCrowd altitude assumptions (100 m and 50 m) are stated but not validated. The 80% classification threshold and the altitude formula are engineering choices without sensitivity analysis. And the abstract implies real-time feasibility, which the limitations section later contradicts.\n\nI would send this to peer review. The dataset alone justifies referee time, and the tracking claims are important enough that the DDCF question needs to be settled publicly. A reviewer should ask for a verification of the filter's per-object re-detection reliability, a simpler gap-filling baseline, and a HOTA decomposition. If those come back clean, this is a useful contribution. As it stands, treat the headline ID-SW numbers with caution.","headline":"A solid engineering contribution with a new public tracking benchmark, but the headline DDCF gains are under-verified and could partly reflect trajectory stitching rather than true identity preservation.","tokens_in":12917,"tokens_out":1656,"would_cite":true,"duration_ms":19109,"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":"This paper claims that a point-oriented SORT tracker augmented with camera-motion compensation, altitude-aware assignment, classification-based trajectory validation, and correlation filters reusing the localizer's own features reduces…","keywords":["drone crowd monitoring","multi-object tracking","trajectory continuity","discriminative correlation filters","point-based localisation","altitude-aware assignment","identity switches","crowd counting"],"falsifier":"Take the full pipeline and disable only the DDCF re-localisation, then additionally suppress detections in a random sample of 10–30 frame windows; if the large ID-switch and Tr-nMAE improvements vanish, the filter is the cause, whereas if they survive, the other modules are carrying the gain. A second check is to crop the filter's re-localised positions and have humans or a strong detector judge whether the same person was found.","tokens_in":11864,"feed_emoji":"🚁","tokens_out":7276,"duration_ms":62335,"temperature":0.7,"pith_summary":"This paper argues that the main obstacle to reliable drone-based crowd monitoring is not detection accuracy but trajectory continuity: people appear as points a few pixels wide, so detector misses and false positives fragment tracks and inflate counting errors. The proposed method keeps the cheap online SORT skeleton but swaps bounding-box overlap for Euclidean point-distance assignment, adds camera-motion compensation, an altitude-dependent matching radius, a classification check before a track is confirmed, and Deep Discriminative Correlation Filters that reuse the localiser's own spatial features. On DroneCrowd and the new UP-COUNT-TRACK dataset this full combination lowers trajectory-counting error to 23% and 15%, respectively, and cuts identity switches from thousands to hundreds, outperforming an offline global-optimisation baseline. A sympathetic reader would take this as evidence that online, nearly free repair of track continuity can replace expensive offline tracking for tiny objects.","feed_headline":"Drone crowd tracker cuts counting error to 15 percent","feed_subtitle":"A cheap set of online tracking fixes keeps tiny people on one trajectory, beating offline optimizers.","key_machinery":"The load-bearing mechanism is the reuse of the localisation network's decoder features: a 544×940×16 spatial feature map from the second-to-last layer of the head is extracted once per frame, and object-sized crops are fed into an ECO-style Deep Discriminative Correlation Filter that re-localises a confirmed trajectory whenever its detection is missing. Alongside it, association uses a circular gate with altitude-dependent radius $T_r = \\max(10, \\frac{100}{\\text{altitude}}\\cdot 10)$ pixels, and a lightweight convolutional classifier must give an average probability above 80% before a candidate track is confirmed. Together these mechanisms decide which points belong to which person and keep that decision alive through detector dropouts.","core_discovery":"The paper's central claim is that a point-oriented tracker can recover long, stable trajectories for sub-ten-pixel people without a new detector or offline optimisation. Replacing IoU-based association with a point-distance circle whose radius is set by flight altitude, compensating drone motion with an affine transform from sparse optical flow, validating unconfirmed tracks with a small classifier, and bridging missed detections with a Deep Discriminative Correlation Filter built on the localisation model's second-to-last-layer features reduces DroneCrowd counting error from 32% to 23% and UP-COUNT-TRACK counting error from 37% to 15%, while dropping identity switches from 6290 to 388 and from 3305 to 287. The method runs online, frame by frame, and the correlation filter adds little compute because it consumes features already computed for detection.","pith_inferences":["The reported ID-switch collapse may be partly a consequence of artificially prolonging trajectories through detection gaps; a direct way to test this is to compare tracks produced with and without the DDCF module on sequences where detector outputs are deliberately suppressed.","If the 16-channel features carry enough identity information for few-pixel objects, the same feature-reuse pattern could extend to other point-based dense prediction tasks, such as vehicle or animal counting, that output spatial feature maps from their