{"id":"66d4dbb8-bc02-4786-a7e6-58806333dd78","arxiv_id":"2508.19094","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A vibrating event camera with an unbalanced mass generates persistent events in static scenes, and an EKF-based pipeline removes the induced motion to produce clean, usable event streams.","lead":"The authors attached a small off-center weight to an event camera so it vibrates like a phone, making it produce events even in a static scene, then used a filter to remove the vibration and recover clean event streams. This offers a low-power way to keep event-based vision working in low-motion conditions, which could matter for robots and drones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4 omits the VIBES rows that are claimed to demonstrate improved edge continuity; the edge-detection half of the central claim is unsupported by the presented data.","rationale":"The reader's conditional verdict is well-founded, and I agree that CONDITIONAL is appropriate. Among the candidate concerns, the missing VIBES rows in Table 4 is the most load-bearing because the central claim explicitly includes \"improves ... edge detection.\" The only quantitative evidence for edge-continuity improvement is Table 4, yet the table lists VIBES nowhere. The text surrounding the table makes a positive assertion that is not backed by the displayed numbers. This is not a peripheral formatting issue: it is the absence of the exact comparison needed to support the claim. The sinusoidal/depth-plane assumptions the reader highlights are real limitations, but they affect generalization beyond the demonstrated settings; within the near-field planar scenes tested, they may hold. The table omission is a presentational gap that can be resolved by supplying the missing data. If the missing rows agree with the text, the concern is reduced to a copy-paste error; if they disagree, the central claim is weakened. Thus the most decisive check is a public release of the edge statistics or a reproducible pipeline that regenerates them.","tokens_in":17020,"tokens_out":13513,"duration_ms":142322,"concrete_test":"Request or reconstruct the missing VIBES rows in Table 4. The procedure is specified in Sec. 4.2: accumulate events over a 33 ms window per status (S-EV, V-EV, VIBES), smooth with a Gaussian, binarize, thin with Zhang–Suen, then extract connected components, average contour length, and junction count. For each of the four scenes, compute these three metrics on the VIBES stream and compare to the S-EV rows already present. If VIBES does not show a decrease in average number of components and an increase in average contour length relative to S-EV across most scenes, the edge-detection claim is not supported. Ideally, publish the raw tables or code so the numbers can be independently recomputed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. 6) states that VIBES yields motion-compensated events that improve edge detection. The quantitative support for this is Table 4 (Sec. 4.4). However, Table 4 contains only S-EV and V-EV rows for every scene; no VIBES rows are present. The caption and text (\"Across most scenes, VIBES reduces edge fragmentation...\") assert the VIBES comparison, but the numbers are absent. Additionally, the text in Sec. 4.4 says \"VIBES produces clean, well-connected edge maps\", but without VIBES rows, no such comparison can be verified. This is a missing support for a core contribution, not a peripheral issue. If the missing rows do not reproduce the claimed improvement, the central claim of improved edge detection fails. The concern is internal to the paper: the presentation omits the very data that would demonstrate the claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VibES, a hardware-software system that mounts a small eccentric rotating mass on an event camera to induce controlled sinusoidal vibration, thereby generating persistent events even in static scenes. The induced motion is modeled as a damped harmonic oscillator, and a pipeline (HASTE tracker, NUFFT frequency initialization, and an EKF) estimates the sinusoidal motion parameters directly from the event stream. The estimated motion is then used to compensate the events, yielding a motion-stabilized event stream. The authors claim improvements in event density, image reconstruction quality, edge detection, scene-frequency estimation, and relative depth awareness, and demonstrate the approach on real-world datasets (including AMI-EV) and synthetic data.","tokens_in":17238,"tokens_out":4886,"duration_ms":46699,"significance":"If the claims are sustained, the work offers a practical, low-power, hardware-simple answer to the well-known 'perceptual fading' problem of event cameras, with a fully model-based, online motion-compensation pipeline that does not require external position encoders. The paper discloses mechanical design parameters, power consumption, and runtime, and it demonstrates hardware independence by reusing the algorithm on the AMI-EV rotating-wedge-prism dataset. The real-world measured entropy, NIQE, variance, and gradient-magnitude improvements over a static event camera are notable. The main weakness is that the quantitative support for the edge-continuity part of the central claim is incomplete, as detailed below.","major_comments":[{"comment":"Table 4 is the only place where edge continuity/fragmentation metrics are reported, yet every row is labeled S-EV or V-EV; there is no VIBES row. The caption states 'Across most scenes, VIBES reduces edge fragmentation and