{"id":"ec41351b-fa50-419a-89f5-53ef0489b1ae","arxiv_id":"2505.07110","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper re-describes DeepSORT and reports a small, undocumented comparison table claiming it beats Fast-RCNN, Mask-RCNN, and YOLOv5 on gesture and eye tracking.","lead":"This paper applies the existing DeepSORT tracking algorithm to hand gesture recognition for human-computer interaction. It claims DeepSORT outperforms detector baselines on a gesture and eye tracking task, but provides no code, data, or detailed experimental protocol.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 uses detection metrics (mAP, Precision, Recall, F1) to compare a tracker against detectors, so it cannot support the central claim of superior tracking accuracy and real-time performance.","rationale":"The reader's concern centers on unverifiability of Table 1. I sharpen this: the table is not merely undocumented; it measures the wrong quantity. mAP, Precision, Recall, and F1 are detection metrics, while DeepSORT is a tracker and the claim is about tracking accuracy and real-time performance. A detector can have high per-frame detection scores while failing in identity preservation, occlusion handling, and trajectory continuity, which are the core of the paper's claim. The paper also fails to report any tracking-specific metric or runtime. Thus the central claim is unsupported even under the most charitable reading of the numbers. The dataset mischaracterization is an additional independence concern. Because these issues confirm the reader's rejection without changing the verdict, I recommend no adjustment.","tokens_in":6603,"tokens_out":3086,"duration_ms":30556,"concrete_test":"Run a proper comparison on the same UT-Interaction sequences: compute MOTA, IDF1, HOTA, identity switches, and FPS for DeepSORT and for each detector coupled with a standard association step (e.g., SORT), using a common evaluation tool such as TrackEval. If DeepSORT does not outperform these baselines on tracking metrics, the superiority claim is unsupported. As a minimal check, inspect whether Table 1 contains any tracking metric at all; if none are present, the experiment as reported cannot support the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DeepSORT can accurately capture and track the user's gesture trajectory and is superior to traditional tracking methods in terms of real-time and accuracy. The only quantitative evidence is Table 1 in Section IV.B, but all four reported metrics (mAP, Precision, Recall, F1-score) are per-frame detection/classification metrics, not tracking metrics. DeepSORT is an online multi-object tracker; its accuracy is normally measured with MOT metrics such as MOTA, IDF1, HOTA, and identity switches. The comparators (Fast-RCNN, Mask-RCNN, YOLOv5) are detectors, not trackers, so Table 1 does not compare tracking methods. No FPS, latency, or other real-time measurement is reported despite the real-time claim. The gesture recognition results in Figures 2 and 3 are qualitative only, with no recognition accuracy or confusion matrix. A secondary issue: the paper describes UT-Interaction as an HCI dataset with eye movements and gestures, but UT-Interaction is a human-human action recognition dataset; this mismatch further weakens external validity. Therefore, even if the numbers in Table 1 are reproducible, they do not verify the strongest claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a DeepSORT-based visual tracking approach for gesture recognition in human-computer interaction. It reviews the DeepSORT pipeline, presents standard Kalman-filter and Hungarian-association equations, and reports a comparative experiment in Table 1 where DeepSORT is claimed to achieve the best mAP, precision, recall, and F1-score against Fast-RCNN, Mask-RCNN, and YOLOv5. Qualitative figures illustrate eye-movement tracking and gesture trajectories. The paper concludes that DeepSORT delivers superior tracking accuracy and real-time performance in dynamic, multi-target HCI scenarios.","tokens_in":6819,"tokens_out":2491,"duration_ms":25592,"significance":"If properly validated, a real-time, robust gesture tracker is a useful component for HCI systems, and the paper's topic is of general interest. However, the manuscript offers no valid experimental support for its central claim. Table 1 uses detection metrics to compare a tracker against detectors, no evaluation protocol or implementation details are given, the dataset is described inaccurately, and the only other evidence is qualitative. The contribution is therefore not established; the significance remains conditional on a completely new, properly designed tracking evaluation.","major_comments":[{"comment":"The central claim of superiority in tracking accuracy rests entirely on Table 1, but the four reported metrics (mAP, precision, recall, F1-score) are per-frame detection or classification metrics, not multi-object tracking metrics such as MOTA, IDF1, HOTA, or identity-switch counts. DeepSORT is a tracker, while the comparators (Fast-RCNN, Mask-RCNN, YOLOv5) are detectors, so the table does not compare tracking methods at all. The abstract's statement that DeepSORT is 'superior to traditional tracking methods in terms of real-time and accuracy' is therefore unsupported by the presented quantitative evidence.","section":"IV.B, Table 