{"id":"7a135477-395e-4335-a5a7-15bdcf99a7ee","arxiv_id":"2506.07878","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Mamba-based spatio-temporal state space network estimates event-camera optical flow with 32 GMACs and 1.11 EPE on DSEC, claiming large compute savings over prior methods.","lead":"The authors present an event-camera optical flow network built around Mamba state space models instead of heavy iterative refinement or transformer components. It reports competitive accuracy on DSEC with far fewer computations, which could benefit real-time robots and drones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline efficiency-accuracy claim rests on an ambiguous evaluation split: Table 1 appears to mix a custom validation split with published test-set numbers, and the 4.5x inference speedup has no reported measurement protocol.","rationale":"The reader's weakest assumption focused on the sufficiency of a single spatio-temporal event volume (§3.1.3). That is a substantive modeling concern, but it is not the single most load-bearing issue for the paper's headline claim. The headline asserts a specific efficiency-accuracy tradeoff: 4.5x faster inference and 8x lower computations than TMA, 2x lower than EV-FlowNet, with competitive performance. Every quantitative component of that claim depends on the comparison being measured under an identical protocol. The paper creates a custom validation split because DSEC test ground truth is unavailable, yet Table 1 reports results on 'DSEC' without clarifying whether baseline numbers were recomputed on that split. If they were not, the reported EPE 1.11 is not comparable to the listed baselines' EPEs, and the claim of 'competitive performance' collapses even if the architecture is sound. The 4.5x inference-speed figure is even less supported, since no timing experiment appears anywhere in the manuscript. These are addressable issues—release code, re-evaluate baselines on the same split, measure latency—so the appropriate verdict remains conditional rather than reject. The reader's single-volume concern is real, but its empirical support would be restored once the comparison is made fair; conversely, if the evaluation is unfair, the single-volume premise is left without the evidence the paper presents for it. Thus the evaluation-protocol ambiguity is the more fundamental threat to the central claim, and agreement with the reader is only partial.","tokens_in":12234,"tokens_out":4601,"duration_ms":51970,"concrete_test":"Re-run the exact validation split described in §4.1 with the authors' code (or requested from the authors) and evaluate the official TMA and E-RAFT checkpoints on that same split on the same GPU; report EPE, AE, 1PE, 3PE, and wall-clock inference time per sample. If the EPE gap to TMA grows from the reported 0.37 px, or if the measured wall-clock speedup differs from 4.5x, the headline claim must be revised. As a stronger alternative, submit the model to the public DSEC test server and compare official test EPE against the published test-set baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method is simultaneously far cheaper and competitive in accuracy on DSEC. Both halves of this claim require that all compared numbers come from the same evaluation protocol. Section 4.1 states: \"Ground truth for test data is not available, so we created a validation split following [2] for training and ablation studies.\" Table 1, however, is captioned \"Evaluation results on DSEC [9] dataset\" and lists baselines such as TMA, E-RAFT, and EV-FlowNet with their published EPE/GMAC values, without stating that these baselines were re-evaluated on the custom split. If the baseline EPEs in Table 1 are taken from the official DSEC test server or original publications while \"Ours\" is evaluated only on the custom validation split, the comparison is not apples-to-apples, and the \"competitive performance\" part of the headline is unsupported. This is the most load-bearing weakness because the architecture's justification for using a single event volume (§3.1.3) is empirically supported only by these very numbers; if the comparison is invalid, that modeling premise also loses its direct evidence. Separately, the abstract's \"4.5x faster inference compared to TMA\" is not backed by any wall-clock timing experiment in the paper: Table 2 reports GMACs and memory but no latency, and no section describes a controlled speed measurement on identical hardware. A factor that appears only in the abstract without methodology cannot support the central efficiency claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatio-temporal state space model (STSSM) for event-based optical flow. The network encodes a single spatio-temporal event volume with a series of STSSM blocks built on Mamba, then predicts low-resolution flow and uses convex upsampling to obtain full-resolution flow, avoiding 4D correlation volumes and iterative refinement. The authors report an EPE of 1.11 on DSEC at 32 GMACs, claim 8x lower computation than TMA and 2x lower than EV-FlowNet with competitive accuracy, and provide ablations over SSM variants, temporal/position encoding, and backbone choices.","tokens_in":12641,"tokens_out":5231,"duration_ms":57922,"significance":"If the quantitative claims hold, the work is a useful step toward efficient event-based optical flow, demonstrating that SSMs can replace expensive correlation volumes and iterative refinement while retaining competitive accuracy. The paper's strengths include a clearly described architecture, extensive ablations across S4, S4D, S5, and Mamba, experiments on both DSEC and BlinkFlow, and a planned public code release. However, the central efficiency-accuracy claim depends on an unambiguous evaluation protocol and a direct timing comparison, both of which are currently missing or ambiguous. The novel application of SSMs to event-based flow is interesting and likely of interest to the workshop audience, but the headline numbers require verification before the claims can be accepted as stated.","major_comments":[{"comment":"The evaluation protocol behind Table 1 is not specified. Section 4.1 states that ground truth for the DSEC test set is unavailable and that you created a validation split following [2] for training and ablation studies, but Table 1 is captioned \"Evaluation results on DSEC [9] dataset\" and does not state whether all rows were evaluated on the same split. If the baseline EPE/AE/3PE values are the published test-set numbers while \"Ours\" is evaluated on the custom validation split, the comparison is not apples-to-apples and the \"competitive performance\" claim is unsupported. Please state for every row the exact split and source of the numbers, and either re-evaluate all baselines on the same split or submit to the official test server.","section":"§4.1 / Table 1"},{"comment":"The abstract claims \"4.5x faster inference compared to TMA\", but no wall-clock latency comparison is reported anywhere. Table 4 reports inference times only for the authors' own SSM variants (6.3-20.9 ms) and Table 2 reports GMACs and memory, not latency. There is no description of a controlled timing experiment on identical hardware for TMA or E-RAFT. Please add a measurement protocol (GPU, batch size, input resolution, number of runs) and actual latencies for all compared methods, or remove the inference-speed claim from the abstract.","section":"Abstract / §4.2"},{"comment":"The quantitative efficiency claims are internally inconsistent. Section 4.2 states a 16x computational reduction versus TMA, while the abstract says 8x lower computations compared to TMA; Table 1 lists TMA at 522 GMACs, which is about 16.3x the 32 GMACs reported for the proposed method, whereas 8x corresponds to Table 2's TMA with 2 iterations (233.6 GMACs). Also, the claim of a \"264% reduction in EPE\" relative to EV-FlowNet is arithmetically wrong: 2.32 to 1.11 is a 52% reduction. Please correct these numbers and state the reference configuration explicitly for each comparison.","section":"§4.2 / Abstract"},{"comment":"The modeling premise that a single sufficiently large event volume ε(tR,tT) suffices for flow estimation, without reference/target views or correlation volumes, is not directly tested. The paper argues this from the event-generation model, but the only empirical support is the DSEC comparison, whose split is ambiguous (see above). An ablation that varies the temporal window size, or compares the single-volume input against a two-volume/correlation-volume variant with the same backbone, would directly test this premise. Without it, the architecture's central design choice rests on an unverified assumption.","section":"§3.1.3"}],"minor_comments":[{"comment":"There are several typos: \"vocel grid\" should be \"voxel grid\", \"demostrate\" should be \"demonstrate\", and \"T A system\" in §3.1.3 appears to be a formatting artifact.","section":"§3.2.1 / Fig. 5"},{"comment":"The table layout is not self-explanatory: the meaning of the row labels \"No\", \"t\", and \"t+p\" under the shared header \"Enc.