{"id":"fbad62bc-83ae-4571-8a03-86d026c8a496","arxiv_id":"2501.03874","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An event camera plus a two-module spiking neural network tracks and reconstructs MNIST and Kanji characters hidden behind strongly scattering media in benchtop experiments.","lead":"This paper combines an event camera with a spiking neural network to track and reconstruct images of moving hidden objects behind a dense scattering material. A generalist reader may care because the approach promises low-latency, low-power imaging through fog, tissue, or turbid media.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The demonstration rests on a static scattering medium; any temporal variation in the phantom itself would flood the DVS with events and likely defeat the learned SNN, so generality to realistic dynamic scattering media is unproven.","rationale":"I agree with the reader's weakest_assumption and would make dynamic scattering the primary condition for acceptance. This is not an internal inconsistency: the paper is explicit about the static medium. But the strength of the claimed advance — tracking and imaging through strongly scattering media — is precisely the regime where the medium can move, as in fog, tissue, or turbid water. The energy-efficiency and real-time claims also depend on sparsity; a fluctuating medium destroys sparsity. The reported 0.9577 SSIM and 15 ms inference are only for the static phantom. No error bars or tracking metric are given, so we cannot distinguish genuine generalization from memorization of the static-background condition; however, that is a secondary weakness. The proposed test would settle whether the static assumption is a benign simplification or a boundary of the method. Verdict should stay CONDITIONAL: accept if the authors add the dynamic-medium test or explicitly scope the claim to static scattering media; reject if the claim is presented without that scope. I would not move to REJECT because the proof-of-concept is coherent and reproducible in principle, and machine-checked proof is not applicable to empirical optics.","tokens_in":21075,"tokens_out":2455,"duration_ms":26738,"concrete_test":"Fix the transmission setup of Fig. 2(a) with the same MNIST targets and phantom, but mount the scattering slab on a piezo shaker or modulate a second diffuser in the illumination path to add controlled optical-density fluctuations (e.g., 0.1–10 Hz, amplitudes causing 10–1000 extra DVS events per ms). Retrain and test the same SNN on target trajectories under these conditions; report tracking MSE and reconstruction SSIM versus fluctuation amplitude. If SSIM drops below a threshold such as 0.7 or tracking error increases markedly, the central claim must be restricted to static or slowly varying media. As a control, repeat with the phantom stationary to verify the setup reproduces the reported performance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism — DVS hardware attention isolating target information from uninformative background — depends on the target being the only source of temporal intensity changes. The paper states 'The turbid medium is taken as static though could in principle be slowly varying in its optical density' (Introduction, near Fig. 1), and all benchtop experiments use a fixed silicone phantom (Methods: Phantom Preparation). Under that condition, the DVS thresholding indeed suppresses the static background and the SNN can learn target-related spikes. But the headline claim concerns 'strongly scattering media' generally; fog, tissue, smoke, and turbid water all exhibit optical-density fluctuations, motion, or decorrelation on the timescales of tracking. If the phantom itself produces events — e.g., from vibration, thermal drift, or particle motion — those events are indistinguishable at the sensor from target events, the attention premise fails, and the trained SNN has never seen this input distribution. The DMD/E-ink targets are also projected light patterns rather than physical objects; however, this proxy is standard and secondary. The load-bearing gap is the untested static-medium assumption, not the network math.