{"id":"d35dadd5-6b4b-431a-bce3-a90e86fc6dab","arxiv_id":"2608.06184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Event-driven micro-dynamics from subtle camera motion separate reflection and transmission layers, achieving state-of-the-art reflection removal.","lead":"This paper introduces EvReflection, a network that uses event camera signals, which record tiny brightness changes at microsecond speed, to separate reflections from clean background images. If it works, it gives cameras a new way to remove glass reflections during slight handheld motion, without needing large camera shifts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The well-posedness theorem in Appendix B.2 fails under the paper's own translation-only motion model: ST and SR are proportional, H is singular, and the claimed uniqueness guarantee does not follow from depth difference alone.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the claimed linear independence of ST and SR in Appendix B.2 is asserted, not established, and under the paper's own translation-only model it is false. This concern is central because the paper's framing, both in the abstract and in Section 3.1, is that event signals of micro-dynamics make reflection removal well-posed. If the uniqueness theorem fails, that central theoretical justification collapses. However, the empirical contribution, an event-guided network, a simulation pipeline, and a real-world dataset, could still be valid even without the theorem, so a conditional verdict remains appropriate. The reader already set CONDITIONAL; my stress-test pass does not move that verdict. The secondary issues noted by the reader, such as the under-specified SIR2 event protocol and small real test set, are real but less load-bearing than the mathematical gap, and they are addressable in revision. I agree with the reader's assessment rather than escalating to rejection, because the paper's practical contribution does not logically depend on the flawed proof, provided it is corrected or removed.","tokens_in":21729,"tokens_out":4746,"duration_ms":57378,"concrete_test":"Instantiate a pure-translation micro-shake trajectory V(t) with three or more time samples and depths ZT != ZR. Build the matrix A from Eq. (2) with (vT)_{ti} = (1/ZT) V(ti) dot d and (vR)_{ti} = (1/ZR) V(ti) dot d, then compute det(A^T A). It will be 0 to machine precision, confirming singularity and refuting the stated positive-definiteness claim. As a follow-up, add a depth-independent rotational term omega(t) common to both layers and check whether det(A^T A) becomes positive; if it does, the theorem must be restated with rotation as an explicit modeling assumption rather than relying on depth difference alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of unique solvability hinges on Appendix B.2, Eqs. (31)-(33), where ST and SR are declared linearly independent because ZT != ZR. But the paper's own model in Eq. (31) is optical flow u_k proportional to 1/Z_k for a translating camera. Within a local window, the camera velocity V(t) is common to both layers, so (vT)_{ti} = (V(ti) dot d) / ZT and (vR)_{ti} = (V(ti) dot d) / ZR. Therefore ST = (ZR/ZT) SR exactly: the two columns of A in Eq. (2) are scalar multiples, rank(A) = 1, det(H) = 0, and the least-squares problem has infinitely many minimizers. Pointwise inequality vT(t) != vR(t) does not imply linear independence of the sampled vectors; that is a non sequitur. This is not a disagreement with an external consensus; it is an internal inconsistency between Eq. (31) and the assertion in Eq. (32). The claim in Section 3.1 that motion parallax strictly guarantees positive definiteness is therefore false under the model the paper itself assumes. A common depth-independent rotational flow component could in principle restore rank 2, but the proof never states or uses such an assumption. The empirical architecture may still work, but the advertised theoretical guarantee of well-posedness does not follow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EvReflection, an event-driven reflection removal method that exploits micro-dynamics captured by event cameras. It proposes a Micro-Dynamics Decoupler (MDD) and a Parallax-Attention Rectifier (PAR), together with a parallax-aware simulation pipeline and a new real-world benchmark dataset (EVR2). The paper reports state-of-the-art results on SIR2 (29.09 dB PSNR) and EVR2 (27.25 dB PSNR), and it includes a theoretical claim in Section 3.1 and Appendix B that event-based motion parallax makes the layer-gradient estimation problem well-posed.","tokens_in":22062,"tokens_out":4558,"duration_ms":58017,"significance":"The core idea is genuinely novel and potentially influential: event cameras provide a new