{"id":"23cc148f-9d52-4418-aaff-4999ac1d20c2","arxiv_id":"2504.14807","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A real-time eye-state classifier for driver drowsiness based on dynamic template matching, Kalman tracking, and HOG+SVM, presented without quantitative evaluation.","lead":"This paper describes a driver drowsiness detection system that tracks the driver's eyes with template matching and a Kalman filter, then uses a support vector machine to classify each eye as open or closed. It is a small engineering paper that reports no quantitative accuracy, speed, or comparison numbers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative evaluation supports the claimed real-time accuracy: Section 3 reports only illustrative frames and unlabeled ROC curves, with no accuracy, AUC, latency, or eyeglass-stratified results.","rationale":"The reader's verdict is REJECT, and I agree with that verdict. However, the identified weakest assumption—training distribution mismatch—is not the single most load-bearing concern. The paper does not report quantitative results for the actual in-vehicle system at all: Fig. 5 shows ROC curves without axis labels or threshold values, Fig. 6 shows illustrative frames with a subjective caption, and no frame rate or processing time is given anywhere. Even if the training data perfectly matched deployment conditions, the central claim of real-time accurate drowsiness detection would still be unverified because its two main predicates, accuracy and speed, are never measured. The training/test mismatch is a plausible cause of failure, but it presumes there is a reported performance to transfer, which there is not. My concrete test asks for a held-out in-vehicle evaluation with ground truth, which would expose both the transfer problem and the missing speed measurement. Since the reader already rejected the paper on the grounds of unsupported claims, my concern does not move the verdict; it strengthens the same rejection with a more direct and more fundamental issue.","tokens_in":5570,"tokens_out":2270,"duration_ms":25235,"concrete_test":"Obtain the authors' trained SVM and tracking pipeline, or reimplement it from Section 2, and run it on a held-out in-vehicle video set with per-frame ground-truth eye states (open/closed), recorded under varying head pose and with subjects both with and without eyeglasses. Report per-frame classification accuracy and ROC AUC, the number of tracking losses per 1000 frames, and the wall-clock frames-per-second on the target embedded platform. If accuracy is not materially higher than chance for eyeglass-wearing or profile-view segments, or if frame rate is below the real-time threshold, the central claim is falsified. If the authors instead provide such a table as an erratum, that would settle the concern by supplying the missing evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the system detects driver eye state accurately in real time, including with eyeglasses and off-angle profile views, and that it has minimal computational complexity. This claim has two testable components: classification/tracking accuracy in the target driving environment, and processing speed. Section 3 reports a 2m59s, 5000-frame tracking video and ROC curves for LBP versus HOG features, but gives no numeric results: no area under the ROC curve, no chosen operating threshold, no false-positive rate, no per-frame eye-state classification accuracy, no tracking-loss count, no frame rate or per-frame processing time, and no comparison against the cited prior systems. The caption of Fig. 6, 'All detections are correct,' is anecdotal. Without these measurements, the claimed real-time performance and robustness to eyeglasses and off-angle views are not supported. The reader's weakest assumption, that the RIP ISL plus personal training set may not represent in-vehicle deployment conditions, is real, but it is secondary: even a perfect train/test match would not rescue the paper, because the paper never reports a deployment-oriented evaluation at all. A held-out in-vehicle measurement with ground truth would simultaneously test the classification accuracy, the tracking robustness, and the transfer assumption. The absence of this measurement is the most load-bearing weakness because every downstream claim—drowsiness alarm reliability, embedded suitability, superiority to prior work—rests on it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a real-time driver drowsiness detection pipeline. The face is detected with an LBP cascade, eyes are localized with a Haar cascade, and eye positions are tracked in subsequent frames by normalized cross-correlation template matching combined with Kalman filtering. A linear SVM with HOG features classifies each eye ROI as open or closed, and a drowsiness alarm is triggered if the eyes remain closed for a specified duration. The authors report qualitative results on in-vehicle and lab videos and compare LBP versus HOG features via ROC curves (Fig. 5), concluding that the system operates in real time with minimal computational complexity and is robust to eyeglasses and profile views.","tokens_in":5850,"tokens_out":1504,"duration_ms":14688,"significance":"If substantiated, the claimed contribution—a low-complexity eye-tracking and eye-state classification pipeline robust to partial occlusion and eyeglasses—would be of practical interest for embedded driver monitoring. The paper's strength is its simple, modular design that reuses well-established components (Viola-Jones detection, Kalman filtering, SVM classification), and the choice of HOG over LBP is a sensible, testable design decision. However, the empirical core of the paper is not reported: there