{"id":"efbdefb1-c07a-4a61-b218-cb817b4c184e","arxiv_id":"2411.10072","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A custom tracker plus three-region counter reaches 97% counting accuracy on an edge device, though the paper's own long-term test shows 92%.","lead":"This paper combines a fine-tuned head detector, a feature-matching tracker, and a three-region counter to count people passing through a doorway from an overhead camera. It reports 97% counting accuracy at 20-27 FPS on a low-power edge computer, targeting smart building and safety applications.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 97% counting-accuracy claim is not reproducible because Section II-D's three-region counting rule never defines how Region B transitions, partial crossings, or oscillations are handled.","rationale":"The reader's weakest assumption—that people walk straight through the doorway without lingering or reversing—is precisely the load-bearing gap. Section II-D asserts that three regions 'surpass the limitations' of two-region systems for boundary oscillations, but it never specifies how the region history handles an A→B→A or B→C→B transition. Since every real crossing passes through B, the counting result depends on an unstated state machine. The paper's own long-term data (Table II, 92% overall) also contradicts the abstract's 97% figure, and the live-test protocol that produced 97% is not described with enough detail to recompute it. The SOTA comparison in Table III likewise relies on the same unreported 97% number. Although the paper contains useful engineering components—fine-tuned SSD-MobileNet head detection, a lightweight feature-matching tracker, and edge deployment data—the central performance claim is not supported by the evidence presented. A simple scripted trial with boundary hesitations would settle the concern immediately; without such a test, the claimed accuracy cannot be accepted.","tokens_in":5206,"tokens_out":3743,"duration_ms":42298,"concrete_test":"Run a controlled overhead-camera trial with 20 scripted doorway behaviors: (1) straight A→C crossings; (2) A→B→A retreats followed by a real entry; (3) A→B→C→B→C oscillations; and (4) a person occluded for more than E frames between A and C. Compare the system's registered entry/exit counts with ground truth. If any retreat-only trial is counted, any oscillating trial is double counted, or any occluded crossing is dropped, then the Section II-D counting rule lacks the required hysteresis and the 97% accuracy claim does not hold for realistic partial or oscillating paths.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—97% accuracy in real-time people counting—depends on the counting algorithm in Section II-D, but that algorithm is underspecified at exactly the point where real doorway trajectories are nontrivial. The text defines an entry as a region-history move from A to C and an exit from C to A, with B as an intermediate 'critical' region, yet it gives no state machine, no hysteresis threshold, and no reset condition. It does not state whether A→B→A is discarded, whether A→B→C→B→C produces one count or two, or what happens when a person is first detected in B (e.g., someone already inside the critical zone at system startup). The motivation for three regions is to avoid oscillations near a single boundary, but no debounce or confirmation rule is provided. Algorithm 1 also leaves E (consecutive-miss deletion threshold) unspecified, so detection dropout near the boundary can split one crossing into two IDs or delete a crossing entirely. Additionally, the manuscript's own Table II reports 92% overall accuracy over two days (4 errors on 50 observations), not the abstract's 97%, and no event counts or error breakdown are given for the live test that allegedly achieved 97%. The counting rule's unstated assumptions and the unreported test statistics jointly make the headline accuracy impossible to verify or reproduce from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a real-time people-counting system based on an overhead camera. The pipeline consists of a fine-tuned SSD MobileNet head detector (day and night models), a MobileNetV2-based feature extractor, a custom feature-matching tracker (Algorithm 1), and a three-region counting rule (Section II-D). The authors report a 97% counting accuracy and 20–27 FPS on an Intel NUC edge computer, and compare favorably with prior works in Table III.","tokens_in":5537,"tokens_out":2549,"duration_ms":26787,"significance":"If the claimed 97% accuracy at 20–27 FPS on a low-power edge device were substantiated, the system would be a practically useful contribution to smart-building and safety applications, especially because the method aims to handle crowded scenes and low-light conditions. The paper, however, provides no reproducible experimental protocol, no dataset release, and no code release, and its own Table II reports 92% accuracy in the only described controlled test. The counting algorithm is underspecified at the exact point where real door-crossing trajectories are nontrivial. These deficiencies make the central performance claim unverifiable from the manuscript as written.","major_comments":[{"comment":"The counting rule is underspecified. The text defines an entry as