{"id":"c983d88b-a738-45d4-a3d1-d154d178e1b6","arxiv_id":"2607.20142","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A real-time webcam system detects EEG cap electrodes and assigns them named 10–20 roles via geometric rules, reaching 0.94 mAP@.5 in leave-one-subject-out evaluation.","lead":"Researchers built a two-stage vision system that detects EEG cap electrodes in a live webcam feed and checks each one's position against facial landmarks in real time. Non-specialists could use it to ensure a cap is placed correctly without a trained EEG technologist.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported mAP@.5=0.94 may be detector-only; with the mandatory adaptive crop, performance at native scale drops to 0.80, so the system-level detection accuracy is unreported.","rationale":"The reader's CONDITIONAL verdict stems from small dataset size, unreleased code, and unresolved Fz miss rates. My independent stress-test identifies a more specific and potentially more consequential concern: the evaluation metric does not clearly characterize the end-to-end system. The paper's abstract and Table 2 report the detector's mAP, while Section 4.3 and Section 5.2 indicate that the crop — an integral part of the system — reduces mAP from 0.94 to 0.80 on native-scale frames. This means the headline number may not reflect what a user would experience with the deployed pipeline. The paper is transparent about the crop's effect in an ablation, but it never integrates this into a system-level performance figure for the actual test distribution. This is a genuine gap in the evaluation, not a fatal error: the method, LOSO protocol, and robustness ablations are otherwise sound, and the limitation is fixable by reporting end-to-end performance and clarifying training details. Therefore, the reader's CONDITIONAL verdict remains appropriate — the paper should be conditioned on providing the missing system-level evaluation and clarifying the crop's role in training and testing. I do not think the concern warrants outright rejection, because the paper's core technical contribution — the two-stage architecture and the honest ablation of the crop trade-off — is still credible; however, the headline accuracy should be interpreted with caution until clarified.","tokens_in":11469,"tokens_out":12250,"duration_ms":115270,"concrete_test":"Re-run the five LOSO folds through the full pipeline as specified in Section 4.3: apply MediaPipe FaceLandmarker, generate the adaptive crop (using the full-frame fallback when no face is found), run YOLO on the upscaled crop, map detections back to full-frame coordinates, and compute mAP@.5 on the held-out folds. Compare this system-level mAP to Table 2. Additionally, inspect the training code to determine whether the detector was trained on full frames or on crops; if full-frame training is confirmed, retrain with crop-based augmentation and re-measure to determine whether the crop-induced performance drop at 1.0× is a train-test mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline detection metric (mAP@.5 = 0.94 ± 0.07, Table 2) is explicitly attributed to the YOLO detector, but the system described in Section 4.3 always applies an adaptive head crop before detection. Section 5.2's distance-robustness experiment (Figure 4) reports that at 1.0× scale (the native, unperturbed frames), with-crop mAP@.5 is 0.80, while the no-crop baseline is 0.94. Since the deployed pipeline applies the crop whenever MediaPipe finds a face, the full system's detection performance on the held-out LOSO frames is never reported; it is plausibly near 0.80 rather than 0.94. The paper does not state whether Table 2 was computed with or without the crop, and the discrepancy is not addressed. If Table 2 used full frames (no crop), the evaluation protocol does not match the deployable system. If it used the crop, the numbers contradict Figure 4. This ambiguity is load-bearing because the central claim — real-time electrode detection and placement validation — is supported by a metric that may not reflect the actual system's accuracy. A system-level mAP should be reported, and the training protocol regarding the crop should be clarified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a real-time, webcam-only system that detects EEG cap electrodes and assigns them to named 10–20 roles. Stage 1 is a YOLO-based single-class electrode detector; Stage 2 is a deterministic geometric role-assignment stage using MediaPipe facial landmarks and an adaptive head crop. The evaluation uses subject-disjoint leave-one-subject-out (LOSO) cross-validation on 1,309 frames from 5 subjects wearing three cap sizes, reporting detector mAP@.5 = 0.94 ± 0.07, per-electrode miss rates, a leave-one-cap-out axis, synthetic robustness ablations (lighting, distance, roll), role-assignment accuracy (90.7% of matched detections), and throughput figures. The authors are transparent about confounds (Small-cap leave-cap-out is subject-confounded), the single-rater and fixed-box annotation convention, the synthetic nature of robustness tests, and the persistent Fz detection difficulty.","tokens_in":11845,"tokens_out":6382,"duration_ms":62473,"significance":"If the reported system-level performance matched the headline detector numbers, this would be a practically useful step toward automated EEG cap placement quality assurance in point-of-care settings. Strengths of the paper are the leakage-free LOSO protocol, the explicit leave-one-cap-out axis, the honest treatment of subject/cap confounding, the reproducible synthetic perturbation methodology, the deterministic and clearly specified geometric role-assignment rule, and the real-time throughput characterization. The main gap is that the reported headline metric appears to be detector-only, while the deployed system includes an adaptive crop whose effect at native scale is non-negligible; the end-to-end placement-validation accuracy, which is the clinically relevant quantity, is not reported.","major_comments":[{"comment":"The headline detection metric is not the deployed system's metric. Figure 4 (left) reports that at 1.0× scale, with the adaptive head crop the mAP@.5 is 0.80, while the no-crop full-frame baseline is 0.94. Table 2's LOSO mAP@.5 = 0.94 ± 0.07 matches the no-crop baseline, and §5.2 states 'Without the crop, mAP@.5 falls from 0.94 ± 0.07 ...' — so Table 2 is evidently detector-only on full frames. However §4.3 states the deployed pipeline always applies the crop when a face is found. Thus the system as actually deployed has not been evaluated on the held-out LOSO folds. Please report system-level mAP@.5 and all-electrodes-found rate with the crop applied exactly as in §4.3, and state unambiguously which configuration produced Table 2.","section":"§5.2 / Figure 4 / Table 2"},{"comment":"End-to-end placement-validation accuracy is missing. The paper's central claim is that the system 'validates their anatomical placement in real time,' but detection and role assignment are reported separately. Table 2 gives the all-electrodes-found detection rate (0.70 pooled; 0.07 on fold s4), and §5.4 gives role-assignment accuracy (90.7% of matched detections). No metric reports the fraction of frames in which the system finds all five electrodes and assigns all five roles correctly, under the deployed crop. This joint per-frame metric is the quantity a clinical user would care about; please add it.","section":"§5.4 / Table 2"},{"comment":"The training/evaluation protocol regarding the crop is unspecified. §4.1 says 'All results in this paper use this configuration,' but it is not clear whether training images were full-frame or cropped. If the detector was trained on full frames only, the inference-time crop is a train/test domain shift that may explain the 1.0× degradation in Figure 4. Please describe the training data transform precisely and, if the crop is used at inference, either include it in training or justify its absence.","section":"§4.1 / §4.3"}],"minor_comments":[{"comment":"The dataset is small (5 subjects, 1,309 frames), annotations come from a single rater, and boxes are fixed-size squares. These limitations are acknowledged, but adding a brief statement on the absence of inter-rater reliability and the potential bias from a single rater would strengthen the reproducibility discussion.","section":"§3 / Dataset"},{"comment":"The fallback to the full frame when no face is found is stated, and the distance experiment shows that performance collapses at 0.3× when MediaPipe fails. It would be helpful to state explicitly in the limitations that the system cannot perform role assignment at all in such cases, and that the operational envelope therefore depends on reliable face detection.","section":"§4.3 / §5.2"},{"comment":"Table 4 reports YOLOv8s at 640 px with 0.86 ± 0.19 mAP@.5, while Table 2 reports YOLOv8s at 960 px with 0.943 ± 0.074. The resolution difference is clear from the text, but a sentence in the Table 4 caption restating that these are 640-px operating points would avoid confusion.","section":"Section 5.6 / Table 4"},{"comment":"The manuscript states that code and dataset are not publicly released. For a computer-vision methods paper, releasing at least the evaluation code and a small sample of anonymized data would substantially aid reproducibility; the current statement is a limitation for verification.","section":"Data and code availability"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the authors are unusually honest about confounds and limitations. The main blocker is the discrepancy between the headline detector-only mAP and the deployed crop-based pipeline; the system-level and