are no numeric accuracy, ROC-AUC, latency, or tracking-failure measurements, and no reproducible evaluation protocol. The central claims of real-time operation, accuracy under eyeglasses, and robustness to off-angle views therefore remain unsupported. The manuscript does not ship code, datasets, or machine-checked derivations; the only quantitative statements are the training set size (9763 eye images) and the length of a sample tracking video (5000 frames).","major_comments":[{"comment":"The ROC curves for LBP and HOG are presented without any numeric values: no AUC, no operating point, no false-positive rate, and no accuracy or error rate. The claim that \"HOG feature outperformed LBP feature\" (Section 3) cannot be assessed, and no statistically meaningful comparison is possible. Please report the full confusion matrix, AUC with confidence intervals, and a chosen operating threshold with the corresponding true-positive and false-positive rates.","section":"Section 3, Fig. 5"},{"comment":"The caption states \"All detections are correct,\" but this is anecdotal; no ground-truth labeling procedure, no frame count, no per-class accuracy, and no stratified results (e.g., with/without eyeglasses, near-frontal/profile, lab/vehicle) are given. The claim that the system \"will be detected accurately even if the driver is wearing eyeglass\" (Section 1) is a central contribution and must be supported by quantitative results on an eyeglass-annotated test set.","section":"Section 3, Fig. 6"},{"comment":"The paper describes a 2m59s, 5000-frame tracking video but reports no tracking success rate, no number of lost tracks, no re-initialization frequency, and no localization error relative to ground truth. Since the tracking step feeds the eye-state classifier, the absence of any tracking accuracy measure makes it impossible to evaluate the end-to-end pipeline. Please report per-frame eye-center error or tracking success/failure counts, and specify the ground-truth annotation method.","section":"Section 3, tracking video"},{"comment":"The paper claims the system \"works in real time with minimal computational complexity\" (Section 4 and Introduction) but reports no frame rate, per-frame processing time, hardware platform, or resolution. This is a load-bearing claim for a driver monitoring system and must be quantified; please provide measured throughput on the target embedded platform and a per-stage timing breakdown (face detection, eye detection/tracking, feature extraction, SVM classification).","section":"Section 3, real-time claim"},{"comment":"The SVM is trained on the RIP ISL Eye Dataset plus a personal collection (9763 images), but the paper never characterizes the test distribution (in-vehicle lighting, camera angle, eyeglasses, partial occlusion) or establishes that the training data cover it. No cross-validation, subject independence, or domain-adaptation procedure is described. Because the paper makes deployment claims, the evaluation must use a held-out set that is stratified by the challenging conditions named in the Introduction; otherwise the claimed accuracy may not transfer to actual driving environments.","section":"Section 3, training/test distribution"}],"minor_comments":[{"comment":"The correlation formula contains notation that is not fully defined; in particular, the summation ranges and the relationship between T', I', and the image coordinates are unclear. Please rewrite the equation with explicit definitions of all variables and bounds.","section":"Section 2.2, Eq. (1)-(2)"},{"comment":"\"The final result of eye detection and tracking is subjected to Kamlan tracking\" appears to be a typo for \"Kalman tracking\". Please correct.","section":"Section 2.2, text"},{"comment":"The description of template update says templates are replaced by the most recent eye templates obtained via Viola-Jones detection, but the trigger condition for re-detection is only \"matching score is less than specified threshold\"; the threshold value is never given. Please state the threshold and how it was chosen.","section":"Section 2.2, text"},{"comment":"The sample tracking frames are shown without any quantitative measure of tracking accuracy or a description of how the bounding boxes were verified. Please add a caption explaining the video content, frame indices, and what the boxes represent.","section":"Section 3, Fig. 4"},{"comment":"Reference [17] misspells \"Gobb's\" (should be \"Dr. Dobb's Journal\"), and reference [18] incorrectly lists the first author as \"Ojha\" (the correct spelling is \"Ojala\"). Please check all references against the original sources.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper is a short workshop-style report with no quantitative evaluation. The central claims are entirely unsupported by numbers, and the missing measurements (AUC, latency, tracking error, eyeglass-stratified accuracy) are not local fixes that can be patched with a few added sentences; they require a new evaluation study. For a journal submission, the absence of any reproducible experiments is a fundamental lack of evidence. The novelty over the cited prior art (e.g., template matching plus Kalman filtering and SVM eye-state classification) is also modest, although the specific combination and the single-eye decision rule are reasonable engineering contributions if properly validated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2504.14807. The paper is a straightforward integration of known components: Viola-Jones face detection, Haar eye detection, dynamic template matching with multiple templates and Kalman filtering, HOG features with an SVM for open/closed classification, and a temporal drowsiness rule. The authors are honest about the components and cite the relevant prior work. The system description is clear enough that someone could reimplement it, aside from the garbled correlation equations (1) and (2), which are not reproducible as printed.