increases contour continuity compared to S-EV,' and the text says 'VIBES produces clean, well-connected edge maps.' These claims are not supported by the presented numbers. The missing rows are load-bearing for the 'improves edge detection' component of the central claim. Please add the VIBES rows, or if the data are not available, restrict the claim to the variance/gradient-magnitude metrics of Table 3 and remove the continuity/fragmentation statements.","section":"Sec. 4.4, Table 4"},{"comment":"The metric definition says 'The average number of components measures the number of components in an edge map, with a greater number showing less fragmented and noisy edges.' This is internally contradictory: more connected components in a thinned edge map typically indicates more fragmentation, not less. Table 4's caption says 'number of connected components (lower is better),' which contradicts the text. Please correct the direction of the metric and ensure the text and table caption agree.","section":"Sec. 4.2, 'Edges continuity and fragmentation' metric"},{"comment":"The predicted depth ratios are 0.41±0.01, 0.59±0.02, and 0.80±0.03 for ground-truth ratios 0.33, 0.50, and 0.66. This is a systematic overestimation of roughly 20% relative error. Calling this 'reliably predicts' is an overstatement without discussing the bias. Please report relative errors and discuss possible sources (e.g., amplitude estimation bias, non-ideal sinusoidal motion, tracker noise) and whether this is acceptable for the intended application.","section":"Sec. 5.2, Relative Depth Awareness"},{"comment":"The paper claims compensation 'across the entire scene' using the EKF-tracked region, but Sec. 3.2.3 states the assumption that all objects in the same scene share one depth plane. Real scenes contain depth variation, and the amplitude of the induced motion depends on depth (Eq. 5). Although the implementation allows multiple trackers, the paper does not describe how multiple trackers are fused or how a spatially varying amplitude map is obtained for whole-image compensation. This limits the general-scene claim. Please either describe the fusion/interpolation scheme or explicitly state the single-depth limitation in the main text and temper the 'across the entire scene' phrasing.","section":"Sec. 3.2.3 and Fig. 1"}],"minor_comments":[{"comment":"The phrase 'with a greater number showing less fragmented and noisy edges' is confusing and should be rewritten once the direction is corrected.","section":"Sec. 4.2"},{"comment":"The text says higher frequencies are affected by aliasing, but Table 5 shows the smallest absolute error at 22.6 Hz. Please clarify how aliasing manifests and why the error does not monotonically increase.","section":"Sec. 5.1"}],"recommendation":"major_revision","confidential_remarks":"The omission of VIBES rows from Table 4 is the key blocking issue; it directly concerns the paper's headline claim of improved edge detection. The cause is unclear, but it must be corrected before the paper can be considered further. The depth-prediction results also show a systematic bias that the authors should address. I see no problems with novelty or citation practices."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the mechanism is genuinely new and appealingly low-tech: glue an off-axis mass to a hobby motor on an event camera, let a foam suspension produce quasi-sinusoidal motion, and compensate it online from the events themselves. Second, one of the loaded claims—that VIBES improves edge continuity—is unsupported by the data as printed.\n\nWhat is new: the specific hardware configuration (rotating unbalanced mass plus foam spring-damper) is new relative to pan-tilt, wedge-prism, and mirror-based stimulation. The compensation pipeline is a sensible assembly of HASTE, NUFFT, and an EKF with a phase-less sinusoidal parametrization; it runs in real time and generalizes to the AMI-EV wedge-prism data, which is a nice hardware-agnostic check. The real-world metrics for entropy, NIQE, variance, and gradient magnitude all point in the right direction, and the reported 65 Mev/s processing speed is concrete.\n\nSoft spots. Table 4, which is supposed to show edge fragmentation and continuity, contains only S-EV and V-EV rows for every scene; there are no VIBES rows. The caption and text claim VIBES reduces fragmentation and increases continuity, but the numbers are absent. That is missing support for a core contribution, not a quibble. The synthetic depth ratios are also consistently off: ground truth 0.33, 0.50, 0.66 map to 0.41, 0.59, 0.80. The text calls these reliable, but the systematic bias suggests the model or simulation has an unaccounted effect. It is probably a correctable offset rather than a refutation of the relative-depth concept, but it needs explanation. The assumptions of globally constant frequency and per-tracker constant depth are reasonable for a prototype but should be stated as limits on generalization. Trackers are manually initialized, which is fine for a prototype but makes the \"persistent\" claim a bit softer.\n\nWho this is for: people working on event-driven perception in static scenes, microsaccade-inspired hardware, and online motion compensation. It is a solid within-subfield contribution, not a paradigm shift. With the missing VIBES rows fixed and the depth bias explained, I would have no problem with it.