{"id":"32be76ef-23fa-4ad5-9ef9-da7bae89bc56","arxiv_id":"2506.07860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An event-camera and eye-tracking system on smart glasses detects a ping-pong ball and predicts its landing point from the player's viewpoint at 200 Hz with a reported 4.5 ms total latency.","lead":"This paper builds a system that tracks a ping-pong ball from a player's smart glasses using event cameras, which measure light changes in microseconds, and uses eye gaze to focus on the ball. It reports reactions in a few milliseconds and better landing-point predictions than the glasses' ordinary 30 fps camera.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8.4 cm learning-based error-reduction claim in §1 is an unsupported cross-model comparison: Table 6 has no 30 Hz DCGN baseline.","rationale":"The reader's weakest assumption is the monocular depth-from-radius sensitivity. That is a real limitation and deserves a sensitivity analysis, but it does not invalidate the relative event-vs-frame comparison because both pipelines share the same depth estimator. The more immediately load-bearing concern is an internal inconsistency in the paper's central quantitative claim: the abstract and contribution list advertise an 8.4 cm error reduction 'with learning-based prediction methods' attributable to the higher update frequency, yet the only support in Table 6 is a cross-model gap (DCGN at 200 Hz versus the physics-based model at 30 Hz). No 30 Hz DCGN baseline is reported, so the improvement cannot be attributed to update rate as claimed. This is a correctness issue in the presentation of the headline numbers, not merely a robustness concern. It warrants a revision: either add the missing 30 Hz DCGN experiment or rewrite the contribution to state that a stronger learned predictor achieves lower absolute error, without claiming this is a frequency-driven gain. The reader's CONDITIONAL verdict remains appropriate; the depth analysis and the DCGN matched-frequency experiment are both conditions for final acceptance.","tokens_in":16487,"tokens_out":13371,"duration_ms":155599,"concrete_test":"Re-run the online forecasting protocol of Table 6 with the DCGN baseline fed only the measurements available to a 30 Hz frame-based pipeline (subsample the estimated ball states to one measurement every 33 ms, or retrain DCGN on 30 Hz-upsampled trajectories). If the RMSE gap between this 30 Hz DCGN and the reported 200 Hz DCGN (0.1072 m) is much smaller than 8.4 cm, the §1 contribution bullet overstates the frequency benefit. Alternatively, analytically recompute the 8.4 cm as 0.1915 - 0.1072 and verify that it compares different model classes; the claim holds only if a matched 30 Hz DCGN shows a similar degradation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The contribution list states that 'with learning-based prediction methods [16], this error reduction further improves to 8.4 cm,' attributing the gain to the higher measurement frequency of the event pipeline. In the online-forecasting table (Table 6), the only 8.4 cm gap is 0.1915 m (Diff. Eq., 30 Hz) minus 0.1072 m (DCGN, 200 Hz). This is a cross-model comparison: DCGN is a different predictor, not the same predictor at a lower update rate. No 30 Hz DCGN condition is reported anywhere in the paper or supplementary material. Therefore the claim that the event camera's 200 Hz updates 'improve' a learning-based predictor by 8.4 cm over frame-based updates is not supported by the experiments; the improvement could be entirely due to DCGN's larger model capacity. This directly weakens the core quantitative contribution, which is positioned as a benefit of event-based update frequency.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a real-time egocentric table-tennis ball trajectory prediction system built on an event camera mounted on Meta Project Aria glasses. The pipeline uses eye-gaze-based region-of-interest cropping, motion compensation, DBSCAN clustering, and circle fitting to detect the ball and estimate its monocular depth from the apparent ball radius. Trajectory prediction is performed by monotonically constrained polynomial regression, an Extended Kalman Filter, and a physics-based differential-equation model (with drag and a simplified bounce model). The authors contribute a synchronized multi-sensor dataset with OptiTrack ground truth, report a 92.59% ball detection rate, a mean computation latency of 1.5 ms and an 'ideal worst-case total latency' of 4.5 ms, and claim that 200 Hz event-based updates reduce