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents an end-to-end neuromorphic optical imaging system that combines a dynamic vision sensor (DVS) with a multistage deep spiking neural network (SNN) to track and reconstruct randomly moving or time-varying targets hidden behind strongly scattering silicone phantoms. In transmission geometry, MNIST-like projected characters move randomly behind a phantom with MFP=72; in reflection geometry, Kanji-MNIST characters with time-varying contrast are imaged behind a phantom with MFP=144. The network has a tracking module (OTM) and a reconstruction module (ORM plus residual refinement), trained end-to-end with surrogate gradients and a hybrid loss. The reported reconstruction metrics are SSIM 0.9577 and MSE 0.0108 (transmission) and SSIM 0.8068 and MSE 0.0583 (reflection), and the energy estimate claims up to 20x lower power than an equivalent ANN. The paper frames the DVS as a hardware attention mechanism that suppresses the static scattering background, with the SNN performing tracking and reconstruction in parallel.","tokens_in":21269,"tokens_out":3484,"duration_ms":36734,"significance":"If the empirical claims are fully supported, the work introduces a genuinely new combination of hardware and algorithmic ideas: event-based sensing as a physical attention layer for scattering media, paired with a spiking reconstruction/tracking network. The benchtop experiments use quantitatively characterized phantoms (MFP=72 and 144), standardized character sets, held-out test samples, and a surrogate-gradient training pipeline. These are concrete and reproducible elements that go beyond a purely conceptual proposal. The energy comparison with an ANN of identical architecture is a useful first-order estimate. However, the current evidence is incomplete in several load-bearing respects: tracking accuracy is never quantified, reported metrics lack variance across the five training iterations, the static-medium assumption is untested, and no baseline comparison is provided. The central idea is promising and largely defensible, but the manuscript as written overstates its validated scope.","major_comments":[{"comment":"The central claim includes tracking, yet tracking accuracy is never quantitatively evaluated. Figure 4(b) shows inferred positions as green crosses versus ground truth as red crosses, but no metric such as center-coordinate root-mean-square error, mean absolute error, or per-time-step error is reported anywhere in the text. Without a quantitative tracking metric on the held-out test set, the tracking claim is supported only by a single visual example. Please add numeric tracking errors over all test trajectories, ideally with per-time-step distributions.","section":"Object Tracking Module / Figure 4(b)"},{"comment":"The paper states that five training iterations were conducted with different initializations, but all reported reconstruction metrics are single point estimates (SSIM 0.9577, MSE 0.0108; SSIM 0.8068, MSE 0.0583). With only five runs, the standard deviation is easy to compute and is needed to assess whether the differences between geometries and configurations are meaningful. Please report mean and standard deviation (or individual values) for all headline metrics across the five iterations.","section":"Methods, Training and Evaluation / Figures 5 and 6"},{"comment":"The method's core hardware-attention mechanism relies on the scattering medium being static: the DVS emits events only for temporal intensity changes, so if the phantom itself had fluctuating optical density, motion, or decorrelation, those events would be indistinguishable from target events at the sensor. The manuscript explicitly states 'The turbid medium is taken as static though could in principle be slowly varying in its optical density,' and all benchtop experiments use a fixed silicone phantom. Yet the abstract and discussion generalize to fog, tissue, smoke, and other strongly scattering media. This assumption is load-bearing: either add experiments with a deliberately perturbed or slowly fluctuating phantom to test robustness, or restrict the claims in the abstract and title to static scattering media.","section":"Introduction (near Fig. 1) and Methods, Phantom Preparation"},{"comment":"No baseline comparison is provided to calibrate the reported reconstruction and tracking results. The paper does not compare against a frame-based camera with a conventional CNN, against the same DVS data processed by a non-spiking network, or against existing speckle-correlation methods for tracking through scattering media. The energy comparison is only between the SNN and an ANN of the same architecture, which is informative but does not address whether the full neuromorphic pipeline outperforms simpler alternatives in accuracy. Please add at least one appropriate baseline on the same benchtop data, or temper the comparative claims of advantage in the Introduction and Discussion.","section":"Discussion / Results comparison"}],"minor_comments":[{"comment":"In Eq. (12), the third loss term is written with weight 'a' (BCE weight) rather than a distinct weight, and the preceding paragraph says the loss comprises MSE and SSIM, while the equation uses BCE, SSIM, and IoU. Please correct the typo and align the text with the equation.","section":"Methods, Eq. (12)"},{"comment":"The Fig. 4 caption reports SSIM 0.9568 