{"id":"625983a3-07c2-45c5-b619-cf8f407a8b76","arxiv_id":"2505.22769","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MAC-Gaze uses IMU-based motion detection and replay-based continual learning to automatically recalibrate mobile gaze trackers, reducing gaze estimation error by 19.9% on RGBDGaze and 31.7% on MotionGaze.","lead":"Mobile gaze trackers lose accuracy as users move, because one-time calibration assumes a fixed head-to-phone geometry. This paper introduces MAC-Gaze, which uses phone motion sensors to detect when recalibration is needed and updates the gaze model continuously, cutting gaze error by up to 31.7% in their tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Replay-ratio grid (Table 7) says best MAC-Gaze is 1.59 cm on RGBDGaze, yet the main result (Table 3) is 1.41 cm; the headline improvement is not internally reproducible.","rationale":"The paper's contribution is an automatic, motion-aware continual calibration system, and the headline evidence is the reported error reduction. An internal inconsistency in the two tables that both claim to report MAC-Gaze's RGBDGaze performance is therefore the most load-bearing concern: it directly affects whether the central quantitative claim is trustworthy. The replay-ratio sweep in Table 7 is explicitly described as testing MAC-Gaze on RGBDGaze, so a 0.18 cm gap between its best average (1.59 cm) and the main result (1.41 cm) cannot be dismissed as a different dataset or a different ablation. The discrepancy is not merely cosmetic; it changes the reported improvement over one-off calibration from roughly 20% to under 10%. A single targeted rerun under the stated optimal configuration would resolve whether Table 3 or Table 7 is the accurate description of the system. The reader identified the IMU-to-gaze causal link as the weakest assumption; I agree that causality is unvalidated, but the paper's own tables provide a more immediate, checkable threat to the headline claim. The correct disposition is therefore conditional: the paper should be accepted only after the authors reconcile Tables 3, 6, and 7 and provide the exact configuration and validation split used for the main result.","tokens_in":24331,"tokens_out":5129,"duration_ms":54658,"concrete_test":"Re-run the RGBDGaze experiment of Table 3 under the exact configuration that Table 7 labels as best (70% replay ratio, balanced memory buffer, same 10% calibration splits, same threshold grid) and report the average. If the result is approximately 1.59 cm, the headline 1.41 cm is not reproducible under the paper's own optimal setting; if it is approximately 1.41 cm, Table 7's entries or the protocol description must be corrected. Also require the authors to state which replay ratio and threshold-selection procedure produced Table 3 and whether validation was held out.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The dominant quantitative claim is the 19.9% RGBDGaze reduction (1.73 to 1.41 cm) and the 31.7% MotionGaze reduction in Table 3 and the abstract. Table 7 (Section 5.4) reports MAC-Gaze replay-ratio sweeps on RGBDGaze across four initial postures; the best average is 1.59 cm at 70% replay. No cell in Table 7 reaches 1.41 cm. Table 6 also lists the full system as 1.41, but without specifying the replay ratio used. Since Table 7 claims to evaluate the same MAC-Gaze design and the same RGBDGaze protocol, the main result is internally inconsistent. If the true best configuration yields 1.59 cm, the claimed advantage over one-off calibration (1.74 cm) drops from 19.9% to about 8.6%, which materially changes the strength of the central claim. A secondary issue is that Oracle Motion-Aware is presented as an upper bound, yet MAC-Gaze beats it on both datasets; the protocol needs to clarify why a continual model with more accumulated data is being compared to a per-motion oracle. The reader's weakest assumption about IMU-trigger causality is legitimate, but it is secondary here: even if IMU triggering were perfect, the reported quantitative gain is not internally reproducible as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MAC-Gaze, a motion-aware continual calibration framework for mobile gaze tracking. It combines a pre-trained visual gaze estimator with an IMU-based activity recognition model and a hybrid trigger that uses supervised classification plus GMM-based outlier detection to decide when to recalibrate. Calibration updates are performed with replay-based continual learning to avoid