impact-point RMSE by 4.8 cm compared with 30 Hz frame-based updates for physics-based prediction, and by 8.4 cm when using a learning-based predictor.","tokens_in":16667,"tokens_out":8270,"duration_ms":94926,"significance":"If the quantitative claims held up, this would be a useful first demonstration of egocentric event-based table-tennis perception, with the additional value of a public dataset and open-source code. The use of OptiTrack ground truth, literature-based aerodynamic constants, and a real event-camera platform are concrete strengths. However, the paper's central quantitative contributions are currently weakened by small-sample evaluation, selection of detection thresholds on the evaluation data, an unsupported cross-model comparison behind the 8.4 cm claim, an unmeasured 'worst-case' latency figure, and an acknowledged but unquantified sensitivity of monocular depth to radius estimation errors. These issues are fixable but require additional experiments and re-analysis before the claims can be considered robust.","major_comments":[{"comment":"The sentence in the contributions list that learning-based prediction methods [16] improve the error reduction to 8.4 cm is not supported by the reported experiments. Table 6 contains a DCGN run at 200 Hz (0.1072 m) and a differential-equation run at 30 Hz (0.1915 m), but no DCGN baseline at 30 Hz or any frame-based update rate appears in the paper or the supplementary material. The 8.4 cm gap is therefore a cross-model comparison, not a demonstration that the event camera's 200 Hz update rate improves a learning-based predictor. The authors should either add a DCGN model fed with 30 Hz measurements or remove and rephrase this claim.","section":"Section 1 and Table 6"},{"comment":"The detection threshold θ1 is selected by scanning values on the same evaluation sequences used to report the 92.59% detection rate, so the reported rate is a training-set maximum rather than an unbiased estimate. The percentages in Table 3 correspond to 27 trajectories (92.59% = 25/27, 88.88% = 24/27, 59.26% = 16/27), and no confidence interval or significance test is provided; the cropping gain in Table 4 (92.59% vs 81.48%) is a difference of three sequences. A validation split or cross-validation, together with bootstrap confidence intervals, is needed before the detection claims can be considered load-bearing.","section":"Section 5.2, Tables 3 and 4"},{"comment":"The claim of a 'worst-case total latency of 4.5 ms' is not supported by the measurements. Table 2 reports the mean computation time (1.497 ms, σ = 0.696 ms), not a worst case, and the 4.5 ms value appears to combine this mean with an ideal theoretical sensing latency derived in Supplementary Section 9 under idealized assumptions about event generation. No measured end-to-end latency from event arrival to predicted output is reported. The authors should report the actual maximum or a high percentile of computation time and, if possible, measured end-to-end latency, and should avoid calling the theoretical composite a worst case.","section":"Section 4 and Table 2"},{"comment":"The entire 3D trajectory estimate rests on monocular depth computed from the apparent ball radius via Z = f W_metric / r_hat, where r_hat is obtained by fitting a circle to three points on the convex hull. Supplementary Section 13 acknowledges that at 2–3 m distance and VGA resolution, a one-pixel error in the estimated radius changes depth by several centimeters, but the paper never quantifies this sensitivity or its propagation into the impact-point RMSE. Without an error-propagation or ablation analysis, it is unclear whether the claimed 4.8 cm and 8.4 cm improvements are robust to the dominant depth-error source. Please add a quantitative sensitivity analysis or explicit covariance modeling.","section":"Section 3.2, Eq. (6), and Supplementary Section 13"},{"comment":"The reported differences between methods are given as mean ± standard deviation over a small number of trajectories, and the standard deviations overlap substantially (e.g., Diff. Eq. 200 Hz at 0.1472 ± 0.0938 m versus Diff. Eq. 30 Hz at 0.1915 ± 0.1705 m). No paired significance tests or effect sizes are reported, so the central claim that 200 Hz updates outperform 30 Hz updates is not statistically established. The authors should add paired tests across the same trajectories or otherwise quantify the uncertainty of the differences.","section":"Section 5.2, Tables 5 and 6"}],"minor_comments":[{"comment":"The contribution bullet says the