and MSE 0.0108 for the single digit '4', while the main text reports SSIM 0.9577 across the full testing dataset; please clarify which number corresponds to which evaluation set so the reader can distinguish the example from the aggregate result.","section":"Figure 4 caption and Results text"},{"comment":"Several typos and formatting errors appear: 'grpup' in the Fig. 4 caption, 'repectedly' after the SSIM/MSE values, 'Kanji MINST' in the Results section, and 'CONVERGENC' in the section heading. These should be corrected.","section":"Various locations"},{"comment":"The data availability statement says data 'will be available in a provided GitHub repository' but no repository link is given in the manuscript; please include the actual link or a clear statement of how to obtain the code and data.","section":"Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the central idea is worth publishing once the empirical support is tightened. The main concerns are not correctness of the SNN training per se, but missing quantitative tracking metrics, missing variance over training runs, the untested static-medium assumption relative to the broad claims, and the absence of baselines. I would encourage the editor to require these additions rather than merely cosmetic revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this one before reading: the paper is a real benchtop demonstration, not vaporware, and the system-level integration is new. The authors combine a DVS event camera with a two-module spiking neural network (tracking + reconstruction) and show results through silicone phantoms with MFP 72 (transmission) and 144 (reflection). That specific combination is not in the prior art they cite, though their own refs 42–44 already did DVS-plus-SNN in turbid media, so it is an extension of an established research program rather than a leap. They also give held-out test metrics (SSIM 0.9577 and 0.8068), which is real evidence.\n\nWhat the paper does well: the optical engineering is careful — calibrated phantoms, MFP quantified, DMD and E-ink proxies acknowledged as such, and the static-medium assumption stated plainly in the introduction. The incremental nature of the architecture (U-Net, BASNet-style refinement) is not hidden. The energy comparison is clearly an estimate from spike rates and standard per-op costs, not a measured result, and they say so.\n\nThe soft spots are mostly about evidence strength, not method. Tracking accuracy is never quantified — the OTM outputs coordinates, but there is no error metric, no plot of tracked vs. true trajectories with numbers. The 5 training iterations are mentioned but no variance or error bars are reported. There is no baseline comparison against a frame-based camera with an ANN, or against a non-spiking reconstruction, which weakens the \"fully neuromorphic\" efficiency claim beyond the paper estimate. The \"fully neuromorphic\" label also overstates what was run: the SNN is simulated on an RTX 3090, not on Loihi or equivalent hardware.\n\nThe stress-test concern about the static-medium assumption is valid but not fatal. The paper explicitly says the medium is static and \"could in principle be slowly varying,\" so this is a scope limitation, not a hidden flaw. The claim \"through strongly scattering media\" should be read as \"through static strongly scattering media.\" Generalization to fog, tissue, or turbid water with dynamic optical density is unproven.\n\nWho should read it: people working in computational imaging through turbid media or event-driven vision who want a concrete integration example. It deserves a serious referee: the experimental setup is solid, the limitation statements are honest, and the missing pieces (tracking metric, error bars, baseline) are additions, not fundamental corrections. I would send it to peer review with a request for those additions, plus code and data — the GitHub link is still a promise.","headline":"A genuinely integrated DVS-plus-SNN demonstration through calibrated dense phantoms, but the empirical evidence is thinner than the claims — track metrics, error bars, and a baseline are missing.","tokens_in":21843,"tokens_out":1475,"would_cite":false,"duration_ms":16475,"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 end-to-end neuromorphic pipeline—event camera plus deep spiking neural network—can track and reconstruct randomly moving hidden objects through strongly scattering media, achieving image similarity above 0.95 in transmission experiments.","keywords":["neuromorphic imaging","event camera","dynamic vision sensor","spiking neural network","scattering media","optical tracking","image reconstruction","turbid media"],"falsifier":"A decisive test is to make the phantom itself fluctuate—slowly stirring the scattering particles