end-to-end metrics need to be reported before acceptance. I would also encourage the editor to weigh the lack of code/data release and the small single-rater dataset when considering the paper's contribution, though these are not technical errors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read of arXiv:2607.20142. The genuinely new bit is that they detect the physical electrodes instead of predicting where they should be from an atlas, then assign roles geometrically from facial landmarks. That closes a loop prior work left open, and it's evaluated with a proper subject-disjoint LOSO protocol, honest error bars, and a separate cap-out axis. The robustness ablations are synthetic but clearly labeled as such. Credit where due: the paper is well-written, transparent about its limitations, and the Fz failure analysis is candid.\n\nThe main soft spot is exactly the stress-test note. The headline mAP@.5=0.94 (Table 2) is reported for the detector, but the deployable pipeline always applies the adaptive head crop when a face is found. Section 5.2 shows at 1.0x scale the with-crop mAP is 0.80 vs 0.94 no-crop. So the actual system's detection performance on the held-out frames is never reported; it's plausibly ~0.80, and the paper doesn't address this. That's load-bearing because the central claim is real-time electrode detection for placement validation.\n\nOther soft spots, in proportion. The dataset is small: 5 subjects, 1,309 frames, all from one cap model, annotated by a single rater, fixed-size square boxes. The all-electrodes-found rate is 0.70, and Subject 4's Fz is missed in 93% of frames; excluding that fold, the average is 0.85, but the paper reports the full number as headline, which is honest but still means the system is not reliable for the vertex in at least one subject. Role assignment is 90.7% on matched detections, Fz leaking heavily. Code and data are not released, which limits reproducibility. No critical red flags beyond the crop/metric mismatch; I don't see fitting to the test set, and the hand-set thresholds are described as heuristics, not tuned against the metrics.\n\nWho's this for: people building practical EEG placement guidance or applying YOLO+landmark pipelines in clinical settings. It deserves a serious referee; the crop ambiguity needs to be resolved, ideally by reporting a single system-level mAP with the actual pipeline, and the authors should release code/data and validate on more subjects/caps before a strong clinical claim. My verdict: conditional accept after major revision, but the contribution is real.","headline":"The direct electrode-detection idea is real and the eval is mostly honest, but the headline mAP comes from the no-crop baseline while the deployed system always crops, so the system-level accuracy is never actually reported.","tokens_in":12268,"tokens_out":2085,"would_cite":false,"duration_ms":18789,"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 webcam-only system detects EEG cap electrodes in real time and checks they sit at the correct 10–20 head positions, reaching 0.94 mAP across held-out subjects.","keywords":["EEG","electrode placement","10-20 system","object detection","YOLO","facial landmarks","real-time guidance","point-of-care"],"falsifier":"Run the pipeline on a held-out subject wearing the Small cap under realistic dim lighting, where MediaPipe locates the face but the detector must rely on appearance alone; the paper predicts a sharp drop in mAP for the Small cap when subject familiarity is also removed, but if the drop vanishes with more subjects or if face detection succeeds and role assignment still fails, the geometric-stage assumption would be contradicted.","tokens_in":11419,"feed_emoji":"🧠","tokens_out":1169,"duration_ms":15941,"temperature":0.7,"pith_summary":"This paper aims to close a safety gap in point-of-care EEG: when a non-specialist dons a cap, misplaced electrodes silently corrupt the recording. The authors propose a two-stage vision system that runs on an ordinary webcam, detects the physical electrodes on the head, and assigns each one a named 10–20 role using facial landmarks as an anatomical anchor. In subject-disjoint tests across five subjects wearing three cap sizes, the detector reaches 0.94 mean average precision (0.96 pooled), and the geometric role-assignment stage correctly labels 90.7% of matched detections. The system runs at near real-time speed on a commodity CPU, making it a plausible low-cost quality-assurance tool for settings where trained EEG technologists are scarce.","feed_headline":"Webcam checks EEG electrode placement in real time","feed_subtitle":"Detector finds cap electrodes and labels their 10–20 roles via facial landmarks, reaching 0.94 mAP on held-out subjects.","key_machinery":"The load-bearing object is the geometric role-assignment stage (Stage 