\" should be defined in the caption. Also, for Vi-T, \"t+p\" (EPE 1.15) improves over \"t\" (EPE 1.34), so the conclusion that position encoding is unnecessary should be scoped to the Mamba branch, not stated as a general finding.","section":"Table 3"},{"comment":"In Table 2, the \"iters\" column for \"Ours\" is set to 1 even though the model is not iterative; clarify what this entry represents. In Table 5, the \"Conv (Baseline)\" row has no value in the N column; indicate whether N is not applicable.","section":"Tables 2 and 5"},{"comment":"The right-hand comparison plot lacks axis labels and units in the caption; please specify what the horizontal and vertical axes represent (e.g., GMACs and EPE) and identify the marker for the proposed method.","section":"Figure 1"},{"comment":"The validation split following [2] is not described concretely. Please list the DSEC sequences used for training and validation, the temporal window size, and the number of evaluation samples so that the results can be reproduced and compared.","section":"§4.1"},{"comment":"Equations (7) and (8) reuse the symbol A for both the continuous-time and discretized state matrices; using a distinct symbol (e.g., Ā) would avoid confusion. Also, the statement \"for the first time we study the use of SSMs for spatio-temporal data\" overclaims novelty given the cited VideoMamba and VMamba works; it should be rephrased as \"for the first time for event-based optical flow\" or similar.","section":"§3.1.3 / §3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a CVPR workshop and the core idea is timely. My main concern is that the headline efficiency and accuracy claims rest on an ambiguous evaluation split and an unreported timing protocol; these are fixable with additional experiments and clarifications rather than being fundamentally flawed. I would not recommend rejection, but the manuscript needs a major revision before the central claims can be trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Lead with the good: the STSSM module is a real design contribution. Spatio-temporal patching plus temporal embedding before a Mamba sequence transform is new for event optical flow, and the convex upsampling head avoids the usual heavy decoder. The compute is genuinely low: 32 GMACs in Table 1. The ablations against S4, S4D, S5, and a Transformer are useful and give credit where due.\n\nThe problems are in the claims, not the architecture. The abstract says 4.5x faster inference versus TMA, but there is no wall-clock experiment anywhere. Table 2 reports GMACs and memory, not latency. So that headline number is unsupported. Also the compute factor is internally inconsistent: the abstract says 8x lower than TMA, the text says 8x lower than E-RAFT and 16x lower than TMA, and Table 1's numbers (522/32) give about 16x. Careless.\n\nThe more load-bearing issue is the evaluation split. Section 4.1 says test ground truth is unavailable and a validation split was created following [2]. Table 1, however, lists baseline EPEs that look like published test-set numbers. If the baselines were not re-evaluated on the same split, then the competitive performance claim is apples-to-oranges. The paper never states that the baselines were rerun. This directly undercuts the single-volume argument in Section 3.1.3, because that argument relies on the same comparison. The reader's weakest-assumption point is right.\n\nI also note the first study of SSMs for spatio-temporal data is overstated given references [26] and [39] apply SSMs to event and neuromorphic data. Code is promised but absent, so 32 GMACs is not independently checkable. To the paper's credit, it does mention the method struggles on occlusions and thin objects in BlinkFlow, which is honest.\n\nBottom line: this is a plausible, efficient architecture that deserves serious referee time, but the empirical claims need a re-run on one protocol plus a real timing measurement before they can be trusted. The reader's conditional verdict is fair. My recommendation: send to peer review, with the evaluation issue as the main request.","headline":"A genuinely cheap Mamba-based flow architecture, but the headline efficiency and accuracy claims rest on an inconsistent comparison and an unsupported inference-time number.","tokens_in":829,"tokens_out":1942,"would_cite":true,"duration_ms":69474,"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":"A state-space model estimates event-camera optical flow with a fraction of the compute of today's best methods while staying close to their accuracy.","keywords":["event-based optical flow","state space model","Mamba","spatio-temporal encoder","convex upsampling","DSEC benchmark","efficient