or vibrating the slab while the target moves—and then measure whether the DVS event stream still encodes the target and whether tracking MSE and reconstruction SSIM hold; if they collapse, the static-medium assumption is load-bearing. A complementary test replaces the DMD-projected 2D character with a solid 3D object moving in depth and checks whether a recognizable reconstruction still emerges.","tokens_in":20799,"feed_emoji":"🎯","tokens_out":8493,"duration_ms":74001,"temperature":0.7,"pith_summary":"This paper attempts to show that a fully neuromorphic imaging pipeline—an event camera feeding a deep spiking neural network—can track and reconstruct visually hidden moving objects through strongly scattering media. The authors argue that the event camera acts as a hardware attention filter: it emits spikes only when light changes, so a moving target creates a signal while the static scattering background stays silent. They combine that sparse spike stream with a two-module SNN that simultaneously estimates the target's coordinates and reconstructs its image, using temporal memory to converge on a sharp picture over time. In benchtop experiments the system tracks randomly moving handwritten digits through a phantom with 72 scattering mean free paths (SSIM 0.9577) and reconstructs time-varying Kanji characters in a 144-MFP reflection geometry (SSIM 0.8068). If correct, this offers a path to real-time, low-power imaging in fog, tissue, and other turbid settings where conventional cameras fail.","feed_headline":"Event camera plus spiking network sees through 72-scatter-length fog","feed_subtitle":"Benchtop tests reconstruct moving objects with 0.96 similarity and about 20x less energy than a conventional network.","key_machinery":"The load-bearing mechanism is the event camera's change-detection front end plus the deep SNN's temporal memory. A dynamic vision sensor (DVS) fires asynchronous per-pixel spikes only when log-intensity crosses a threshold, so a moving target in a static scattering medium produces events while the diffuse background produces almost none; this is the hardware-level attention that isolates the target. The SNN then processes those spikes through leaky integrate-and-fire neurons with trainable membrane decay, stateful synapse filters, and a last-layer reset in the tracking module, which lets the network accumulate evidence over time and produce both coordinates and images.","core_discovery":"The central claim is that combining event-driven sensing with a deep spiking neural network yields simultaneous tracking and image reconstruction of targets that are optically invisible through dense scattering media. Photons that have traversed a turbid phantom are detected by a dynamic vision sensor, which converts intensity changes into asynchronous positive and negative spikes; preprocessing turns these into spiking tensors of shape [time, polarity, height, width]. A tracking module (convolutional SNN encoder plus linear mapping) outputs normalized [x,y] center coordinates, while a reconstruction module (spiking U-Net encoder-decoder with skip connections, followed by a residual refinement module) outputs a probability map of the target; both run in parallel over time steps. The network is trained end-to-end with a surrogate gradient and a hybrid loss combining BCE, SSIM, and IoU for reconstruction and MSE for tracking. On the test set, transmission experiments with random MNIST digit motion at 72 mean free paths reach SSIM 0.9577 and MSE 0.0108, and reflection experiments with Kanji characters at 144 mean free paths reach SSIM 0.8068 and MSE 0.0583. The authors also estimate the SNN consumes about 20x less energy than an architecture-matched ANN (7.99 mJ vs 142.32 mJ).","pith_inferences":["If the static-medium condition is relaxed, a slowly varying phantom would inject its own events into the DVS; the attention mechanism would then need a motion prior or contrast filter to separate target from medium dynamics, which the paper does not address.","The DMD-projected characters are planar and pre-specified; a real solid object moving in three dimensions would add occlusion, perspective, and depth motion that the current two-module network has not been shown to handle.","The energy advantage is tied to spike sparsity; in scenes with dense, continuous background dynamics the AC/MAC ratio and the 20x figure would shrink, so the comparison should be re-measured at higher background event rates.","A natural extension is to feed the tracking module's coordinate output back into the reconstruction module as a positional prior, which could sharpen reconstructions at early time steps where the current system is blurry."],"forward_implications":["In transmission through a phantom with 72 scattering mean free