movement from Region A to Region C and an exit from C to A, with B as a 'critical' intermediate region, but it never defines how partial or oscillating paths are handled. In particular, the manuscript does not state whether A→B→A is ignored, whether A→B→C→B→C yields one count or two, how an object first detected in B is treated, or what debounce/hysteresis rule prevents repeated counting at the boundary. Since the counting accuracy is the central claim of the paper, this missing state machine makes the reported 97% accuracy irreproducible.","section":"Section II-D"},{"comment":"The abstract and Section III claim an overall accuracy of 97% from 'live video testing,' yet the only controlled test described in detail, the two-day continuous run in Table II, reports an overall accuracy of 92.00% (4 errors on 50 actual observations, with predicted totals 26/26 versus actual 26/24). No event counts, duration, ground-truth protocol, or error breakdown are given for the live test that allegedly achieved 97%. The discrepancy between 97% and 92% is unexplained and must be resolved with a full description of both test protocols.","section":"Section III, Table II"},{"comment":"The tracking algorithm depends on several free parameters whose values and selection procedure are never reported: the feature-distance threshold T, the spatial-distance threshold D, and the consecutive-miss deletion threshold E. In addition, the placement of the two horizontal boundary lines that define Regions A, B, and C is not specified, nor is the feature crop size beyond the mention of (120,120,3). Without these parameter values, the experimental results cannot be reproduced or compared with alternative settings.","section":"Section II-C, Algorithm 1"},{"comment":"The fine-tuning details for the two SSD MobileNet head detectors are absent. The manuscript says the models are 'self-trained' and fine-tuned to distinguish heads from chairs, trolleys, and bags, but it does not report the size or composition of the training dataset, the annotation procedure, the number of training iterations, or any detection-level accuracy on a held-out test set. The examples in Figure 1 are illustrative but not a quantitative evaluation. This omission weakens the generalization claim for 'any complex environment.'","section":"Section II-A"},{"comment":"The state-of-the-art comparison in Table III is not reliable. The baseline 'Chinthaka et al. [9]' shares two co-authors with the present paper (B. K. K. Ng and C. Yuen), and the accuracy/FPS values for the baselines are cited from prior publications with, presumably, different hardware, camera setups, and evaluation protocols. The claimed 2% improvement over the best baseline is therefore not an independent, matched comparison. The table should state the source and conditions of each baseline result and discuss whether the comparisons are apples-to-apples.","section":"Table III"}],"minor_comments":[{"comment":"There is a typo: 'we introduce the the proposed object detection model' should read 'we introduce the proposed object detection model.'","section":"Section II, first paragraph"},{"comment":"Equation (2) defines accuracy as (Total observations - Error) / Total observations, but the term 'Error' is never defined precisely. Clarify whether an error is a wrong in/out count, a missed count, or any discrepancy per event.","section":"Section II-D"},{"comment":"The FPS results in Table I are averages without standard deviations or the number of runs; adding this information would make the real-time claim more credible.","section":"Section III, Table I"},{"comment":"The text references a real-time deployment result via link [12], but the figure caption does not explain the on-screen overlay ('# of Ins') or the ID labels A and C. A brief explanation in the caption would improve readability.","section":"Figure 3"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claims are not verifiable from the provided text: the 97% accuracy figure is contradicted by the paper's own Table II, the counting algorithm is underspecified, and the comparison with prior work is compromised by shared co-authors with one baseline. These are load-bearing issues that cannot be fixed by minor edits; they require new experiments and a full methodological description. I also note that the paper appears to be a preliminary workshop-style report rather than a complete journal submission, and it lacks a data/code availability statement that would be essential for the reported results to be reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. My take: this is a coherent, modest engineering paper that overstates its one headline number. The 97% accuracy in the abstract is contradicted by the paper's own Table II, which reports 92% overall across two days (4 errors on 50 observations). The live test that supposedly hit 97% has no described protocol—no event count, no error definition, no scene description—so the number is not reproducible from the manuscript.