\n\nWhat the paper does well: it makes a sensible engineering choice to track each eye independently and fuse with face geometry knowledge, and it reports training data counts (9763 eye images). The LBP vs HOG comparison is a legitimate question, and the ROC curves in Fig. 5 are at least directionally informative.\n\nThe soft spot is not small: there are no numbers anywhere. No AUC, no accuracy on the 5000-frame test video, no false positive rate, no tracking loss count, no frame rate or per-frame processing time, no baseline comparison against any of the cited systems. The claim that the system works in real time with minimal computational complexity is asserted in the conclusions, not measured. \"All detections are correct\" in Fig. 6 is a caption, not an evaluation. The training set is heterogeneous, and the test environment (camera offset to the left, possible profile views, eyeglasses) is described but never paired with ground truth or a held-out evaluation.\n\nThe citation pattern looks fine; the single self-citation [16] is for color face detection and is not load-bearing.\n\nMy verdict matches the stress-test: the missing evaluation is load-bearing. Even if the train/test distribution mismatch were resolved, the paper still would not support its central claim because it never reports a deployment-oriented measurement. This is an incomplete paper as submitted, not a fundamentally unsound one.\n\nWho is this for? Someone working on embedded driver monitoring might find the system architecture interesting as a starting point, but they cannot compare it to alternatives. It deserves peer review only if the venue expects to ask for substantial additional experiments; as a short paper it would need heavy revision. I would not cite it in its current form.\n\nRecommendation: if I were the editor, I would not desk reject it, but I would send it to a rigorous reviewer who will explicitly demand quantitative results for the 5000-frame sequence, a frame-rate measurement, and a comparison to at least one prior system. Without those, it cannot be accepted.","headline":"A competent engineering write-up of a standard eye-tracking pipeline that omits every quantitative result needed to support its real-time accuracy claim.","tokens_in":6389,"tokens_out":2313,"would_cite":false,"duration_ms":20484,"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 driver-monitoring system can detect drowsiness in real time by tracking eyes with dynamic template matching and classifying each eye as open or closed with a support vector machine.","keywords":["driver drowsiness detection","eye tracking","dynamic template matching","Kalman filter","support vector machine","histogram of oriented gradients","face detection","driver monitoring system"],"falsifier":"Run the system in a car with the camera mounted off to the driver's left, as the paper describes, over varied lighting and with and without eyeglasses, and record eye-tracking success and open/closed classification accuracy on a labeled frame set. If accuracy drops materially below the reported HOG ROC curve, or if the tracker re-initialization rate rises sharply during ordinary head motion, the claimed real-world accuracy does not hold.","tokens_in":5374,"feed_emoji":"😴","tokens_out":8962,"duration_ms":75575,"temperature":0.7,"pith_summary":"The paper claims that a driver-drowsiness system can be built from a small set of classical computer-vision components: face detection, normalized cross-correlation template matching for eye tracking, a Kalman filter for smoothing, and a linear support vector machine on HOG features to label each eye as open or closed. The system is designed to run in real time on modest hardware because only the eye regions are processed, and it is meant to keep working when the driver's face is at an angle or only one eye is visible. If the eye state is reported closed for a fixed time interval, an alarm is generated, which is the mechanism for waking a drowsy driver. The paper asserts this design detects eye state accurately even when the driver wears eyeglasses, and that HOG features outperform LBP features for this classification. If correct, a low-cost camera plus this pipeline could alert drivers before they fall asleep, addressing a large share of fatigue-related crashes.","feed_headline":"Eye tracker detects closed eyes in real time and alerts drowsy drivers","feed_subtitle":"The pipeline tracks eyes with template matching, classifies them with SVM, and alarms after prolonged closure.","key_machinery":"The load-bearing mechanism is the pairing of dynamic template matching with an SVM classifier. When tracking starts, several eye templates, typically 10 to 20, are stored; on each new frame, normalized cross-correlation measures how well each template matches the search region, the best match locations are averaged, and a Kalman filter smooths the result to produce the tracked eye position. If the correlation falls below a threshold, the system re-detects the face and eyes and refreshes its templates. The tracked eye region is preprocessed with gamma correction, difference-of-Gaussian filtering, and contrast equalization, then described by a 540-dimensional HOG feature vector and scored by the SVM as open or closed. A temporal rule, closed for a specified period, converts those frame-level labels into a drowsiness alarm.","core_discovery":"The central claim is that normalized cross-correlation-based online dynamic template matching, combined