catastrophic forgetting. The system is evaluated on the RGBDGaze dataset and a newly collected MotionGaze dataset, reporting 19.9% and 31.7% error reductions over traditional calibration baselines, together with ablations and a generalization study using iTracker as the backbone.","tokens_in":24648,"tokens_out":5645,"duration_ms":56739,"significance":"If the quantitative claims hold, the work is significant: it addresses a practically important gap in mobile gaze tracking—deciding when to trigger recalibration and how to update the calibrator without forgetting previous motion conditions—and it contributes a new multimodal dataset with IMU, image, and gaze data. The ablations isolating the hybrid trigger, replay, and motion-based timing are useful, as is the demonstration that the approach generalizes to a second backbone. The main limitation is that the headline number is not internally reproducible from the reported tables, which must be resolved before the claimed improvement can be assessed.","major_comments":[{"comment":"Section 5.4 reports a replay-ratio sweep for MAC-Gaze on RGBDGaze whose best average error is 1.59 cm at a 70% replay ratio, yet Table 3 and Table 6 report 1.41 cm for the full MAC-Gaze system, and no cell in Table 7 reaches 1.41 cm. The paper does not state which replay ratio or protocol difference produces the 1.41 cm result. This makes the headline 19.9% improvement internally unreproducible; the abstract's 1.73 cm baseline also matches neither the 1.74 cm one-off calibration row nor the 1.76 cm no-calibration row in Table 3. If the correct value were 1.59 cm, the improvement over one-off calibration would be only about 8.6%, materially changing the central claim.","section":"Section 5.4, Table 7; Section 5.1, Table 3; Abstract"},{"comment":"The Oracle Motion-Aware baseline trains a separate calibration model for each motion using only 10% of that motion's data, with no replay or accumulation across motions. It is therefore not an upper bound in terms of available training data: MAC-Gaze uses a growing replay buffer containing data from all previously encountered motions. Beating the oracle (1.41 vs 1.54 cm on RGBDGaze; 1.92 vs 2.23 cm on MotionGaze) may reflect the oracle's smaller calibration set rather than superior triggering. The oracle should be given matched replay/accumulation, or MAC-Gaze should be compared with equal total calibration data, to support the claim that MAC-Gaze outperforms perfect motion detection.","section":"Section 5.1, Oracle Motion-Aware definition"},{"comment":"The grid search over detection thresholds is described as 'select the best setting' without stating whether the selection uses the held-out validation subset or the test participants; Section 3.4 says thresholds were determined on a held-out validation subset, but Section 4.4.3 does not confirm this or report the selected values. Similarly, Section 5.4 sweeps replay ratios on RGBDGaze and identifies 70% as best, but does not say whether the main results in Table 3 use that ratio or a different one. Please report the selected threshold and replay-ratio values, specify the validation split, and clarify the relationship between Table 7 and Table 3; otherwise the reported error reductions are open to outcome-dependent selection.","section":"Section 4.4.3 and Section 5.4"},{"comment":"The 'when to calibrate' contribution asserts that IMU-detected motion changes are a reliable proxy for gaze-error degradation, but the paper only reports trigger counts and qualitative timelines. There is no quantitative evaluation of whether triggered recalibrations coincide with measured increases in gaze error, or of false-trigger rates relative to error degradation; the matched-count time-based comparison in Table 6 is a useful start but does not provide precision/recall of the trigger. Figure 7 shows participants p11 and p17 whose error increases after recalibration, which the paper attributes to temporal mismatch, but no aggregate statistic is given. Please add a per-trigger or per-segment analysis to validate the association between detected motion changes and actual calibration need.","section":"Section 5.2, Section 5.3, Figure 7"}],"minor_comments":[{"comment":"Section 3.4 uses τ for the consensus ratio and τ_i for the GMM likelihood threshold, but Section 4.4.3 refers to 'the two thresholds for outlier detection τ and ρ' and never defines ρ; the threshold notation and the selected values should be aligned.","section":"Section 3.4 and