event-based pipeline reduces error by 4.8 cm compared with frame-based updates, but Table 6 gives 0.1915 − 0.1472 = 4.3 cm for the non-EKF variant and 0.1915 − 0.1432 = 4.8 cm for the EKF variant. Please specify which variant is used so the number is reproducible.","section":"Section 1 and Table 6"},{"comment":"The update frequency for DCGN is listed as '–'; please specify whether DCGN is fed with event-derived noisy states or ground-truth states, and at what rate, since this affects interpretation of its relative performance.","section":"Table 5, DCGN row"},{"comment":"The caption contains the typo 'DCGM'; it should read 'DCGN'.","section":"Supplementary Section 11, Figure 7 caption"},{"comment":"The table lists a horizon of 0.12 s, while the text refers to T = 0.1 s; please harmonize these values.","section":"Supplementary Table 7 and Figure 6"},{"comment":"The sentence 'On overview of the method is displayed in Figure 3' should read 'An overview of the method is displayed in Figure 3'.","section":"Section 3"},{"comment":"The values of the geometric bounds P_min, P_max, A_min, A_max, the number of batches M, the polynomial degree d, and the EKF noise covariances are not reported; providing these would improve reproducibility.","section":"Supplementary Section 8 and Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is already accepted at a CVPR workshop; for an archival journal version, the quantitative claims need the strengthening described above. The small sample size and test-set threshold selection are the main statistical risks, and the 8.4 cm learning-based claim should not appear in its current form. That said, the system demonstration and dataset are valuable, and the central approach is defensible, so major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for the system and the dataset. It is the first egocentric event-based table tennis tracker/predictor, built on Aria glasses with a mounted event camera, and the authors release code and data. The foveated cropping around eye gaze is a sensible idea and clearly cuts latency and event count. The hardware sync and OptiTrack ground truth are real work. Credit where it is due: the engineering is solid and the limitations section is candid about the small ball size and gaze dependency. Now the soft spots. The biggest one is the 8.4 cm claim. The intro says the event pipeline improves a learning-based predictor by 8.4 cm compared to frame-based updates, but that number comes from subtracting Diff Eq at 30 Hz (0.1915 m) from DCGN at 200 Hz (0.1072 m). Those are different models. There is no 30 Hz DCGN baseline anywhere. So the improvement cannot be attributed to update frequency; it is just DCGN being a stronger predictor. The 4.8 cm physics-based comparison is within-model and holds up (0.1915 vs 0.1432). The 8.4 cm line should be retracted or re-run with the same predictor at both rates. Other issues are less damning but real. The detection threshold theta_1 is chosen by looking at Table 3, which is evaluation data; that is tuning on the test set. The 92.59% detection rate is 25 out of 27 sequences with no confidence interval. The 4.5 ms worst-case latency is an ideal theoretical value, not a measured worst case. The depth-from-radius sensitivity in Section 13 is acknowledged but never quantified end-to-end; at 2-3 m, one pixel of radius error is tens of centimeters of depth error. Spin is ignored, which the authors admit and which is acceptable for a first egocentric system. None of these are fatal for a workshop paper. They are fixable, and the dataset plus the egocentric setup are genuinely new. I would send it to peer review if it were not already accepted, because the community gets value from the data and the system description. The letter verdict is conditional: the core idea is sound, the comparison claims need correcting, and the evaluation could be tightened. I would cite this for the dataset, not for the error-reduction numbers.","headline":"A genuinely useful egocentric event-camera system and dataset, but the headline 8.4 cm learning-based improvement is a cross-model comparison that the experiments do not support.","tokens_in":759,"tokens_out":823,"would_cite":true,"duration_ms":30756,"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":"An egocentric event-camera system predicts ping pong trajectories at 200 Hz with a worst-case total latency of 4.5 ms, beating frame-based updates by 4.4 cm at the impact point.","keywords":["event cameras","egocentric vision","table tennis","trajectory prediction","foveated