cue for reflection removal under subtle handheld motion, and the release of code and a real dataset would support follow-up work. The empirical gains over RGB-only baselines are large, and the ablations in Section 4.4 and Appendix C consistently show that the event modality and the proposed modules help. The explicit limitation section (static scenes, low light, computational cost) is a useful honesty check. However, the advertised theoretical guarantee is not established under the paper's own motion model, and several evaluation choices prevent the empirical claims from being fully trusted as reported.","major_comments":[{"comment":"The proof that H is positive definite asserts that ZT != ZR implies the velocity-magnitude vectors ST and SR are linearly independent. Under the model actually written in Eq. (31), where optical flow for a translating camera is proportional to 1/Z, the two magnitude vectors are exactly proportional: S_T = (Z_R / Z_T) S_R. Therefore the columns of A in Eq. (2) are scalar multiples, rank(A) = 1, H is singular, and the least-squares problem has infinitely many minimizers. Pointwise inequality v_T(t) != v_R(t) does not imply linear independence of the sampled vectors, so the claimed uniqueness guarantee does not follow. The authors need to either state an explicit additional assumption (for example, a depth-independent rotational flow component, or a time-varying depth ratio) and prove independence under that assumption, or temper the well-posedness claim in Section 3.1.","section":"Section 3.1 and Appendix B.2, Eqs. (31)-(33)"},{"comment":"The simulation pipeline in Section 4.1 states that images from both PASCAL VOC and SIR2 are used as source layers, while Section 4.3 says the model is trained exclusively on 7,643 pairs from PASCAL VOC. If any SIR2 images are used to generate synthetic training pairs, then evaluating on the SIR2 benchmark (Table 1) is contaminated by training data leakage. Please clarify the exact composition of the training set and confirm that no SIR2 test image, or close variant, appears during training.","section":"Section 4.1 vs Section 4.3"},{"comment":"The paper does not specify how event streams are generated for the static SIR2 benchmark images. Because the method is trained on events synthesized with the same parallax-aware pipeline, the SIR2 numbers could reflect a simulator-specific parallax cue rather than a robust property of real event streams. Please provide the full protocol: virtual depths for transmission and reflection, the camera trajectory C(t), glass thickness or offset parameters, event thresholds, and the exact set of source images used to create each test event stream. Ideally, also evaluate on events captured by a real event camera for at least a subset of SIR2-style scenes.","section":"Section 4.1, SIR2 evaluation protocol"},{"comment":"The EVR2 test set is selected by a 'performance-guided stratified sampling' strategy that keeps only 11 of 140 scenes, chosen by mapping the error distribution of a baseline model. This procedure is vulnerable to selection bias unless the rule is fixed before seeing the proposed model's outputs and is applied only to a development partition. Please report the full 140-scene aggregate results or a random, pre-registered split, and specify the selection criterion in enough detail that the benchmark is reproducible.","section":"Section 4.1, EVR2 test-set construction"},{"comment":"The real-world comparison treats baselines inconsistently: methods marked with an asterisk are evaluated with official pre-trained models, while unmarked methods are retrained on EVR2. Differences in training data, resolution, and augmentation make the reported margin over the asterisked baselines difficult to interpret. Please retrain all baselines under the same protocol, or clearly justify why official weights are the appropriate comparison, and report which training split and preprocessing each baseline used.","section":"Table 2, baseline retraining consistency"}],"minor_comments":[{"comment":"The text alternates between 'EVR2' and 'EVR 2' (with a space), and Table 2 headers such as 'EVR2-T3' and 'EVR 2-T5' are inconsistent; please normalize the notation throughout.","section":"Section 4.1 and Table 2"},{"comment":"The notation 'I · E' in Eq. (1) is ambiguous because I is a scalar intensity and E is defined as a continuous event signal; please clarify whether this denotes pointwise multiplication and define the domain of each quantity explicitly.","section":"Section 3.1, Eq. (1)"},{"comment":"The paper says the total loss 'strictly adheres' to RDNet's objective but does not give the weights of the MSE, gradient, and perceptual terms; please provide the exact loss formula and weights for