\n\nRecommendation: worth sending to peer review. A serious referee should ask for the missing rows, raw event data or code, and an explanation of the depth bias.","headline":"The core idea is sound and the hardware is refreshingly simple; the paper deserves a serious referee, but the edge-detection claim currently rests on a table that does not contain VIBES numbers, and the depth results have a systematic bias the text glosses over.","tokens_in":17737,"tokens_out":2197,"would_cite":true,"duration_ms":21763,"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":"Vibrating eccentric motor sustains event streams in static scenes, and an online filter removes the vibration to recover clean events.","keywords":["event cameras","induced vibration","motion compensation","extended Kalman filter","microsaccade-inspired sensing","image reconstruction","edge detection","relative depth estimation"],"falsifier":"Take the vibrating prototype and record its actual trajectory with an external high-speed camera or optical tracker while events are being captured; fit the best single sinusoid of the paper's model to that trajectory and check the residual in image coordinates over one oscillation period. If the residual exceeds about one pixel for targets within the paper's stated 47.3 cm sensing range, motion compensation will leave systematic ghosting or blur on reconstructed edges. A second decisive test: place two textured patterns at clearly different depths and run the pipeline with a single tracker; t","tokens_in":16917,"feed_emoji":"📳","tokens_out":6659,"duration_ms":60255,"temperature":0.7,"pith_summary":"Event cameras measure only intensity changes, so a rigidly mounted camera in a static scene stops producing events and the visual signal fades. VIBES attaches a small rotating unbalanced mass to the camera body, creating forced harmonic vibration that keeps pixels firing the way tiny involuntary eye movements keep human vision refreshed. The paper's central claim is that this injected motion is a predictable single-frequency sinusoid whose parameters can be estimated online from the event stream alone, then subtracted to recover the static scene. If that claim holds, event-based perception can run persistently in static or slow scenes, and the same estimated motion parameters double as cues for scene-object vibration frequency and relative depth. The authors build a foam-suspended prototype and report, across four real-world datasets, higher event density, sharper accumulated frames, more connected edges, and better reconstructed images than a non-vibrating event camera.","feed_headline":"Vibrating motor keeps event cameras awake in static scenes","feed_subtitle":"A cheap shaker motor and an online motion filter restore sharp events and edges where static event cameras go silent.","key_machinery":"Central machinery is the sinusoidal camera-motion model v' = v + A_y cos(omega t + phi_y), which treats induced vibration as planar circular translation of the real camera around an imaginary static camera. Everything else hangs on it: the model makes the injected motion a sparse, predictable signal, lets the non-uniform FFT recover its frequency from irregular event timestamps, gives the Extended Kalman Filter a stable state (theta, omega, a_y, b_y, v) that avoids phase wrapping, and makes the per-track amplitude A_y proportional to f*Y0/Z so amplitude ratios across trackers map directly to relative depth. The EKF parametrization a_y sin theta + b_y cos theta + v keeps the observation model","core_discovery":"The paper claims that a consumer DC motor with an off-axis mass can replace bulky optical or pan-tilt hardware for microsaccade-style event stimulation. The motion is modeled as v' = v + A_y cos(omega t + phi_y): the physical camera circulates around a virtual static camera, so every tracked point moves on a sinusoid whose amplitude encodes scene depth. The pipeline — HASTE tracker, NUFFT frequency initialization, and an Extended Kalman Filter that tracks angular position, frequency, sine/cosine amplitudes, and the compensated coordinate — estimates and removes that motion in real time (~65 Mev/s) without position encoders or calibration. On their own three scenes plus the public AMI-EV data","pith_inferences":["If the sinusoidal assumption is robust, the shaker could be made adaptive: turn it off when the scene is already dynamic and re-engage when event rate drops, preserving the camera's low-power advantage.","Because the amplitude formula A = fY0/Z ties perceived motion to depth, the induced vibration effectively creates a virtual stereo baseline; with the physical vibration amplitude calibrated once, the same estimator could yield absolute depth, not just ratios.","The single-depth-plane assumption per tracker is the main obstacle for arbitrary scenes; a natural extension is per-pixel or depth-segmented EKF instances, which the paper's multi-tracker design already allows.","The hardware is cheap and compact enough that vibration-based event stimulation could become an integrated feature of commodity event camera modules; the open question is whether the foam suspension remains a clean single sinusoid at higher amplitudes and over long-term wear."],"forward_implications":["Static or quasi-static scenes no longer starve event cameras of data; the paper's measurements show entropy stays high and stable while a non-vibrating camera's entropy drops.","Downstream event algorithms — image reconstruction and edge detection in the paper; tracking and SLAM by extension — receive a motion-corrected stream rather than raw vibration, so they can treat the camera as approximately static.","The compensation stack is