Section 4.4.3"},{"comment":"Figure 7 labels the y-axis 'MSE(cm)' and the text calls the metric MSE, while Eq. (2) and the rest of the paper use mean Euclidean error; please unify the terminology.","section":"Figure 7 and Eq. (2)"},{"comment":"There are a few typos: 'more fine-course' should likely be 'more fine-grained', and the keyword list contains 'Continue Learning' instead of 'Continual Learning'.","section":"Section 6.3 and Keywords"},{"comment":"The caption of Figure 5 refers to 'True Label' and 'Weak Label' while the text describes ground-truth versus classifier predictions; the distinction between 'Weak Label' and classifier prediction is not explained.","section":"Figure 5 caption"},{"comment":"Table 5 reports average data length per participant (3,869 frames for RGBDGaze, 17,521 for MotionGaze) without stating units or whether frames are evenly distributed across postures, which would aid interpretation of the calibration-frequency comparison.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a human-computer-interaction venue and addresses a timely problem. The central idea is promising, but the inconsistency between Table 7 and Table 3 must be resolved, and the oracle baseline and threshold-selection protocol need to be made precise. I would be willing to re-review a revised version that clarifies these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is good and the paper is worth taking seriously, but the headline number does not survive internal cross-checking. On RGBDGaze, Table 3 reports MAC-Gaze at 1.41 cm average error; Table 7, which sweeps replay ratios for the same system and same dataset, bottoms out at 1.59 cm at 70% replay. No cell in Table 7 reaches 1.41 cm. If 1.59 cm is the real best configurable result, the claimed 19.9% improvement over one-off calibration (1.74 cm) becomes roughly 8.6%. That is not a cosmetic typo; it changes the strength of the central claim. This needs to be explained or corrected before the paper can be accepted.\n\nWhat is genuinely new is the combination: IMU-triggered recalibration with replay-based continual learning for mobile gaze calibration, plus a hybrid trigger that uses a supervised activity classifier together with GMM clustering on raw IMU data to catch fine-grained posture changes. The MotionGaze dataset is a useful contribution even if it is described in the authors' prior work. The ablations are informative: removing replay hurts badly, and time-based recalibration with the same number of calibrations does much worse, which supports the central \"when to calibrate matters\" thesis. Table 4's diagonal versus off-diagonal one-off results also give clean evidence that motion-state mismatch degrades calibrated gaze error.\n\nThe softer spots are real but secondary. The Oracle Motion-Aware baseline is not a true upper bound: it trains a separate calibrator on 10% of each motion, while MAC-Gaze accumulates data across the sequence, so beating Oracle is not as surprising as the paper implies. The replay ratio appears to be selected after looking at the test sweep rather than on a validation split; the paper does not say otherwise. The evaluation description in Section 4.3 leaves ambiguous whether the 10% used for recalibration is disjoint from the 90% of frames used for testing; if not, results leak. Finally, the link between IMU deviation and gaze error is assumed from prior work rather than validated per participant, but that is secondary to the numerical inconsistency.\n\nThe paper is aimed at mobile gaze researchers and people working on adaptive calibration. It deserves a serious referee, not a desk reject. I would send it to review and ask the authors to resolve the Table 3/Table 7 discrepancy, clarify the calibration/test split, redefine or reposition the Oracle, and release code and data to make the numbers independently checkable.","headline":"The core idea is solid and worth refereeing, but the main accuracy claim is internally inconsistent: Table 3 reports 1.41 cm while Table 7, the replay-ratio sweep on the same dataset, never goes below 1.59 cm.","tokens_in":25185,"tokens_out":3283,"would_cite":true,"duration_ms":37378,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A phone's own motion sensors can decide when gaze tracking needs recalibration, and replay-based continual learning makes the fix stick, cutting average error from 2.81 cm to 1.92 cm on naturalistic data.","keywords":["gaze estimation","mobile gaze tracking","continual calibration","IMU