reproducibility.","section":"Section 4.2, loss function"},{"comment":"The MDD iteration count is listed as M=12 in Table F.1 but the main text in Section 3.3 only says 'N iterations' and never defines N; please reconcile the notation and state the value used in the experiments.","section":"Appendix F, Table F.1"},{"comment":"The limitation paragraph correctly notes that the method struggles in strictly static scenes, but the Introduction and Abstract should be rephrased so that the need for micro-dynamics is stated up front rather than only in the final limitation section.","section":"Conclusion, Limitation paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good empirical fit for the venue and the dataset/code release is valuable, but the theoretical well-posedness proof is internally inconsistent with the paper's own translation-only flow model, and the SIR2/EVR2 evaluation choices raise contamination and fairness concerns. I would not recommend rejection if the authors can correct or qualify the theorem and tighten the evaluation protocol, but the current claims outrun the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is the first paper to use event cameras for reflection removal, and the empirical gains are plausible and large. But the advertised theoretical guarantee of unique solvability is false under the paper's own motion model, and the evaluation has several gaps that need closing before the numbers can be trusted.\n\nWhat's genuinely new: the task formulation (micro-dynamics from events as a cue for layer separation), the MDD+PAR architecture, the parallax-aware simulation pipeline, and the EVR2 real dataset. The authors report +1.63 dB on SIR2 and +1.28 dB on EVR2 over strong RGB baselines, with ablations showing the event branch is worth ~1 dB and the attention fusion another ~0.2 dB. That is a meaningful result if it holds. They also ship code, dataset, and pretrained models, which is exactly what a subfield needs.\n\nThe soft spots are real, in increasing order:\n\n- The SIR2 event protocol is unspecified. SIR2 images are static; you need to know how the event streams were generated (which simulator settings, which motion model). Easy to fix, but essential for reproducibility.\n- The EVR2 test set is only 11 scenes, and they were chosen by a \"performance-guided\" strategy based on a baseline's error map. That looks like selecting a favorable test split. A random or fixed split of more scenes would be much more convincing.\n- Baseline comparisons are uneven: asterisked methods use official pretrained models while others are retrained on EVR2. That stacks the deck, especially for DAI*, which drops about 6 dB from its SIR2 behavior.\n- Most seriously, the theoretical proof in Appendix B.2 does not go through. Equation (31) says optical flow is inversely proportional to depth for a translating camera. Under that model, with a common camera velocity, the two columns of A are scalar multiples (S_T = (Z_R/Z_T) S_R), so rank(A)=1 and H is singular. Pointwise v_T(t) != v_R(t) does not imply linear independence of the sampled vectors. The claim that depth difference alone guarantees a positive definite H is false. A rotational flow component could fix the rank, but that is not the model used in Section 3.1.\n\nThe circularity concern about the simulator encoding the same parallax assumption is minor; the method is supervised and evaluated on real data.\n\nWho should read this? Anyone working on reflection removal or event-based vision. It deserves a serious referee, but the referee should ask for major revision: fix or retract the theorem, disclose event generation for SIR2, enlarge and randomize the test split, and retrain all baselines under one protocol. The empirical core might survive; the theory as written will not.\n\nMy recommendation: send it to review, with a clear request for those revisions.","headline":"First event-camera reflection removal paper with a promising new dataset and large claimed gains, but the theoretical uniqueness proof is internally inconsistent and the evaluation has fixable protocol gaps.","tokens_in":22559,"tokens_out":5371,"would_cite":true,"duration_ms":62584,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Event cameras break the reflection-removal ambiguity by exploiting depth-dependent micro-motion","keywords":["event camera reflection removal","motion parallax","micro-dynamics","layer separation","event-driven network","EVR2 dataset","parallax-aware simulation","cross-modal attention"],"falsifier":"The cleanest check is to re-derive the velocity sequences under the paper's own pure-translation model: if the two depth layers have speeds that differ only by a constant factor, the matrix A has proportional columns and the