hardware-agnostic: it runs on the authors' motor prototype and on the public AMI-EV wedge-prism data with comparable quality, suggesting one software pipeline can serve different shaker designs.","The same frequency estimator can double as a vibration sensor for scene objects, with measured errors below 0.1 Hz for 7.5-22.5 Hz targets.","Multiple EKF trackers yield amplitude ratios that predict relative depth (predicted 0.41/0.59/0.80 for true 0.33/0.50/0.66 in simulation), opening a parallax-based depth cue."],"supporting_citations":[{"why":"Supplies the closest prior system (rotating wedge prism), the public AMI-EV dataset for cross-hardware validation, and the main comparison for motion-compensation quality.","marker":"[24]"},{"why":"HASTE event tracker extracts the centroid trajectories that feed frequency and EKF estimation, so the pipeline depends on it.","marker":"[1]"},{"why":"NUFFT library estimates the dominant oscillation frequency from asynchronously sampled event trajectories, initializing the EKF.","marker":"[4]"},{"why":"E2VID is used to reconstruct images from 10 ms event windows; its outputs are scored with NIQE to compare VIBES against static sensing.","marker":"[38]"},{"why":"Supplies the variance-based sharpness rationale and contrast-maximization context used to justify the accumulated-frame metrics.","marker":"[18]"},{"why":"The DC hobby motor is the actual vibration source in the hardware prototype; its parameters feed the motion model.","marker":"[14]"},{"why":"V2E converts synthetic Blender frames to realistic events for the relative-depth validation experiments.","marker":"[9]"},{"why":"Defines NIQE, the no-reference image quality metric used to evaluate reconstructed frames.","marker":"[33]"}],"fun_headline_variants":["Shaker motor rouses dormant event cameras","Cheap vibration keeps event cameras firing in static scenes","Vibrating mass sustains event camera output, filter cleans it","Low-cost motor induces motion for event cameras, then compensates","Event cameras stay alert via tiny vibrating motor"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the induced vibration is a single sinusoid with one global frequency and phase, and that each tracker sees one depth plane; real foam-suspended motor vibration and multi-depth scenes violate both.","fun_headline_variants_meta":{"raw":{"variants":["Shaker motor rouses dormant event cameras","Cheap vibration keeps event cameras firing in static scenes","Vibrating mass sustains event camera output, filter cleans it","Low-cost motor induces motion for event cameras, then compensates","Event cameras stay alert via tiny vibrating motor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":9.4e-05,"raw_usage":{"total_tokens":799,"prompt_tokens":676,"completion_tokens":123,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":60}},"tokens_in":420,"tokens_out":123,"duration_ms":2875,"temperature":1.0,"reasoning_tokens":60,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:56:48.940747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the vibrating prototype and record its actual trajectory with an external high-speed camera or optical tracker while events are being captured; fit the best single sinusoid of the paper's model to that trajectory and check the residual in image coordinates over one oscillation period. If the residual exceeds about one pixel for targets within the paper's stated 47.3 cm sensing range, motion compensation will leave systematic ghosting or blur on reconstructed edges. A second decisive test: place two textured patterns at clearly different depths and run the pipeline with a single tracker; t","supporting_citations":[{"cited_title":"Microsaccade-inspired event camera for robotics","cited_arxiv_id":null,"evidence_quote":"Supplies the closest prior system (rotating wedge prism), the public AMI-EV dataset for cross-hardware validation, and the main comparison for motion-compensation quality."},{"cited_title":"Alzugaray and M","cited_arxiv_id":null,"evidence_quote":"HASTE event tracker extracts the centroid trajectories that feed frequency and EKF estimation, so the pipeline depends on it."},{"cited_title":"exponential of semicircle","cited_arxiv_id":null,"evidence_quote":"NUFFT library estimates the dominant oscillation frequency from asynchronously sampled event trajectories, initializing the EKF."},{"cited_title":"Events-to-video: Bringing modern computer vision to event cameras","cited_arxiv_id":null,"evidence_quote":"E2VID is used to reconstruct images from 10 ms event windows; its outputs are scored with NIQE to compare VIBES against static sensing."},{"cited_title":"A unifying contrast maximization framework for event cam- eras, with applications to motion, depth, and optical flow es- timation","cited_arxiv_id":null,"evidence_quote":"Supplies the variance-based sharpness rationale and contrast-maximization context used to justify the accumulated-frame metrics."},{"cited_title":"Geared hobby motor","cited_arxiv_id":null,"evidence_quote":"The DC hobby motor is the actual vibration source in the hardware prototype; its parameters feed the motion model."},{"cited_title":"V2e: From video frames to realistic dvs event camera streams","cited_arxiv_id":null,"evidence_quote":"V2E converts synthetic Blender frames to realistic events for the relative-depth validation experiments."},{"cited_title":"completely blind","cited_arxiv_id":null,"evidence_quote":"Defines NIQE, the no-reference image quality metric used to evaluate reconstructed frames."}],"review_version":1}