{"id":"03fc66df-0333-4ed9-b2b1-c0e8eedc0dcf","arxiv_id":"2509.03430","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A headset-integrated system detects touch contact and hover distance on everyday surfaces by analyzing shadows cast by worn infrared emitters, without instrumenting the surface or hands.","lead":"EclipseTouch uses infrared LEDs on a headset to cast shadows of a finger on any surface, then reads those shadows with a camera to tell when the finger is touching or hovering. A compact AI model turns this into a practical way to get touch input on desks, walls, or skin while wearing AR glasses.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Untested fast-motion compositing: 12.5ms frame stacking and ambient-subtraction residuals may invalidate the 98.0% accuracy under realistic swipe speeds.","rationale":"The reader's weakest assumption—that the synchronized frame-subtraction pipeline and 12.5ms compositing hold during motion—is precisely the load-bearing concern I identified. The paper asserts rather than demonstrates the 'essentially simultaneous' nature of the subframes, and the evaluation does not include a worst-case motion analysis. This is a correctness risk for the central claim because fast swipes are a core touch interaction, and errors in the compositing pipeline would directly corrupt the learned classifier's input. The concern is concrete and testable: a speed-binned accuracy study would settle it. I nevertheless recommend keeping the reader's CONDITIONAL verdict (UNCHANGED) because the concern is an untested boundary, not a demonstrated failure, and the paper already includes a substantial evaluation across materials, lighting, and orientations. The reader's other points (aggregation and confounded orientation) are valid but secondary; the temporal compositing assumption is the most fundamental threat to the headline number.","tokens_in":20338,"tokens_out":12431,"duration_ms":140784,"concrete_test":"Collect a focused dataset in which participants perform swipes at a range of peak fingertip speeds (e.g., 0.5, 1.0, 1.5 m/s, measured with an external side camera), then report touch classification accuracy per speed bin using the existing LED 3&4 model. If accuracy drops by more than ~2% from the 98.0% baseline at speeds typical of touch input, the simultaneity assumption is violated and the claim needs qualification. A secondary check: measure the group delay of the 30-frame smoother on annotated touch-onset timestamps; if touch registration is delayed by >100ms, the 98.0% accuracy is not indicative of real-time responsiveness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 98.0% accuracy figure is produced by a pipeline that composites five subframes spanning 12.5ms (Section 3.4) and subtracts the 'no LEDs on' subframe from each illuminated subframe (Section 3.5). This pixel-wise subtraction only cancels ambient light if the ambient shadow pattern is identical in the subtracted frames; during fast hand motion, the finger moves between subframes, so the ambient shadow in the no-LED frame is spatially displaced from the ambient shadow in the LED frames. The subtraction then leaves bright/dark residuals at old and new shadow positions, and the LED shadow channel is misaligned with the fingertip patch center (defined from the LED1 subframe, Section 3.6). The paper asserts that 12.5ms is 'short enough' to stack frames 'as though they were taken at essentially the same moment in time' (Section 3.4), but no hand-speed measurements, worst-case error analysis, or speed-binned accuracy are reported. Since touch interaction includes rapid swipes and taps, the 98.0% figure cannot be assumed to hold for real interaction dynamics. Additionally, the 30-frame mean filter (Section 3.7) smooths over 187-375ms, which may mask per-frame errors but introduces an unquantified latency that is not addressed in the paper's performance claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"EclipseTouch proposes a headset-integrated, active-infrared shadow-casting method for detecting touch contact and estimating hover distance on uninstrumented surfaces. The system captures a five-subframe sequence at 400 FPS (no-LED plus four IR illuminators), subtracts the no-LED frame to suppress ambient shadows, composites the subframes into one 80 FPS multi-channel image, extracts fingertip patches with MediaPipe, and feeds them to a FastViT-based network that jointly outputs touch state and hover distance. The main evaluation uses 10 participants, 1,008,000 frames, 11 non-skin materials plus skin, three lighting levels, and two orientations, with leave-one-participant-out cross-validation. The headline results are a 98.0% touch-segmentation accuracy and a 6.9 mm hover-distance mean error, supported by an ablation of 26 illuminator configurations.","tokens_in":20637,"tokens_out":4907,"duration_ms":51353,"significance":"If the reported results hold, this is a meaningful step toward practical ad hoc touch input for AR/VR headsets: the sensing is entirely headset-worn, requires no per-user or per-surface calibration, works in darkness, and supports on-skin input. The empirical effort is unusually broad for this area, including a large multi-participant dataset, 26 illuminator-geometry ablations, three lighting levels, and both horizontal and