activity recognition","replay-based continual learning","hybrid calibration trigger","MotionGaze","catastrophic forgetting"],"falsifier":"A recording with continuous gaze ground truth and IMU data in which a user makes large gaze shifts while holding the phone perfectly still would settle the trigger's necessity: if gaze error rises sharply with no IMU deviation, the motion signal is missing the relevant change. The converse test, IMU deviations that produce no rise in gaze error, would reveal over-triggering.","tokens_in":24142,"feed_emoji":"👀","tokens_out":8732,"duration_ms":92076,"temperature":0.7,"pith_summary":"Mobile gaze tracking on smartphones drifts as users shift posture, change how they hold the device, or move while walking; one-off calibration cannot keep up. The paper aims to show that calibration can be made continual and automatic: a phone's IMU sensors detect when the user's motion state changes, and a replay-based continual learner updates the gaze calibrator without forgetting earlier motion conditions. The reported result is that this MAC-Gaze system lowers average gaze error from 1.73 cm to 1.41 cm on the RGBDGaze dataset and from 2.81 cm to 1.92 cm on the newly collected MotionGaze dataset, compared with traditional one-off calibration. A sympathetic reader would care because this is a practical route to keeping gaze interfaces usable during everyday phone interaction without repeated manual calibration.","feed_headline":"Phone motion triggers recalibration, cutting gaze error by 31.7%","feed_subtitle":"System uses IMU sensors and replay learning to cut average error from 2.81 cm to 1.92 cm on naturalistic data.","key_machinery":"The load-bearing mechanism is the calibration trigger plus the replay buffer working as a pair. The hybrid trigger first applies temporal majority voting to the supervised activity predictions of the IMU model, so a single misclassification does not fire a recalibration; it then scores each raw IMU reading by its maximum likelihood under a Gaussian mixture model built from previously stored motion samples, and declares a new fine-grained motion pattern when the outlier ratio in the voting window passes a threshold. When a trigger fires, the continual calibration module updates a two-layer MLP using a loss that supervises both the new calibration samples and replayed samples from the buffer, which prevents catastrophic forgetting. The paper reports that removing the buffer raises error from 1.41 cm to 1.73 cm, and replacing the hybrid trigger with the classifier alone raises it to 1.58 cm, which is the direct evidence that both halves of the mechanism carry the result.","core_discovery":"The paper's central claim is that the two open questions in mobile gaze calibration—when to recalibrate and how to recalibrate—can be answered together with motion sensing and replay-based continual learning. For 'when,' MAC-Gaze combines supervised activity recognition (an IMU encoder-decoder with a classification head) with unsupervised Gaussian mixture clustering on raw IMU samples; recalibration is triggered when the stable classified activity changes or when the current IMU stream largely falls outside previously seen GMM components. For 'how,' a small MLP calibrator is retrained on new calibration data together with a memory buffer of samples from earlier motion states, balancing adaptation with retention. On RGBDGaze and MotionGaze, the full system outperforms no calibration, one-off calibration, and classifier-only triggering, and on RGBDGaze it also outperforms an oracle that knows every motion transition in advance.","pith_inferences":["A direct test of the paper's implicit causal claim would compare gaze error and IMU outlier scores continuously; if error rises before the IMU trigger fires, the motion signal is lagging the true domain shift.","The choice to cluster raw IMU readings rather than learned latent features implies that the activity-classification representation discards exactly the micro-motion information calibration needs, which motivates future work on hierarchical or multi-scale representations that preserve both.","An extension the paper leaves implicit is combining the motion trigger with implicit visual signals such as touch, reading, or on-screen content so that users who change gaze behavior without changing phone motion can still trigger recalibration.","The per-participant fluctuation seen for highly dynamic users suggests that motion-predictive models, rather