vision","eye gaze","ball detection","latency analysis"],"falsifier":"Record a held-out rally with motion capture, run the published pipeline, and compare the depth estimates from Eq. (6) to the measured 3D positions; the claim would be falsified if the spread of depth residuals from radius fitting is as large as the 4.4-8.4 cm accuracy gaps that the 200 Hz rate is said to create.","tokens_in":16189,"feed_emoji":"🏓","tokens_out":9716,"duration_ms":104659,"temperature":0.7,"pith_summary":"The paper claims that a head-worn event camera, paired with the eye tracker of smart glasses, can detect and track a ping pong ball from the player's own viewpoint at 200 Hz with a worst-case total latency of 4.5 ms. This is presented as the first egocentric, event-based table tennis trajectory prediction system. By cropping event processing to an 80 by 80 window around the gaze reprojection, the pipeline cuts computation by a factor of 10.81 and improves detection from 81.48% to 92.59%. Over a 0.2 s forecasting horizon, the high-rate physics-based predictor reaches 0.1472 m RMSE at the impact point, compared with 0.1915 m for the same predictor fed at 30 Hz, and the learning-based DCGN baseline reaches 0.1072 m. The paper's case is that event cameras overcome the bandwidth-latency trade-off of frame cameras for fast ball sports.","feed_headline":"Ping pong ball tracked at 200 Hz on eye-tracking glasses","feed_subtitle":"Foveated event cropping cuts compute 10.8x and beats 30 FPS cameras by 4.4 cm at impact point","key_machinery":"The load-bearing mechanism is foveated event windowing: events are kept only inside an $80\\times80$ pixel region around the reprojection of the wearer's eye gaze, then ego-motion is compensated with gyroscope data, a normalized timestamp image is thresholded to remove static background, DBSCAN clustering groups the remaining dynamic events, and the cluster whose convex hull best matches a circle (circularity $\\gamma_j = P^2/(4\\pi A)$) is declared the ball. Depth is recovered from the fitted radius by Eq. (6), and the resulting range measurements are regressed with monotonic polynomial fits and fed to a differential-equation predictor with drag; an EKF bootstraps the initial state. The same measurements can instead be fed to a DCGN, a deep conditional generative model that learns the distribution of future trajectories.","core_discovery":"On its own terms, the discovery is that egocentric event-based perception is enough for real-time ball trajectory forecasting: a monocular event camera on the head, using only events in a gaze-centered window, can produce 200 Hz state measurements whose physics-based future impact prediction is more accurate than the same predictor with 30 Hz frame-rate updates. The depth channel comes from the apparent radius of the ball via $Z = f W_{\\mathrm{metric}}/\\hat r$ after fitting a circle to three convex-hull points, and trajectories are propagated with gravity and quadratic drag, optionally bootstrapped by an extended Kalman filter. The system reaches a 92.59% ball detection rate and a 4.5 ms ideal worst-case total latency, and the paper reports that foveated cropping alone reduces the average number of processed events from 7157 to 735 per detection window.","pith_inferences":["Because depth depends on a one-pixel-accurate radius estimate at 2-3 m, the paper's acknowledged but unquantified radius sensitivity may be the real bottleneck; a higher-resolution event sensor or a stereo depth check would likely improve impact-point error more than further event-rate increases.","The gaze-follows-ball assumption means the system will fail when a player briefly looks away; an audio-based opponent-hit trigger, which the paper mentions as future work, could widen or re-center the ROI before the ball is in view.","Neglecting spin is a deliberate simplification; if spin were co-estimated from the event stream, the bounce model and longer-horizon predictions would have a new error source and a new opportunity.","The 4.5 ms latency number is an ideal worst case from CPU-only processing on one laptop platform; deploying on embedded glasses hardware will likely move the computational latency, even if the event-rate advantage remains."],"forward_implications":["At 200 Hz the physics-based predictor reduces impact-point RMSE to 0.1472 m, a 4.4 cm improvement over the 30 Hz version of the same pipeline; with DCGN the improvement is 8.4 cm.","The detection pipeline's worst-case total latency, 4.5 ms including perception and computation, is