heads.","The altitude rule uses only a scalar radius; incorporating full camera pose from GPS, IMU, and gimbal angles could generalise the assignment gate to non-horizontal cameras and curved flight paths.","The order-of-magnitude ID-switch reduction suggests many identity switches in drone datasets are artifacts of detection gaps rather than association ambiguities; detector training might benefit from a temporal continuity loss that rewards stable features across frames."],"forward_implications":["Trajectory-counting error falls to 23% on DroneCrowd and 15% on UP-COUNT-TRACK, so counting people by unique trajectories becomes a practical route to crowd-size estimates.","Identity switches drop from 6290 to 388 on DroneCrowd and from 3305 to 287 on UP-COUNT-TRACK, which would make per-person movement and behaviour analysis feasible instead of just counting.","Because the enhanced online tracker beats the offline greedy global-optimisation method on UP-COUNT-TRACK and is competitive on DroneCrowd, the paper implies offline processing is not required for good trajectory continuity.","The altitude-dependent assignment radius lets a single configuration handle flights from roughly 29 m to 100 m above ground, removing a tuning step for different drone heights.","Reusing localiser features avoids a second feature-extraction network, but the paper notes the overall pipeline remains computationally demanding and is not yet real-time on board a drone."],"supporting_citations":[{"why":"Supplies the SORT baseline with Kalman filter and Hungarian assignment that the point-distance tracker modifies.","marker":"Bewley et al. (2016)"},{"why":"Provides the ECO Efficient Convolution Operators framework that the DDCF module adapts for re-localisation.","marker":"Danelljan et al. (2017)"},{"why":"The Dot Localisation method whose detections and second-to-last-layer spatial features the tracker consumes, and the source of the UP-COUNT dataset extended here.","marker":"Ptak and Kraft (2025)"},{"why":"Source of the Camera Motion Compensation module using sparse optical flow and an affine transform to correct Kalman predictions.","marker":"Aharon et al."},{"why":"Contributes the DroneCrowd benchmark, the T-mAP evaluation procedure, and the object-size based altitude assumptions.","marker":"Wen et al. (2021)"},{"why":"The globally optimal greedy offline tracker used as the state-of-the-art comparison baseline.","marker":"Pirsiavash et al. (2011)"},{"why":"Defines HOTA, the primary multi-object tracking metric reported in the evaluation.","marker":"Luiten et al. (2021)"}],"fun_headline_variants":["Point-based drone tracker cuts counting error to 15%","Drone tracker reduces crowd counting errors by 59%","Online drone tracker beats offline optimizers with point matching","Drone tracker keeps tiny people on track with point association","Drone tracker uses point centroids to reduce ID switches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme leans on the assumption that the 16-channel feature map from the localisation head's second-to-last layer still distinguishes one few-pixel person from another person and from background during missed-detection intervals; if those features are not identity-discriminative at that scale, the correlation filter can drift and the ID-switch and counting gains would reflect prolonged but wrong trajectories rather than true tracking.","fun_headline_variants_meta":{"raw":{"variants":["Point-based drone tracker cuts counting error to 15%","Drone tracker reduces crowd counting errors by 59%","Online drone tracker beats offline optimizers with point matching","Drone tracker keeps tiny people on track with point association","Drone tracker uses point centroids to reduce ID switches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3199,"prompt_tokens":949,"completion_tokens":2250,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2170}},"tokens_in":565,"tokens_out":2250,"duration_ms":17891,"temperature":1.0,"reasoning_tokens":2170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:34:24.559192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the full pipeline and disable only the DDCF re-localisation, then additionally suppress detections in a random sample of 10–30 frame windows; if the large ID-switch and Tr-nMAE improvements vanish, the filter is the cause, whereas if they survive, the other modules are carrying the gain. A second check is to crop the filter's re-localised positions and have humans or a strong detector judge whether the same person was found.","supporting_citations":[{"cited_title":", author Ge, Z","cited_arxiv_id":null,"evidence_quote":"Supplies the SORT baseline with Kalman filter and Hungarian assignment that the point-distance tracker modifies."},{"cited_title":", author Bhat, G","cited_arxiv_id":null,"evidence_quote":"Provides the ECO Efficient Convolution Operators framework that the DDCF module adapts for re-localisation."},{"cited_title":", author Du, D","cited_arxiv_id":null,"evidence_quote":"Contributes the DroneCrowd benchmark, the T-mAP evaluation procedure, and the object-size based altitude assumptions."},{"cited_title":", author Ramanan, D","cited_arxiv_id":null,"evidence_quote":"The globally optimal greedy offline tracker used as the state-of-the-art comparison baseline."},{"cited_title":", author Osep, A","cited_arxiv_id":null,"evidence_quote":"Defines HOTA, the primary multi-object tracking metric reported in the evaluation."}],"review_version":1}