paths, the system tracks random x-y trajectories and reconstructs the moving character with SSIM 0.9577, so the same pipeline should apply to other optically hidden moving targets such as vehicles in fog.","In reflection with 144 mean free paths, spatially fixed but optically time-varying characters are reconstructed with SSIM 0.8068, indicating the method works when the illuminating light is incoherent and the detector sees double-pass scattering.","The measured energy estimate for the SNN engine is about 20 times lower than an equivalent ANN (7.99 mJ vs 142.32 mJ), which makes real-time, battery-limited field deployment plausible.","Because spikes are the sole currency of information from sensor to output, the approach is compatible with emerging neuromorphic hardware and with wireless transmission of sparse events."],"supporting_citations":[{"why":"Supplies the surrogate-gradient training method that lets the deep SNN be trained end-to-end with backpropagation.","marker":"[61]"},{"why":"Provides the U-Net encoder-decoder structure with skip connections that the reconstruction module is built on.","marker":"[65]"},{"why":"Inspires the residual refinement module that learns residual differences to sharpen reconstructed boundaries.","marker":"[66]"},{"why":"Motivates the stateful synapse filters added after SNN blocks to give the network temporal memory.","marker":"[67]"},{"why":"Establishes the dynamic vision sensor as the asynchronous event camera hardware that generates the input spike trains.","marker":"[28]"},{"why":"Provides the survey context for event-based vision that frames the DVS attention mechanism.","marker":"[29]"},{"why":"Supplies the AC/MAC accounting method used for the SNN-vs-ANN energy comparison.","marker":"[55]"},{"why":"A representative prior speckle-correlation method that tracks, but does not image, moving objects through scattering media; the paper contrasts its own simultaneous tracking-and-imaging result against it.","marker":"[7]"}],"fun_headline_variants":["Neuromorphic camera+SNN tracks and images through dense fog","Event camera and spiking net see through scattering media, save 20x energy","SNN with event camera reveals moving targets behind turbid media","Tracking and imaging via spikes: neuromorphic vision through scattering","Neuromorphic imaging: event-driven camera and SNN pierce turbid media"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the scattering medium stays static while the target is the only dynamic light source; if the medium fluctuates or background motion creates comparable events, the event camera's attention mechanism no longer isolates target information and the learned spiking network loses its input signal.","fun_headline_variants_meta":{"raw":{"variants":["Neuromorphic camera+SNN tracks and images through dense fog","Event camera and spiking net see through scattering media, save 20x energy","SNN with event camera reveals moving targets behind turbid media","Tracking and imaging via spikes: neuromorphic vision through scattering","Neuromorphic imaging: event-driven camera and SNN pierce turbid media"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3489,"prompt_tokens":1028,"completion_tokens":2461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":2367}},"tokens_in":644,"tokens_out":2461,"duration_ms":18070,"temperature":1.0,"reasoning_tokens":2367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:45:24.718674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to make the phantom itself fluctuate—slowly stirring the scattering particles or vibrating the slab while the target moves—and then measure whether the DVS event stream still encodes the target and whether tracking MSE and reconstruction SSIM hold; if they collapse, the static-medium assumption is load-bearing. A complementary test replaces the DMD-projected 2D character with a solid 3D object moving in depth and checks whether a recognizable reconstruction still emerges.","supporting_citations":[{"cited_title":"O., Mostafa, H","cited_arxiv_id":null,"evidence_quote":"Supplies the surrogate-gradient training method that lets the deep SNN be trained end-to-end with backpropagation."},{"cited_title":"Exploiting Neuron and Synapse Filter Dynamics in Spatial Temporal Learning of Deep Spiking Neural Network","cited_arxiv_id":"2003.02944","evidence_quote":"Motivates the stateful synapse filters added after SNN blocks to give the network temporal memory."},{"cited_title":"& Delbruck, T","cited_arxiv_id":null,"evidence_quote":"Establishes the dynamic vision sensor as the asynchronous event camera hardware that generates the input spike trains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the survey context for event-based vision that frames the DVS attention mechanism."}],"review_version":1}