1"},{"comment":"No experimental protocol is described: there is no training/test split, no mention of how detections were produced for DeepSORT, no hyperparameters, no hardware details, and no runtime or latency measurements. The text claims real-time performance but never reports FPS, processing time, or any other timing metric. The absence of error bars or variance, combined with differences as small as 0.02 in mAP, makes the reported comparisons unverifiable and plausibly within noise. This lack of protocol prevents reproducibility and invalidates the comparative conclusions.","section":"IV.B, experimental protocol"},{"comment":"The description of the UT-Interaction dataset is factually incorrect. UT-Interaction is a standard human-human interaction and action recognition dataset containing activities such as handshaking, pointing, and hugging, recorded from a surveillance-style viewpoint. It is not an HCI dataset with eye movements, mouse operations, or touchscreen gestures as claimed in the paper. This mismatch undermines the external validity of the experiments and raises a serious concern about whether the reported results were actually obtained on the stated data.","section":"IV.A, dataset description"},{"comment":"Figures 2 and 3 provide only qualitative screenshots of eye-movement and gesture trajectories. No recognition accuracy, confusion matrix, trajectory error metric, or comparison against baseline tracking methods is reported for these qualitative results. Consequently, the paper does not demonstrate that the proposed system can actually recognize the gestures (swipe, click, zoom) or that its tracking trajectories are accurate, even though these are the claimed outcomes.","section":"IV.B, Figures 2 and 3"}],"minor_comments":[{"comment":"There is a typographical error: the section begins with 'In In this study,' which should be corrected to 'In this study.'","section":"III, first paragraph"},{"comment":"The text states that 'DeepSORT achieves the highest performance across all metrics,' but in Table 1 YOLOv5 has a higher recall (0.85) than DeepSORT (0.84). This factual inconsistency should be corrected.","section":"IV.B, Table 1 discussion"},{"comment":"Several equations contain garbled or incomplete notation, such as the state vector and the subscript mismatches in the matching-cost formulation; the mathematics would need to be typeset cleanly to be checked. The references to OC-SORT, CMC, and AW are mentioned as components of the framework but are never defined or evaluated, so their inclusion is unclear.","section":"III, equations and notation"},{"comment":"The paragraph describing the dataset repeats claims about 'eye movements' and 'operating a mouse, touchscreen' that do not correspond to any known public dataset; please clarify the actual data source and annotation format.","section":"IV.A, dataset description"}],"recommendation":"reject","confidential_remarks":"The manuscript's reference list contains many self-citations to journals and arXiv preprints that appear unrelated to the claimed contributions, and the dataset description does not match the publicly known UT-Interaction dataset. These issues, together with the absence of any usable experimental protocol, suggest that the current manuscript is not suitable for publication in its present form. A resubmission would require a fundamentally new experimental evaluation with proper tracking metrics, a correct dataset, and reproducibility details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper applies DeepSORT to gesture recognition, and nothing more. What it does well is accurately summarize how DeepSORT works—Kalman prediction, appearance features, Hungarian association. That is a textbook description, but it is correct, so I'll credit it. That is also the only thing worth crediting.\n\nThe central claim is that DeepSORT “outperforms traditional tracking methods in real-time and accuracy.” The only quantitative support is Table 1, and the stress-test note is exactly right: the four metrics there are mAP, Precision, Recall, and F1-score, which are detection/classification metrics. DeepSORT is a tracker; Fast-RCNN, Mask-RCNN, and YOLOv5 are detectors, not trackers. No MOTA, IDF1, HOTA, identity switches, or FPS/latency is reported. So the table does not test the claim being made. Even if the numbers are faithfully reproduced, they don't measure tracking quality or real-time performance.\n\nThe experiments are otherwise qualitative: a few trajectory plots with no recognition accuracy, no confusion matrix, no protocol, no code, no error bars. The UT-Interaction dataset is described as an HCI dataset with eye movements and gestures, but it is widely known as a human-human action recognition dataset—the description is factually wrong, which further undercuts external validity.\n\nThere are also signs of citation padding: references to LLM compression, data center scheduling, and graph-based anomaly detection have no bearing on gesture tracking. They inflate the bibliography without supporting the work. The manuscript also contains typos like “In In this study” and garbled math, but that's minor next to the missing evidence.\n\nIs this a serious paper? No. The method is routine, the evaluation is misaligned with the claim, and there's no reproducibility. A desk reject is the right call, not peer review. A serious referee would spend an afternoon writing what I just wrote in five minutes.