vertical orientations. The paper also reports a very fast 0.47 ms inference time, and the authors are commendably explicit about failure modes such as glass, mirrors, very dark IR surfaces, and direct sunlight. The central phenomenon—that structured shadows from headset-mounted LEDs encode fingertip-surface distance—is well motivated and is a credible alternative to depth-camera and RGB-only approaches.","major_comments":[{"comment":"The frame-compositing and ambient-subtraction pipeline assumes the hand is effectively stationary across the 12.5 ms five-subframe firing sequence. Section 3.4 asserts that this duration is 'short enough' to treat the subframes as simultaneous, but no hand-speed measurements, speed-binned accuracy, or worst-case displacement analysis are provided. During a fast swipe, the finger can move many pixels between the no-LED frame and the later LED subframes, leaving subtraction residuals and misaligning the shadow channel with the fingertip patch center. Since the training data includes continuous touch/drag behavior, the 98.0% figure may partially absorb this artifact, but the claim that the method works for realistic interaction dynamics is not yet supported. Please report motion statistics from the study or add a speed-controlled evaluation.","section":"§3.4, §3.5, §4"},{"comment":"The 30-frame mean filter over the 80 FPS output creates a 375 ms smoothing window, which introduces an unquantified output latency of roughly half the window for step changes in touch state. The conclusion claims EclipseTouch runs 'with low latency,' but only the 0.47 ms neural-network inference time is reported. For touch input, event timing is as important as classification accuracy; a threshold crossing delayed by 150–200 ms is noticeable. Please report end-to-end latency (capture, compositing, tracking, inference, smoothing) and the effect of the smoothing window on tap-onset timing.","section":"§3.7, §6/Conclusion"},{"comment":"The lighting and orientation results are based on only two non-skin materials (white wall and patterned wallpaper), not the full set of 12 materials. Section 5.4 reports vertical-orientation accuracy of 95.8% overall and 93.5% for the patterned wallpaper, substantially below the horizontal numbers. The summary statement 'Across all materials (including skin), in both orientations and all three lighting conditions' is therefore stronger than the factorial design supports. I ask the authors to rephrase the claim to describe pooled/segmented coverage, or, ideally, to add at least one more material to the orientation/lighting cells to support the cross-condition claim.","section":"§4, §5.3, §5.4"},{"comment":"The illuminator-geometry ablation uses six participants for training and four for testing, and the best configuration (LED 3 & 4) is then used in the leave-one-participant-out evaluation that includes those same four test participants. This creates a mild selection leak: the choice of geometry is informed by participants who later appear in the final 'unseen user' numbers. The effect is probably small because the configuration is physically motivated, but it should be acknowledged, or the geometry selection should be nested inside cross-validation, to keep the headline accuracy clean.","section":"§5.1, §5"}],"minor_comments":[{"comment":"Typo: 'Finger patches contain contain N channels' duplicates 'contain'.","section":"§3.7"},{"comment":"The text says 'we only use the five fingertip points' but then describes using wrist, MCP, and PIP joints for normalization. Please clarify which keypoints are actually used.","section":"§3.6"},{"comment":"The terms 'Single Channel' and 'Multiple Channel' are not defined in the text. Clarify that 'channel' refers to the number of subframes (LED groups) input to the network as separate image channels.","section":"§5.1 / Figure 8"},{"comment":"The phrase 'none of our materials were fully reflective, like glass' is slightly ambiguous—consider 'fully reflective, such as glass'.","section":"§5.2"},{"comment":"The skin model is trained separately from the surface model; the overall 'across all materials' statement should state this explicitly to avoid implying a single model handles skin and non-skin surfaces.","section":"§5.5"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems paper with an unusually thorough evaluation for the area. The main risks are not about the basic feasibility of IR shadow-casting touch sensing; they are about the strength of the generalization claims under realistic motion and output latency. I would support publication after the motion-speed validation, latency analysis, and rephrasing of the cross-condition claims. The geometry-selection overlap is minor but should be cleaned up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline take: this is a genuine, well-engineered increment in headset-only ad hoc touch sensing, and the evaluation is broad enough that I trust the central result more than the stress-test note suggests. The 98.0% figure is real for the conditions they tested; the open questions are about how far the conditions generalize, not about whether the numbers are manufactured.