\n\nWhat the paper does well: it builds a sensible low-cost overhead-camera people counter using SSD-MobileNet head detection, MobileNetV2 embeddings, and a greedy feature-distance tracker with a miss-count threshold. The practical touches are real: fine-tuning the detector to reject chairs and bags, using the camera's IR mode to switch between day and night detection models, and downsampling inputs for edge inference. The FPS figures (20-27 on a NUC CPU) are believable for this pipeline. The two-day long-term monitoring is the most credible part of the evaluation, but it shows 92%, not 97%.\n\nThe soft spots are where the claims live. The counting logic in Section II-D is underspecified at exactly the tricky part. The rule says an entry is A-to-C and an exit C-to-A, with B as a 'critical' region, but there is no state machine for A->B->A, A->B->C->B->C, starting inside B, or detection dropout splitting a crossing into two IDs. Algorithm 1's thresholds T, D, and E are never given. The SOTA comparison is three entries, one of which is the same group's prior work, so the '2% improvement' is not independently established. No modern trackers (DeepSORT, ByteTrack) are compared, and no public benchmark is used.\n\nIf you weigh the central claim, the paper fails its own evidence bar. But the engineering is mostly sensible, and the two-day data gives a starting point. A serious referee could ask for the missing protocol, threshold values, and trajectory handling, and the paper might become useful after heavy revision. I would not cite it as is, but I would not mind seeing a revised version. For peer review: yes, send it out—not because it is ready, but because a competent reviewer can turn the vague accuracy claim into a concrete list of missing details. For a reading group, maybe, as an example of how system papers should report test protocols.","headline":"A sensible edge-counting system whose headline 97% accuracy is contradicted by its own 92% long-term test and is not reproducible from the text.","tokens_in":6014,"tokens_out":3061,"would_cite":false,"duration_ms":30149,"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":"A single overhead camera can count people entering and exiting a doorway in real time at 97% accuracy while running on a low-power edge computer.","keywords":["people counting","overhead camera","object tracking","edge computing","head detection","SSD MobileNet","feature matching","region history"],"falsifier":"Take a known ground-truth scenario where people step into the doorway's boundary zone and then step back out without fully entering, and compare the system's reported count with the true number of entries and exits; if the system registers changes that did not happen, the counting rule fails exactly in the lingering-and-reversing case.","tokens_in":5056,"feed_emoji":"🚶","tokens_out":6960,"duration_ms":64244,"temperature":0.7,"pith_summary":"This paper aims to show that a single overhead camera and a low-power edge computer can count people entering and exiting a doorway in real time without sacrificing accuracy for speed. The proposed pipeline combines a fine-tuned SSD MobileNet head detector, a MobileNetV2 feature extractor that produces a 1024-dimensional embedding per head, a custom feature-distance tracking algorithm, and a three-zone counting rule where movement from the outside zone to the inside zone registers an entry and the reverse registers an exit. The authors report 97% counting accuracy at 20-27 FPS on an Intel NUC mini PC with no GPU, which would beat the accuracy-FPS trade-off of prior door-monitoring systems. A two-day long-term test reports an overall accuracy of 92%, with per-day values of 89.66% and 95.24%, so the headline figure refers to separate live video testing.","feed_headline":"Overhead camera system counts people at 97% accuracy in real time","feed_subtitle":"Runs at 20-27 FPS on a GPU-less mini PC, beating the usual accuracy-versus-speed trade-off.","key_machinery":"The load-bearing component is the region-history attribute attached to each tracked head object, together with the assignment-based tracking algorithm. The camera's field of view is divided into three horizontal zones: A (outside), B (critical boundary), and C (inside); each person-object stores the sequence of zones it has occupied, and an entry is counted only when the sequence moves from A to C, an exit only from C to A. Tracking is framed as an assignment problem on an $m \\times n$ feature-distance matrix $M$, where the algorithm greedily pairs the closest registered and new detections subject to a distance threshold $T$ and a spatial cap $D$, and each unassigned old object increments a miss counter $eCount$ that removes the object after $E$ consecutive misses. The head detector is a fine-tuned SSD MobileNet with separate day and night models, and the feature extractor is a compressed MobileNetV2 that takes a $120\\times120\\times3$ crop and outputs a $1\\times1024$ embedding.","core_discovery":"The central claim is that the combination of a fine-tuned single-shot head detector, a lightweight embedding-based tracker, and a three-region history rule solves the accuracy-versus-speed problem in people counting. On the paper's terms, the method is state-of-the-art because it keeps accuracy at 97% while running at 20-27 FPS on edge hardware, where prior work either ran fast with lower accuracy (86.32% at 45 FPS) or achieved high accuracy at impractically low frame rates (93.10% at 0.6 FPS, 95% at 5 FPS). The counting rule is designed to avoid the