below the 10-50 ms window cited for human visual reaction, and far below the 66 ms perception-only latency of a 30 FPS camera.","Foveated cropping lowers the number of events processed per detection from 7157 to 735, a 10.81x reduction, while raising detection rate from 81.48% to 92.59%.","Short prediction windows up to 33 ms favor physics-based fitting over DCGN, while longer horizons and online re-estimation favor DCGN, which reaches 0.1072 m in online forecasting."],"supporting_citations":[{"why":"Supplies the ego-motion compensation and timestamp-image thresholding method, plus the low-latency obstacle-detection pipeline this system adapts.","marker":"[7]"},{"why":"Demonstrates event-based agile object catching whose detection pipeline and latency analysis are the template for the ball detector.","marker":"[8]"},{"why":"Provides the DCGN learning-based trajectory predictor used as the strongest comparison baseline in the online forecasting evaluation.","marker":"[16]"},{"why":"Introduces motion-compensated mean timestamp images used to identify moving objects after ego-motion compensation.","marker":"[20]"},{"why":"Establishes the sensing-latency model for event cameras that underlies the 4.5 ms worst-case total latency claim.","marker":"[6]"},{"why":"Supplies the smart glasses with eye gaze, IMU, and synchronized sensor data; without it the egocentric foveated setup would not exist.","marker":"[4]"},{"why":"Represents the prior event-based table tennis perception work that operates in static rather than egocentric setups, defining the gap this paper claims to close.","marker":"[41]"}],"fun_headline_variants":["Foveated event camera on glasses predicts ping pong in 4.5 ms","Gaze-centered events track ball at 200 Hz, beat 30 FPS","Egocentric event vision cuts compute 10.8x for ball tracking","Event-based eye-tracking glasses predict ball trajectory in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a circle fitted to three points on the event cluster accurately gives the ball's apparent radius, which at 2-3 m turns a one-pixel radius error into tens of centimeters of depth error, a sensitivity the paper acknowledges but never quantifies.","fun_headline_variants_meta":{"raw":{"variants":["Foveated event camera on glasses predicts ping pong in 4.5 ms","Gaze-centered events track ball at 200 Hz, beat 30 FPS","Egocentric event vision cuts compute 10.8x for ball tracking","Event-based eye-tracking glasses predict ball trajectory in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1531,"prompt_tokens":982,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":467}},"tokens_in":598,"tokens_out":549,"duration_ms":6908,"temperature":1.0,"reasoning_tokens":467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:25:00.101737+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a held-out rally with motion capture, run the published pipeline, and compare the depth estimates from Eq. (6) to the measured 3D positions; the claim would be falsified if the spread of depth residuals from radius fitting is as large as the 4.4-8.4 cm accuracy gaps that the 200 Hz rate is said to create.","supporting_citations":[{"cited_title":"Dy- namic obstacle avoidance for quadrotors with event cameras","cited_arxiv_id":null,"evidence_quote":"Supplies the ego-motion compensation and timestamp-image thresholding method, plus the low-latency obstacle-detection pipeline this system adapts."},{"cited_title":"Event-based agile object catching with a quadrupedal robot","cited_arxiv_id":null,"evidence_quote":"Demonstrates event-based agile object catching whose detection pipeline and latency analysis are the template for the ball detector."},{"cited_title":"Real time trajectory prediction using deep conditional generative models.IEEE Robotics and Automation Letters, 5(2):970–976, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the DCGN learning-based trajectory predictor used as the strongest comparison baseline in the online forecasting evaluation."},{"cited_title":"Event-based moving object detection and tracking","cited_arxiv_id":null,"evidence_quote":"Introduces motion-compensated mean timestamp images used to identify moving objects after ego-motion compensation."},{"cited_title":"How fast is too fast? the role of perception latency in high- speed sense and avoid.IEEE Robotics and Automation Let- ters, 4(2):1880–1887, 2019","cited_arxiv_id":null,"evidence_quote":"Establishes the sensing-latency model for event cameras that underlies the 4.5 ms worst-case total latency claim."}],"review_version":1}