than reactive triggers, are the next step for the most active device-handling patterns."],"forward_implications":["Motion-triggered recalibration reduces average 2D gaze error from 1.73 cm to 1.41 cm on RGBDGaze and from 2.81 cm to 1.92 cm on MotionGaze, beating one-off calibration in every tested posture.","When recalibration happens matters as much as how often: giving participants the same number of calibrations at fixed time intervals yields 1.71 cm versus 1.41 cm for motion-triggered calibration.","Replay-based learning is what prevents forgetting: without the memory buffer, error jumps from 1.41 cm to 1.73 cm.","The calibration approach wraps around different gaze backbones; with iTracker as the base model, error drops by 22.0% on RGBDGaze and 40.1% on MotionGaze.","Starting posture does not change the outcome: one-way ANOVA across four initial calibration postures gives p = 0.782, meaning the continual learner adapts regardless of task order."],"supporting_citations":[{"why":"Supplies the RGBD smartphone gaze dataset used to evaluate the system.","marker":"[2]"},{"why":"Supplies the phone IMU activity dataset used to train the supervised motion detector.","marker":"[6]"},{"why":"Supplies the large mobile gaze dataset used to pretrain the backbone gaze estimator.","marker":"[35]"},{"why":"Supplies the authors' own multimodal gaze-plus-IMU dataset used as the second evaluation benchmark.","marker":"[39]"},{"why":"Supplies the supervised autoencoder architecture that the IMU motion detector adapts.","marker":"[1]"},{"why":"Supplies the efficient transformer-based backbone used to extract visual gaze features.","marker":"[48]"},{"why":"Supplies the survey evidence that motivates replay-based continual learning for this task.","marker":"[30]"}],"fun_headline_variants":["Motion-aware calibration cuts gaze error 31.7% on naturalistic data","Gaze tracking recalibrates on the fly, slashing error by 31.7%","IMU-driven recalibration keeps mobile gaze accurate as you move","Continual learning recalibrates gaze when motion shifts, error down 31.7%","Phone's motion sensors trigger gaze recalibration, boosting accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that phone motion sensed by the IMU is a reliable indicator of when the head-eye-camera relationship behind gaze error has changed, a link the paper adopts from prior work rather than validating directly per participant.","fun_headline_variants_meta":{"raw":{"variants":["Motion-aware calibration cuts gaze error 31.7% on naturalistic data","Gaze tracking recalibrates on the fly, slashing error by 31.7%","IMU-driven recalibration keeps mobile gaze accurate as you move","Continual learning recalibrates gaze when motion shifts, error down 31.7%","Phone's motion sensors trigger gaze recalibration, boosting accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000503,"raw_usage":{"total_tokens":2487,"prompt_tokens":1006,"completion_tokens":1481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":1394}},"tokens_in":622,"tokens_out":1481,"duration_ms":12588,"temperature":1.0,"reasoning_tokens":1394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:00:45.731429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A recording with continuous gaze ground truth and IMU data in which a user makes large gaze shifts while holding the phone perfectly still would settle the trigger's necessity: if gaze error rises sharply with no IMU deviation, the motion signal is missing the relevant change. The converse test, IMU deviations that produce no rise in gaze error, would reveal over-triggering.","supporting_citations":[{"cited_title":"Rgbdgaze: Gaze tracking on smartphones with rgb and depth data","cited_arxiv_id":null,"evidence_quote":"Supplies the RGBD smartphone gaze dataset used to evaluate the system."},{"cited_title":"Eye tracking for everyone","cited_arxiv_id":null,"evidence_quote":"Supplies the large mobile gaze dataset used to pretrain the backbone gaze estimator."},{"cited_title":"A supervised autoencoder for human activity recognition with inertial sensors","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised autoencoder architecture that the IMU motion detector adapts."},{"cited_title":"Continual learning in sensor-based human activity recognition: An empirical benchmark analysis.Information Sciences 575(2021), 1–21","cited_arxiv_id":null,"evidence_quote":"Supplies the survey evidence that motivates replay-based continual learning for this task."}],"review_version":1}