Hessian is singular, which would contradict the claimed positive-definiteness. An experimental falsifier is to record EVR2-style scenes with a purely translational slide (no rotation or speed variation) and measure whether EvReflection's margin over RGB baselines shrinks or vanishes, since the theory predicts the motion cue should then be uninformative.","tokens_in":21522,"feed_emoji":"","tokens_out":3800,"duration_ms":36066,"temperature":0.7,"pith_summary":"The paper claims that reflection removal, which is ill-posed for a single image because infinitely many transmission/reflection pairs sum to the same observation, becomes well-posed when the camera records event signals during subtle handheld motion. Because the glass reflection and the background behind it sit at different depths, a tiny camera shake moves the two layers differently, and event cameras capture that differential motion at microsecond resolution. The paper builds a network, EvReflection, that turns event streams into layer-specific motion priors and uses those priors to suppress reflection artifacts in an RGB image, reporting state-of-the-art results on both the synthetic SIR2 benchmark (29.09 dB) and a new real-world EVR2 dataset (27.25 dB). If these results hold, event cameras supply a physically grounded cue that lets reflection removal work under casual handheld motion instead of requiring large camera baselines or strong static-image priors.","feed_headline":"Event camera shake makes reflections removable","feed_subtitle":"Subtle handheld motion separates transmission from reflection, beating RGB methods by over 1.2–1.6 dB.","key_machinery":"The load-bearing object is the linear system A x = b built from event observations, where the rows of A record the scalar velocity magnitudes of the transmission and reflection layers at successive instants inside a micro-second window. The claim that this system has a unique solution rests on the Hessian H = A^T A being positive definite, which the paper equates with the two layers' motion magnitude vectors S_T and S_R being linearly independent, i.e., with real motion parallax v_T(t) != v_R(t). On the network side, the carrying mechanism is the Micro-Dynamics Decoupler, a dual-branch recurrent module that iteratively separates the mixed event stream into distinct transmission and reflection dynamics priors, and the Parallax-Attention Rectifier, a cross-modal attention block that uses those priors as query keys to decide which spatial regions of the RGB feature map belong to transmission and which to reflection.","core_discovery":"The central claim is that event signals of micro-dynamics make reflection removal well-posed. The paper models the observed image as a sum of transmission and reflection layers, relates the event stream to the temporal derivative of log intensity, and derives an Event-Gradient Constraint: the measured signal equals the negative gradient of each layer projected along its own motion. Aggregating several event observations within a short window yields a linear system Ax=b whose Hessian A^T A, the paper argues, is positive definite when the two layers move at different speeds. The paper states that distinct micro-dynamics guarantee a unique solution, resolving the ambiguity that plagues single-image methods. It then instantiates this theory as EvReflection, whose Micro-Dynamics Decoupler (MDD) separates the event stream into transmission and reflection motion priors and whose Parallax-Attention Rectifier (PAR) uses those priors to spatially modulate RGB features and remove artifacts. The empirical claim is a gain of 1.63 dB over the best RGB competitor on SIR2 (29.09 versus 27.46 dB) and 1.28 dB on the real-world EVR2 dataset (27.25 versus 25.97 dB).","pith_inferences":["The paper's own motion model for a translating camera makes the linear-independence argument questionable: when flow magnitude scales as 1/Z under pure translation, the two layers' speed vectors differ only by a constant ratio, so they are proportional and the Hessian is singular. The unique-solution claim therefore needs the camera to rotate or the trajectory to vary in a way that produces genuin","A directly testable extension would be a controlled experiment in pure lateral translation on the EVR2 rig: if the theory's necessary condition fails under pure translation, performance should drop measurably compared with trajectories containing rotation, even though depth differences are present.","The method seems likely to transfer to event-guided de-raining, de-hazing, or glare suppression, where the offending layer occupies a distinct depth plane, but those tasks would need their own simulators and benchmarks because reflection has the special property that the virtual layer