oscillation errors of two-zone systems by requiring a full outside-to-inside transition before registering an entry, and the tracking algorithm tolerates short detection gaps through a miss counter. The long-term laboratory test shows the system maintains a 92% overall accuracy across two days of continuous operation.","pith_inferences":["A natural stress test is to have people linger in the boundary zone and reverse direction; the paper does not specify how the region history handles A-B-A paths, so this is the most likely place for the 97% claim to break down in practice.","The reported 97% live accuracy versus 92% long-term accuracy suggests the figure is scenario-dependent; reporting error breakdowns by entry vs exit and by crowd size would make the method's true operating envelope clear.","The same assignment-plus-region-history logic could be transferred to other top-down counting tasks, such as objects on conveyor belts or vehicles in lanes, by retraining the detector on the new object class.","The dual day/night model selection could be replaced by an on-the-fly brightness estimator, making the system camera-agnostic rather than tied to the IR-mode signal of one camera model."],"forward_implications":["At 20-27 FPS on a GPU-less mini PC, the system can be deployed for continuous door monitoring in office buildings, retail stores, and transit stations where power and cost are constrained.","The day/night dual-model detection means the same hardware can provide 24-hour counting without additional lighting, using the camera's IR mode to switch models.","Because tracking is feature-based rather than bounding-box-based, the method tolerates occlusions and brief detection misses as long as they stay under the E-frame threshold.","In long-term operation, the method reports 92% overall accuracy, with the two-day test showing errors can be as low as one miscount per day when traffic is light.","The accuracy-FPS combination implies that prior trade-offs, such as 93% at 0.6 FPS or 95% at 5 FPS, are not fundamental limits for overhead-camera counting."],"supporting_citations":[{"why":"Provides the SSD single-shot detector architecture used for head detection.","marker":"[10]"},{"why":"Provides the MobileNetV2 architecture used to extract 1024-dimensional head embeddings for tracking.","marker":"[11]"},{"why":"Baseline depth-sensor counting method (86.32% accuracy, 45 FPS) that the paper compares against.","marker":"[7]"},{"why":"Baseline CNN-based head-detection counting method (93.10% accuracy, 0.6 FPS) that the paper claims to outperform in speed.","marker":"[8]"},{"why":"Baseline cluster-pruning edge-AI counting method (95% accuracy, 5 FPS) that is the closest accuracy competitor.","marker":"[9]"}],"fun_headline_variants":["97% accuracy people counting now runs on low-power edge cameras","Real-time crowd counting hits 97% accuracy on a GPU-less mini PC","Overhead camera AI tracks people at 27 FPS with 97% accuracy","97% accurate crowd counting in real time on edge hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 97% figure rests on the assumptions that people cross the doorway in one direction without reversing in the boundary zone and that the detector never drops a person long enough to hit the miss threshold E.","fun_headline_variants_meta":{"raw":{"variants":["97% accuracy people counting now runs on low-power edge cameras","Real-time crowd counting hits 97% accuracy on a GPU-less mini PC","Overhead camera AI tracks people at 27 FPS with 97% accuracy","97% accurate crowd counting in real time on edge hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":2961,"prompt_tokens":830,"completion_tokens":2131,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":2066}},"tokens_in":446,"tokens_out":2131,"duration_ms":15366,"temperature":1.0,"reasoning_tokens":2066,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:00:00.331068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a known ground-truth scenario where people step into the doorway's boundary zone and then step back out without fully entering, and compare the system's reported count with the true number of entries and exits; if the system registers changes that did not happen, the counting rule fails exactly in the lingering-and-reversing case.","supporting_citations":[{"cited_title":"Ssd: Single shot multibox detector,","cited_arxiv_id":null,"evidence_quote":"Provides the SSD single-shot detector architecture used for head detection."},{"cited_title":"Benchmark data and method for real-time people counting in cluttered scenes using depth sensors,","cited_arxiv_id":null,"evidence_quote":"Baseline depth-sensor counting method (86.32% accuracy, 45 FPS) that the paper compares against."},{"cited_title":"Passenger flow estimation based on convolutional neural network in public transportation system,","cited_arxiv_id":null,"evidence_quote":"Baseline CNN-based head-detection counting method (93.10% accuracy, 0.6 FPS) that the paper claims to outperform in speed."},{"cited_title":"Cluster pruning: An efficient filter pruning method for edge ai vision applica- tions,","cited_arxiv_id":null,"evidence_quote":"Baseline cluster-pruning edge-AI counting method (95% accuracy, 5 FPS) that is the closest accuracy competitor."}],"review_version":1}