with Kalman filtering, can hold onto eye positions across frames well enough that a support vector machine with histogram of oriented gradients features can then classify each eye as open or closed; a temporal persistence rule, closed for a specified amount of time, then declares sleepiness and sounds an alarm. Face detection runs first using a fast LBP cascade, eyes are found inside the face with a Haar cascade, and eye templates are continuously refreshed from recent detections so the tracker adapts to changing appearance. The SVM is trained offline on 9,763 labeled eye images, of which 4,180 are closed and 5,583 are open, and the experiments compare LBP and HOG features, with HOG giving the better ROC curve. The paper reports the full system runs in real time with minimal computational complexity, can decide from a single visible eye, and detects eye state accurately for drivers wearing eyeglasses.","pith_inferences":["A natural extension not explored in the paper is lifting the same tracking-plus-classification skeleton to other driver state cues, such as head pose, yawning, or gaze direction, by replacing only the SVM label set.","Because no per-driver calibration is reported, a practical deployment would likely need a short personalization step or a much larger multi-driver training set to close the gap between offline training data and real in-vehicle conditions.","A directly testable extension is measuring how tracking success and classification accuracy degrade as camera-driver geometry, lighting, or glasses type changes; the claim of accurate detection with eyeglasses would be stronger if reported separately for glasses and no-glasses subsets.","The decision rule only looks at eye closure, so a driver who is awake but distracted, for example looking down at a phone, would not be flagged; adding a head-pose cue would extend coverage beyond the paper's stated scope."],"forward_implications":["The pipeline can run on embedded hardware because the only expensive step is face and eye detection, which is re-run only when tracking is lost.","Because single-eye monitoring is allowed, the system remains usable when the driver's face is turned and only one eye is visible to the camera.","Refreshing templates after re-detection lets the tracker recover from brief tracking failures and adapt to slow lighting or pose changes.","Frame-by-frame open/closed labels plus a temporal decision window give a simple, tunable alarm policy: adjusting the required closure period trades false alarms against missed drowsiness."],"supporting_citations":[{"why":"Supplies the face detector and the eye-detection cascade that initialize and re-initialize the tracking loop.","marker":"[15]"},{"why":"Provides the software implementation used for face and eye localization throughout the system.","marker":"[17]"},{"why":"Defines the LBP texture feature used as the baseline classifier feature for comparison.","marker":"[18]"},{"why":"Defines the HOG feature descriptor used by the SVM that labels each eye open or closed.","marker":"[19]"},{"why":"Contributes the labeled open/closed eye image set used to train the SVM classifier.","marker":"[20]"}],"fun_headline_variants":["Real-time eye tracker alarms on prolonged eye closure","SVM and template tracking detect drowsy drivers in real time","Closed eyes for too long triggers driver fatigue alarm","Instant driver drowsiness detection via eye state monitoring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The SVM is trained offline on a fixed set of eye images, and the blanket accuracy claims assume those images represent real in-vehicle conditions, including off-angle views, changing lighting, and eyeglasses, with no reported domain adaptation or per-driver calibration.","fun_headline_variants_meta":{"raw":{"variants":["Real-time eye tracker alarms on prolonged eye closure","SVM and template tracking detect drowsy drivers in real time","Closed eyes for too long triggers driver fatigue alarm","Instant driver drowsiness detection via eye state monitoring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2264,"prompt_tokens":849,"completion_tokens":1415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":1353}},"tokens_in":465,"tokens_out":1415,"duration_ms":9731,"temperature":1.0,"reasoning_tokens":1353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:14.983639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the system in a car with the camera mounted off to the driver's left, as the paper describes, over varied lighting and with and without eyeglasses, and record eye-tracking success and open/closed classification accuracy on a labeled frame set. If accuracy drops materially below the reported HOG ROC curve, or if the tracker re-initialization rate rises sharply during ordinary head motion, the claimed real-world accuracy does not hold.","supporting_citations":[{"cited_title":"International Journal of Computer Vision 57(2), 137--154 (2004)","cited_arxiv_id":null,"evidence_quote":"Supplies the face detector and the eye-detection cascade that initialize and re-initialize the tracking loop."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the software implementation used for face and eye localization throughout the system."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Defines the LBP texture feature used as the baseline classifier feature for comparison."},{"cited_title":"IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Defines the HOG feature descriptor used by the SVM that labels each eye open or closed."},{"cited_title":"IEEE Int","cited_arxiv_id":null,"evidence_quote":"Contributes the labeled open/closed eye image set used to train the SVM classifier."}],"review_version":1}