inference","event camera"],"falsifier":"Run the same STSSM architecture with an added second, temporally shifted event volume feeding a small correlation lookup (as in iterative-refinement methods) while keeping compute similar; if the EPE on DSEC drops by more than the roughly 0.3 pixel margin the paper reports against E-RAFT, that would show the single-volume premise, not the state-space model, is doing the accuracy work.","tokens_in":12097,"feed_emoji":"⚡","tokens_out":8788,"duration_ms":92109,"temperature":0.7,"pith_summary":"Event cameras record brightness changes as asynchronous events, opening the door to very fast motion estimation; the obstacle is that the most accurate neural networks for this task are computationally heavy. This paper tries to establish that a network built around a Spatio-Temporal State Space Model (STSSM) can estimate optical flow with accuracy close to the best iterative-refinement methods while using a fraction of their compute. On the DSEC driving benchmark the authors report an average endpoint error of about 1.11 pixels at 32 GMACs, which they state in the abstract as 4.5x faster inference and 8x lower computation than TMA, and 2x lower computation than EV-FlowNet. If that holds, event-based flow becomes practical for real-time, low-power systems where the accurate methods are too slow and the fast methods are too inaccurate.","feed_headline":"8x less compute for event-camera optical flow via state-space model","feed_subtitle":"Matches near-top accuracy on DSEC at 32 GMACs, fast enough for real-time robotics.","key_machinery":"The core module is the Spatio-Temporal State Space Model (STSSM) block. It takes a spatio-temporal event volume (a voxel grid discretizing events into time bins, height, and width), cuts it into 3D patches of size m by k by k per channel, flattens them through a linear projection into a 1D sequence, adds a learnable temporal embedding to separate the temporal from the spatial axes, runs the sequence through a selective state-space model (the paper uses Mamba-style SSM layers) for linear-time sequence transformation, and reprojects the sequence back into a feature volume. Four such blocks form an encoder that progressively reduces spatial resolution while increasing channels and compressing time to a single frame; a flow head plus mask head then performs RAFT-style convex upsampling to full resolution, replacing the heavy decoder used in U-Net architectures.","core_discovery":"The central claim is that optical flow can be estimated from event data by a single spatio-temporal voxel volume processed with state-space models (SSMs), removing the need for 4D correlation volumes and reference/target view pairs that characterize current iterative-refinement methods. The authors argue that within a sufficiently large time window the events alone determine the flow, and they build a network whose encoder stacks four STSSM blocks, each patching the volume, projecting to a 1D sequence, adding a temporal embedding, applying selective state-space layers, and reprojecting to a feature volume, ending in a 2D feature map from which a lightweight head predicts flow plus convex upsampling masks for full-resolution output. On the DSEC benchmark they report an average endpoint error of about 1.11 pixels at 32 GMACs, with the abstract claiming 4.5x faster inference and 8x lower computation than TMA and 2x lower than EV-FlowNet; the experimental section separately quotes 8x and 16x reductions against E-RAFT and TMA. The paper also presents ablations indicating that Mamba-style SSMs outperform transformer and 3D-CNN encoders for this task at lower cost.","pith_inferences":["The single-volume assumption, if correct, extrapolates to other event-based tasks that currently pair reference and target views: depth, egomotion, and deblurring could each be reformulated as single-volume mapping problems and inherit the same compute savings; that is an inference, not a claim of the paper.","Because SSMs scale linearly in sequence length, the architecture may tolerate much longer temporal windows than the paper tests, potentially improving accuracy on slow motion or at low event rates; this is testable by scaling the input time dimension on DSEC, which the paper does not do.","The reported speedups versus TMA depend on operating points (iteration counts) and hardware; a same-GPU wall-clock benchmark varying iteration counts would more directly verify the 4.5x figure, and the paper's Table 2 already hints the comparison is best-case for the proposed method.","The qualitative BlinkFlow results reveal weakness