\n\nWhat is new: past shadow-casting touch systems were stationary (PlayAnywhere, Matsubara) or required a wrist LED (ShadowTouch). EclipseTouch puts the illuminators and camera on the headset, uses synchronized frame subtraction to suppress ambient shadows, and demonstrates the combination across 12 materials, three lighting levels (including dark), and two orientations with leave-one-participant-out evaluation and an ablation of 26 illuminator configurations. The hover-distance and multitouch supplementals are nice bonuses, and the limitations section is candid about glass, mirror, direct sunlight, and skin performance.\n\nThe soft spots are real but not fatal. First, the headline blends per-condition numbers that vary meaningfully: skin 94.7%, vertical patterned wallpaper 93.5%. The aggregate 98.0% is honest but hides those tails. Second, the orientation comparison is thin—only two materials, and the hand-distance confound the authors themselves note. Third, no code or data is released, so the exact number can't be independently checked yet. That matters, but this is typical for UIST hardware prototypes, and the protocol is described in enough detail that replication is conceivable.\n\nOn the stress-test: the fast-motion compositing worry doesn't land as a demonstrated flaw. The 12.5ms stack is short relative to typical hand speeds, and the data collection included continuous drags, not only slow taps. The 30-frame mean filter is a more legitimate concern—at 80 FPS that's nearly 375ms of smoothing, which the paper doesn't discuss in the latency section. But the accuracy numbers may well be computed on the same filtered signal, making it a consistency issue rather than a hidden failure. The right response is a speed-binned experiment in a revision, not a rejection.\n\nWho benefits: anyone working on AR/VR input, wearable sensing, or ad hoc surface interaction. The paper deserves a serious referee and a solid acceptance if the authors address the per-condition reporting and release artifacts.\n\nRecommendation: accept for peer review, with expectations of revision rather than desk reject.","headline":"A well-engineered headset-only shadow-casting touch system whose 98.0% headline is honest for the tested conditions, with soft spots in per-condition reporting, orientation evaluation, and unreleased artifacts; the fast-motion concern is plausible but not a demonstrated flaw.","tokens_in":21142,"tokens_out":2424,"would_cite":true,"duration_ms":25209,"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":"The paper claims that a headset-mounted active infrared shadow-casting system can estimate touch contact and hover distance on uninstrumented surfaces with 98.0% touch accuracy and 6.9 mm mean hover error, with no surface, user, or environm","keywords":["ad hoc touch surfaces","infrared shadow casting","touch segmentation","hover distance estimation","mixed reality input","ambient shadow suppression","egocentric vision","on-skin touch"],"falsifier":"Run EclipseTouch during fast swipes and flicks over an instrumented surface with ground-truth contact, plotting touch accuracy against finger speed; if the 12.5 ms frame stack no longer behaves as one instant, accuracy should fall noticeably below 98% at high speeds.","tokens_in":20201,"feed_emoji":"👆","tokens_out":8393,"duration_ms":78301,"temperature":0.7,"pith_summary":"EclipseTouch is a headset-only way to turn ordinary surfaces into touchpads. The paper claims that by pulsing infrared LEDs around the headset in sync with a camera, the finger casts a structured shadow whose shape encodes both whether the finger is touching and how far it hovers. Subtracting a no-LED frame from each lit frame removes shadows cast by room lights, so the system works in bright, typical, and near-dark lighting. On 12 materials, including skin, in horizontal and vertical orientations, and across three lighting conditions, the reported mean touch accuracy is 98.0% (SD=0.3%), with no per-user or per-surface calibration. The result matters because it suggests existing headsets with IR cameras and illuminators could enable free-form surface touch through a software update.","feed_headline":"Infrared shadows turn any surface into a 98-percent-accurate touchpad","feed_subtitle":"No calibration, no wristband: the headset's own IR LEDs read touch and hover distance even in the dark.","key_machinery":"The mechanism is active infrared shadow casting with synchronized frame subtraction. LEDs offset from the camera cast a shadow whose gap from the fingertip shrinks as the finger nears the surface and vanishes at contact. The load-bearing signal is the difference image: each lit subframe minus the unlit subframe leaves only the headset-cast shadow, stripped of ambient shadows. The model is a 3.3M-parameter hybrid vision transformer built on a backbone the authors call FastViT; it takes fingertip-centered patches from the LED channels plus a finger identity, and outputs one logit that serves both as touch probability and as hover distance.","core_discovery":"The central discovery is that shadow casting, long used for touch sensing with fixed cameras, can be compressed into a single worn device if the shadow source is actively controlled and ambient shadows are subtracted away. Four infrared LEDs fire one at a time in a 2.5 ms sequence while a global-shutter camera snaps a frame per step; the lit frames and one unlit frame are composited into one multi-channel image. Because light adds