2), which converts anonymous detections into named anatomical roles. It uses the inter-eye distance from MediaPipe FaceLandmarker as a subject-specific reference scale (d_ref), then classifies each detection as a temporal electrode (F9/F10) if it lies beyond a horizontal margin past the outer eye corner and above a crown threshold, and assigns inner electrodes by greedy nearest-neighbour proximity to eyebrow landmarks. This deterministic rule is what turns a generic electrode detector into a placement validator, and it is evaluated directly against the detector's output.","core_discovery":"The paper's central claim is that electrode placement can be verified continuously during cap donning using only a single RGB camera, without depth sensors, markers, or post-hoc scanning. The key move is to separate detection from naming: a class-agnostic YOLO detector finds electrode-like blobs, and a deterministic geometric rule assigns each blob to a specific 10–20 role (FP1, FP2, Fz, F9, F10) based on facial landmarks. This design avoids needing per-role training data and lets the system check the electrodes' actual positions rather than assuming they match an atlas. The authors also show that a landmark-driven head crop is the decisive factor for maintaining accuracy at distance, liftin","pith_inferences":["The 10–20 naming rule depends on a roughly frontal head pose and on the inter-eye distance being a reliable scale anchor; a natural extension would be to test pose robustness beyond the synthetic roll sweep, e.g., with real side or tilted head captures.","The paper's evaluation focuses on five fixed electrode roles; extending to the full 10–20 system (e.g., C3/C4, O1/O2) would require either more landmark anchors or a richer geometric model, but the same architecture should transfer.","If the system were deployed, a prospective clinical study comparing placement correctness (measured by a gold standard like optical scanning) with and without real-time guidance would be the decisive test of whether the vision loop actually improves EEG signal quality."],"forward_implications":["Non-specialist operators could receive real-time feedback on whether each electrode is in its correct 10–20 position during donning, potentially reducing the error rate that currently corrupts point-of-care EEG.","The two-stage design (detection + geometric naming) can be ported to other electrode caps or extended to additional 10–20 sites without retraining the detector, since naming is rule-based.","The landmark-driven head crop establishes a general principle: for small-object detection on the head, using facial landmarks to upscale the region of interest is more effective than simply increasing input resolution, and it works at no inference cost.","A compact mobile backbone (YOLOv10n) reaches the real-time band on a CPU at 640 px with near-equal accuracy, suggesting the system can run on a phone or low-cost device in low-connectivity settings."],"fun_headline_variants":["Webcam verifies EEG electrode positions in real time","Single camera labels EEG electrodes without markers","Real-time EEG cap check: YOLO finds electrodes, geometry names them","EEG cap electrodes detected and named by webcam at 19 FPS","Zero-marker EEG cap validation via facial landmarks + YOLO"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline assumes MediaPipe can locate the subject's face in the webcam frame; if face detection fails (poor lighting, extreme pose, occlusion, or distance beyond 0.3× scale), the adaptive crop is abandoned, the anatomical reference frame is lost, and the system cannot assign electrode roles.","fun_headline_variants_meta":{"raw":{"variants":["Webcam verifies EEG electrode positions in real time","Single camera labels EEG electrodes without markers","Real-time EEG cap check: YOLO finds electrodes, geometry names them","EEG cap electrodes detected and named by webcam at 19 FPS","Zero-marker EEG cap validation via facial landmarks + YOLO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000412,"raw_usage":{"total_tokens":1996,"prompt_tokens":796,"completion_tokens":1200,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":1114}},"tokens_in":540,"tokens_out":1200,"duration_ms":9947,"temperature":1.0,"reasoning_tokens":1114,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T10:37:41.854051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a held-out subject wearing the Small cap under realistic dim lighting, where MediaPipe locates the face but the detector must rely on appearance alone; the paper predicts a sharp drop in mAP for the Small cap when subject familiarity is also removed, but if the drop vanishes with more subjects or if face detection succeeds and role assignment still fails, the geometric-stage assumption would be contradicted.","supporting_citations":[],"review_version":1}