\n\nWho might still get value? Someone looking for a compact recitation of DeepSORT's equations. That's not a research contribution.","headline":"A routine DeepSORT application with no verifiable experiments: the only quantitative evidence compares detection metrics against detectors, not tracking metrics, so the central claim collapses.","tokens_in":7271,"tokens_out":1405,"would_cite":false,"duration_ms":15308,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DeepSORT, a hybrid Kalman-filter and deep-feature tracker, is claimed to capture hand and eye gesture trajectories in real time and to beat Fast-RCNN, Mask-RCNN, and YOLOv5 in accuracy.","keywords":["DeepSORT","visual tracking","gesture recognition","human-computer interaction","multi-object tracking","Kalman filter","data association","real-time interaction"],"falsifier":"Re-run the four models on the UT-Interaction dataset with identical detections, a fixed evaluation script, and repeated runs, and check whether DeepSORT actually reaches mAP 0.79, precision 0.89, recall 0.84, and F1 0.87 while beating YOLOv5's 0.77, 0.88, 0.85, and 0.86; if the reported ordering or margins do not reproduce, the central claim fails.","tokens_in":6459,"feed_emoji":"🖐️","tokens_out":8524,"duration_ms":73834,"temperature":0.7,"pith_summary":"This paper sets out to establish that DeepSORT, a multi-object tracker combining Kalman-filter motion prediction with deep-learning appearance features and Hungarian-assignment data association, can serve as the visual-tracking core of gesture-based human-computer interaction. It claims that this tracker captures hand and eye trajectories in real time, keeps identities stable through occlusion and fast motion, and handles several targets at once. The supporting comparison in Table 1 reports DeepSORT with mAP 0.79, precision 0.89, recall 0.84, and F1 0.87, ahead of YOLOv5, Fast-RCNN, and Mask-RCNN. If correct, the result matters because vision-based interaction could replace mice, keyboards, and touchscreens in dynamic settings such as augmented reality, virtual reality, and robotics.","feed_headline":"DeepSORT beats three rivals in gesture tracking","feed_subtitle":"Kalman-filtered deep tracking claims top marks for precision and real-time hand and eye control.","key_machinery":"The load-bearing object is the DeepSORT algorithm itself, defined here as a multi-object tracker that pairs a Kalman filter for motion prediction with a convolutional-neural-network appearance descriptor and the Hungarian assignment procedure, which matches detections to existing tracks. In the paper's pipeline, each frame yields detections with bounding boxes and feature vectors; the Kalman filter predicts where each target will be next, a cost matrix combines motion and appearance distances, and the Hungarian assignment decides which detection continues which track. The paper also adds supporting modules around this core: a REID module for identity maintenance, OC-SORT for refined motion modeling, CMC for cross-modal consistency, and AW for attention weighting. What this machinery does for the argument is turn per-frame detections into persistent trajectories, which is exactly what gesture recognition needs to distinguish a swipe from a click and to keep a hand identity stable across occlusions.","core_discovery":"The central claim, stated on the paper's own terms, is that a DeepSORT-based pipeline is an accurate and real-time method for tracking the user's gestures in interactive systems. The system first detects targets in each frame as bounding boxes, extracts appearance features with a deep convolutional network, predicts the next state with a Kalman filter, and then matches detections to existing tracks by solving a cost matrix with the Hungarian algorithm; this cascade turns raw frames into continuous, labeled hand or eye trajectories. The paper reports that DeepSORT reaches the best scores in its four-model comparison and that qualitative plots for swipe, click, zoom, and eye-gaze show stable, well-attributed trajectories. It attributes the advantage to the hybrid of motion prediction and appearance re-identification, which the paper says resists occlusion, motion blur, and identity switches in multi-target scenes.","pith_inferences":["Because the paper reports no training details, evaluation protocol, or error bars for Table 1, the numerical ordering should be treated as preliminary until the experiment is reproduced under a disclosed protocol.","A natural extension the paper leaves implicit is using the same tracking backbone for AR/VR pointer prediction or assistive interfaces, where trajectory continuity rather than per-frame detection accuracy determines usability.","The claimed advantage depends on appearance re-identification, so a sharper test would use a benchmark with frequent identity switches and visually similar targets; failure there would localize the limit of the approach."],"forward_implications":["Gesture-driven interfaces can be built on DeepSORT as the tracking layer, with hand trajectories for swipe, click, and zoom extracted and visualized in real time.","Because the method is designed to hold identities through occlusion and motion blur, interactive systems can keep tracking a hand even