lies behind the glass.","The ablation numbers suggest that motion priors matter most on thick glass, so future work could specialize the attention mechanism to the larger displacement regimes rather than treating all thicknesses uniformly."],"forward_implications":["Handheld devices with event-RGB hybrid sensors could remove reflections from casual photographs without requiring the user to move the camera deliberately or capture multiple views with a large baseline.","The EVR2 dataset, the first real-world event-based reflection removal benchmark, gives the field a common test bed stratified by glass thickness, and the parallax-aware simulator offers a way to generate paired training data for other event-guided layer-separation tasks.","Reflection removal performance degrades gracefully as glass thickens: the method still leads competitors on 8 mm glass, where double-reflection ghosting is strongest.","If the theoretical argument is correct, similar event-gradient constraints could be applied to other additive superposition problems, such as removing rain, snow, or glare, where the degradations lie at different depths than the scene."],"supporting_citations":[{"why":"Supplies the event camera model: events trigger on log-intensity change, giving the asynchronous brightness-change signal the method exploits.","marker":"(Gallego et al., 2020)"},{"why":"Provides the brightness constancy / optical flow constraint that the paper substitutes into the event model to obtain the Event-Gradient Constraint.","marker":"(Horn & Schunck, 1981)"},{"why":"Earlier event-based flow work that connects events to spatial gradients, supporting the event-gradient formulation.","marker":"(Benosman et al., 2013)"},{"why":"ESIM, the event simulator used to convert the parallax-warped synthetic videos into event streams for training.","marker":"(Rebecq et al., 2018)"},{"why":"Supplies the voxel-grid event representation the paper refines to 20 temporal bins.","marker":"(Zhu et al., 2019)"},{"why":"RDNet is the strongest RGB baseline and the source of the strictly-adhered-to loss objective, so it sets the baseline the paper must beat.","marker":"(Zhao et al., 2025)"},{"why":"SIR2 is the standard synthetic benchmark on which EvReflection reports its 29.09 dB result.","marker":"(Wan et al., 2017)"},{"why":"FocalNet is the pre-trained context encoder backbone that supplies the RGB semantic features for the rectifier.","marker":"(Yang et al., 2022)"},{"why":"DAI is the latest single-image competitor whose 27.46 dB average the paper surpasses by 1.63 dB.","marker":"(Hu et al., 2026)"},{"why":"The RAFT architecture is adapted as the motion extractor inside the Micro-Dynamics Decoupler.","marker":"(Teed & Deng, 2020)"}],"fun_headline_variants":["Event cameras exploit micro-motion to remove reflections","Micro-dynamics from events break reflection ambiguity","Event-driven motion cues peel reflections off images","Subtle motion from events clears reflections, +1.6 dB","Event signals reveal layer motion for reflection removal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument depends on the claim that subtle camera shake makes the reflection and transmission layers move at different speeds over time, so the measured event signals can pick out each layer's motion; if the two layers' motions are proportional or too weak to detect, the uniqueness argument collapses and the event cue degenerates.","fun_headline_variants_meta":{"raw":{"variants":["Event cameras exploit micro-motion to remove reflections","Micro-dynamics from events break reflection ambiguity","Event-driven motion cues peel reflections off images","Subtle motion from events clears reflections, +1.6 dB","Event signals reveal layer motion for reflection removal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1582,"prompt_tokens":1006,"completion_tokens":576,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":504}},"tokens_in":622,"tokens_out":576,"duration_ms":7620,"temperature":1.0,"reasoning_tokens":504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:03:45.641912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The cleanest check is to re-derive the velocity sequences under the paper's own pure-translation model: if the two depth layers have speeds that differ only by a constant factor, the matrix A has proportional columns and the Hessian is singular, which would contradict the claimed positive-definiteness. An experimental falsifier is to record EVR2-style scenes with a purely translational slide (no rotation or speed variation) and measure whether EvReflection's margin over RGB baselines shrinks or vanishes, since the theory predicts the motion cue should then be uninformative.","supporting_citations":[],"review_version":1}