on occlusions and thin objects; an independent benchmark with dense independent moving objects and occlusions would clarify whether the single-volume state-space design has a systematic failure mode."],"forward_implications":["If the central claim holds, accurate event-based optical flow becomes feasible on embedded or mobile processors, since 32 GMACs per inference is within reach of real-time hardware.","The single-volume design removes the 4D correlation volume, the main space and time bottleneck of iterative-refinement methods, so frame-rate gains scale with camera resolution rather than with the flow search range.","The ablations showing Mamba-style SSMs beating a ViT encoder in the same head suggest attention may be unnecessary for event-based motion; future event-vision backbones could standardize on state-space layers.","Replacing the U-Net decoder with convex upsampling cuts most of the decoder cost (the paper notes EV-FlowNet's decoder is about 80 percent of its compute), so later efficient flow networks can follow the same lightweight-head pattern."],"supporting_citations":[{"why":"DSEC: the benchmarking dataset all accuracy and efficiency numbers are measured on.","marker":"[9]"},{"why":"E-RAFT: supplies the event voxel-grid representation and is the main iterative-refinement baseline; its encoder is also the Conv baseline in ablations.","marker":"[10]"},{"why":"Mamba: the selective state-space model used inside every STSSM block for the seq-to-seq transformation.","marker":"[12]"},{"why":"S4: the structured state-space model whose formulation grounds the SSM theory and serves as an ablation variant.","marker":"[14]"},{"why":"TMA: the temporal-motion-aggregation iterative method against which the central efficiency claims (8x lower compute, 4.5x faster inference) are made.","marker":"[20]"},{"why":"RAFT: source of the convex upsampling and mask-head design that replaces the heavy decoder.","marker":"[30]"},{"why":"EV-FlowNet: the U-Net baseline with roughly 80 percent decoder overhead, used for the 2x compute comparison.","marker":"[36]"}],"fun_headline_variants":["Event flow: 8x less compute with state-space model","State-space network does event flow at 8x lower cost","8x compute drop for event optical flow via SSM","Event camera flow: SSM cuts compute 8x, keeps accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single event voxel volume over a large enough time window contains all the information needed to estimate the flow; if this is false, the network's accuracy would collapse no matter how good the state-space backbone is, and the paper does not directly test it.","fun_headline_variants_meta":{"raw":{"variants":["Event flow: 8x less compute with state-space model","State-space network does event flow at 8x lower cost","8x compute drop for event optical flow via SSM","Event camera flow: SSM cuts compute 8x, keeps accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1609,"prompt_tokens":1022,"completion_tokens":587,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":638,"tokens_out":587,"duration_ms":7015,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:23:28.041972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same STSSM architecture with an added second, temporally shifted event volume feeding a small correlation lookup (as in iterative-refinement methods) while keeping compute similar; if the EPE on DSEC drops by more than the roughly 0.3 pixel margin the paper reports against E-RAFT, that would show the single-volume premise, not the state-space model, is doing the accuracy work.","supporting_citations":[{"cited_title":"Dsec: A stereo event camera dataset for driv- ing scenarios","cited_arxiv_id":null,"evidence_quote":"DSEC: the benchmarking dataset all accuracy and efficiency numbers are measured on."},{"cited_title":"E-raft: Dense optical flow from event cam- eras","cited_arxiv_id":null,"evidence_quote":"E-RAFT: supplies the event voxel-grid representation and is the main iterative-refinement baseline; its encoder is also the Conv baseline in ablations."},{"cited_title":"Efficiently mod- eling long sequences with structured state spaces","cited_arxiv_id":null,"evidence_quote":"S4: the structured state-space model whose formulation grounds the SSM theory and serves as an ablation variant."},{"cited_title":"TMA: Temporal Motion Aggregation for Event-based Optical Flow","cited_arxiv_id":"2303.11629","evidence_quote":"TMA: the temporal-motion-aggregation iterative method against which the central efficiency claims (8x lower compute, 4.5x faster inference) are made."}],"review_version":1}