when it briefly disappears behind an object or moves quickly.","The multi-target capability means several hands, users, or body parts can be tracked simultaneously, supporting shared-screen and multi-user interaction scenarios.","Following the paper's comparison, DeepSORT would be the recommended choice over Fast-RCNN, Mask-RCNN, and YOLOv5 for HCI tracking tasks among the four tested models.","Eye tracking and gesture tracking can share one DeepSORT-based pipeline, since both are presented as instances of the same tracking mechanism."],"supporting_citations":[{"why":"Cited as evidence that DeepSORT offers superior accuracy and robustness in complex multi-target tracking.","marker":"[5-7]"},{"why":"Cited for the claim that basic SORT struggles with occlusions and identity switches, which motivates the deep extension.","marker":"[14]"},{"why":"Cited for the claim that CNN-based appearance features enable robust re-identification and improved tracking accuracy.","marker":"[15]"},{"why":"Supplies the adaptive signal-processing and responsiveness design principles the method says it builds on.","marker":"[17]"},{"why":"Supplies the re-identification module that maintains target identities across frames.","marker":"[18]"},{"why":"Supplies the OC-SORT motion modeling used to refine state prediction within the tracking pipeline.","marker":"[19]"},{"why":"Supports the gesture-recognition context in which the tracking performance is evaluated.","marker":"[20]"},{"why":"Guides the cross-modal feature integration and attention-weighting modules added to the framework.","marker":"[21]"}],"fun_headline_variants":["DeepSORT tops gesture tracking in four-model test","Real-time gesture tracking wins with DeepSORT's hybrid approach","DeepSORT beats rivals on precision and real-time hand tracking","DeepSORT handles occlusion, blur in multi-target gesture tracking","Kalman-deep fusion boosts gesture tracking in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Table 1 comes from a real, methodologically sound experiment in which the four models were evaluated on the UT-Interaction dataset under comparable conditions; the paper gives no training or evaluation details, so if that table is not a fair side-by-side tracking comparison, the superiority claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["DeepSORT tops gesture tracking in four-model test","Real-time gesture tracking wins with DeepSORT's hybrid approach","DeepSORT beats rivals on precision and real-time hand tracking","DeepSORT handles occlusion, blur in multi-target gesture tracking","Kalman-deep fusion boosts gesture tracking in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":3010,"prompt_tokens":975,"completion_tokens":2035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1952}},"tokens_in":591,"tokens_out":2035,"duration_ms":12887,"temperature":1.0,"reasoning_tokens":1952,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:23:29.537143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the four models on the UT-Interaction dataset with identical detections, a fixed evaluation script, and repeated runs, and check whether DeepSORT actually reaches mAP 0.79, precision 0.89, recall 0.84, and F1 0.87 while beating YOLOv5's 0.77, 0.88, 0.85, and 0.86; if the reported ordering or margins do not reproduce, the central claim fails.","supporting_citations":[{"cited_title":"Research Hotspots and Trends of User- Centered Human-Computer Interaction: A Bibliometric Analysis,","cited_arxiv_id":null,"evidence_quote":"Cited for the claim that basic SORT struggles with occlusions and identity switches, which motivates the deep extension."},{"cited_title":"Making Cognitive Ergonomics in the Human– Computer Interaction of Manufacturing Execution Systems Assessable: Experimental and Validation Approaches to Closing Research Gaps,","cited_arxiv_id":null,"evidence_quote":"Cited for the claim that CNN-based appearance features enable robust re-identification and improved tracking accuracy."},{"cited_title":"A Visual Communication Optimization Method for Human- Computer Interaction Interfaces Using Fuzzy Logic and Wavelet Transform,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive signal-processing and responsiveness design principles the method says it builds on."},{"cited_title":"Deep-ReID: Deep features and autoencoder assisted image patching strategy for person re- identification in smart cities surveillance,","cited_arxiv_id":null,"evidence_quote":"Supplies the re-identification module that maintains target identities across frames."},{"cited_title":"AppleYOLO: Apple yield estimation method using improved YOLOv8 based on Deep OC-SORT,","cited_arxiv_id":null,"evidence_quote":"Supplies the OC-SORT motion modeling used to refine state prediction within the tracking pipeline."},{"cited_title":"Deep Learning Approach for Hand Gesture Recognition: Applications in Deaf Communication and Healthcare,","cited_arxiv_id":null,"evidence_quote":"Supports the gesture-recognition context in which the tracking performance is evaluated."},{"cited_title":"Human-Computer Interaction in Smart Devices: Leveraging Sentiment Analysis and Knowledge Graphs for Personalized User Experiences,","cited_arxiv_id":null,"evidence_quote":"Guides the cross-modal feature integration and attention-weighting modules added to the framework."}],"review_version":1}