on the sensor, subtracting the unlit frame from each lit frame cancels ambient shadows, leaving only the headset's own structured shadow. A compact vision transformer on fingertip patches jointly predicts touch state and hover distance, reaching 98.0% touch accurac","pith_inferences":["Editorial inference: the 12.5 ms frame stack is treated as one instant, so fast swipes or head motion should stress the subtraction; measuring accuracy versus finger speed would map where the 98% figure degrades.","Editorial inference: because shadow visibility, not illumination power, seems to drive accuracy (distant LEDs beat the near-inline LED), adding LEDs far from the camera is likely more valuable than brighter LEDs in a commercial design.","Editorial inference: since mirrors and glass are excluded, fusing the shadow channel with fingertip reflection images is a natural path to cover glossy surfaces.","Editorial inference: the paper's synthetic-data plan is plausible because shadow casting is well modeled in graphics engines; a purely synthetic training corpus could replace the million-frame human study and expand material coverage."],"forward_implications":["Existing headsets with two IR illuminators and IR-sensitive cameras could gain ad hoc surface touch through a software update, since a single well-placed LED already reaches 96.7% accuracy.","Touch input no longer requires instrumenting surfaces or wearing rings, wristbands, or fingertip sensors; the headset alone suffices.","Because the model jointly outputs hover distance, systems can offer pre-touch feedback while the finger is still above the surface, unlike depth-camera approaches that need a large lift to separate hover from contact.","The method keeps working in near-total darkness because it brings its own illumination, a regime where headset-camera touch methods previously failed.","Leave-one-participant-out evaluation on unseen users supports the no-calibration claim: the reported 98.0% is not tuned per person or per surface."],"supporting_citations":[{"why":"Closest prior system: a wrist-mounted LED and headset camera for shadow touch, which EclipseTouch must match without a wearable.","marker":"[32]"},{"why":"Introduced shadow shape analysis for finger contact detection, the conceptual foundation this system miniaturizes.","marker":"[64]"},{"why":"The authors' prior headset-camera touch system that fails in darkness; its stated 0% dark accuracy motivates active IR illumination.","marker":"[38]"},{"why":"Depth-camera ad hoc touch system used as a comparison showing why depth sensing is noisy for hover separation.","marker":"[15]"},{"why":"Depth-plus-IR head-mounted touch method used as a comparison for combined sensing approaches.","marker":"[69]"},{"why":"Provides the FastViT backbone that the model reparameterizes to 3.3M parameters for 0.47 ms inference.","marker":"[59]"},{"why":"Provides the hand-tracking keypoints used to center and normalize fingertip patches fed to the model.","marker":"[72]"}],"fun_headline_variants":["Headset's IR shadows read touch on any surface","Subtract ambient shadows, detect touch anywhere","Active IR shadow casting yields 98% touch accuracy","No calibration: worn IR emitters create touch surfaces","Global shutter + IR LEDs = touch detection on any surface"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that subtracting the no-LED image from each lit image leaves only the headset's own shadow; if the hand or head moves during the 12.5 ms capture window, or lighting changes between frames, the subtraction stops lining up and the model receives corrupted input.","fun_headline_variants_meta":{"raw":{"variants":["Headset's IR shadows read touch on any surface","Subtract ambient shadows, detect touch anywhere","Active IR shadow casting yields 98% touch accuracy","No calibration: worn IR emitters create touch surfaces","Global shutter + IR LEDs = touch detection on any surface"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2619,"prompt_tokens":664,"completion_tokens":1955,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":1880}},"tokens_in":408,"tokens_out":1955,"duration_ms":14944,"temperature":1.0,"reasoning_tokens":1880,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:52:55.564962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EclipseTouch during fast swipes and flicks over an instrumented surface with ground-truth contact, plotting touch accuracy against finger speed; if the 12.5 ms frame stack no longer behaves as one instant, accuracy should fall noticeably below 98% at high speeds.","supporting_citations":[{"cited_title":"Krueger, Thomas Gionfriddo, and Katrin Hinrichsen","cited_arxiv_id":null,"evidence_quote":"Closest prior system: a wrist-mounted LED and headset camera for shadow touch, which EclipseTouch must match without a wearable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced shadow shape analysis for finger contact detection, the conceptual foundation this system miniaturizes."},{"cited_title":"Wilson, and Hrvoje Benko","cited_arxiv_id":null,"evidence_quote":"Depth-plus-IR head-mounted touch method used as a comparison for combined sensing approaches."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FastViT backbone that the model reparameterizes to 3.3M parameters for 0.47 ms inference."}],"review_version":1}