{"id":"7ccf3060-da89-467d-982f-dc32ae47bdb8","arxiv_id":"2506.03512","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EDCFlow combines temporally dense, multi-scale feature differences at high resolution with a low-resolution cost volume to achieve accurate, efficient event-based optical flow.","lead":"An event-camera optical flow network, EDCFlow, fuses high-resolution pixel-difference motion cues with a low-resolution cost volume to estimate flow at 1/4 resolution with lower cost than prior models. It reports better or comparable accuracy than E-RAFT, TMA, and IDNet on DSEC and MVSEC, and can refine the outputs of existing RAFT-style event flow networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear-motion warp in Eq. (4) is the load-bearing risk: if the reference window is temporally offset or motion is accelerated, the difference maps encode systematic misalignment.","rationale":"The reader's weakest_assumption identified the linear-motion assumption, and I agree it is a genuine risk. My analysis sharpens it: Eq. (4) may be biased even under constant velocity if the reference window v_0 is temporally offset from t, because the scaling assumes v_0 sits at time t while v_i sits at t + (i/g)dt. The paper's text is ambiguous about window durations, but the reference stream E_{t-dt->t} is described as a separate interval of size dt ending at t, which would shift the temporal centroids. This would make the difference maps encode a systematic misalignment that the network must learn to undo, weakening the claimed complementarity and generalization. However, the central claim is about DSEC/MVSEC performance, and the reported results are plausibly correct under the implemented (if under-specified) pipeline. The concern does not invalidate the conditional verdict; it adds a concrete robustness condition. Reproducibility and benchmark tuning also remain conditions, as the reader noted. I recommend no verdict change but advise the authors to clarify the temporal alignment and provide a synthetic nonlinear-motion evaluation.","tokens_in":17529,"tokens_out":15839,"duration_ms":160094,"concrete_test":"Generate a synthetic event dataset with known constant-velocity motion. Access the intermediate feature maps F_i and warp each toward F_0 using the ground-truth endpoint flow and the scaling in Eq. (4). Measure mean feature discrepancy (e.g., L1 distance) after warping; if it is large even when the endpoint flow is exact, the temporal alignment/scaling is wrong. Then run an acceleration sweep (0, 5, 10, 20 m/s^2) and compare EPE against TMA; a rapid EPE increase with acceleration would confirm that the linear-motion assumption is the limiting factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism relies on Eq. (4): f_{0->i} = (i/g) f^{k-1} is used to warp every intermediate feature map F_i toward the reference F_0 before computing difference maps. This presumes both (a) constant-velocity motion over the whole span and (b) that the temporal index i corresponds to a fraction i/g of the true flow interval. The paper's construction of v_0 from the reference stream E_{t-dt->t} (a full-duration interval ending at t) while v_1..v_g partition E_{t->t+1} suggests v_0's temporal centroid is not at time t; under that reading, the scaling is systematically off even for constant-velocity motion. If so, the difference maps contain misalignment-induced residuals rather than clean motion cues, and the reported accuracy partly reflects the network learning to compensate for a fixed warp bias. The authors acknowledge the linear-motion assumption as a limitation in the supplementary future-work section, but the concern is more severe than a generalization caveat: it affects the validity of the core difference-motion signal on the very benchmarks used. A quantitative check is needed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EDCFlow, an event-based optical flow network that fuses high-resolution (1/4) temporally dense feature difference maps with a low-resolution (1/8) cost volume. The target event stream is split into g time windows; each window's features are warped to the reference using a flow scaled by the temporal index, multi-scale difference maps are computed and adaptively fused with correlation features, and a GRU iteratively updates the flow. Experiments on DSEC and MVSEC report state-of-the-art or comparable accuracy together with lower model size and MACs than prior RAFT-like event-based methods, and the method is also proposed as a plug-and-play refinement module for existing RAFT-like networks.","tokens_in":17764,"tokens_out":11093,"duration_ms":117392,"significance":"If the reported results hold, EDCFlow provides a useful accuracy/efficiency operating point for event-based optical flow, in particular by estimating directly at 1/4 resolution without a high-resolution cost volume. The ablation study is a genuine strength: each component (difference branch, correlation branch, multi-scale attention, and channel attention) is shown to contribute, and the plug-and-play refinement experiments are practically valuable. The main caveats are that the core linear-motion warp underlying the difference layer is under-validated, and the reported accuracy differences are small and come from single-run evaluations. These issues are addressable with targeted experiments and uncertainty quantification, so the central claim is plausible but not yet fully established.","major_comments":[{"comment":"The linear-motion scaling f^{k-1}_{0->i} = (i/g) f^{k-1} is the mechanism that warps every intermediate feature map F_i to the reference before computing difference maps, so it is load-bearing for the central claim. The temporal alignment of v_0 and v_1,...,v_g is never defined explicitly: the paper does not state how dt relates to g or where the temporal centroid of each window lies. On the natural reading (v_0 over [t-dt, t] and v_1,...,v_g partitioning [t, t+1] with g*dt = 1), the specific worry that v_0's centroid is offset from t does not produce a systematic error under constant-velocity motion, because f_{0->g} spans the same one-unit duration as the supervised flow f^{t->t+1}. However, the constant-velocity assumption itself is load-bearing and is only acknowledged as a future-work limitation. I ask for (a) an explicit statement of the temporal indexing and the exact relationship between f_{0->g} and the supervised flow, and (b) a quantitative validation on trajectories with acceleration or rotation, for example synthetic events with known nonlinear motion or a variant that estimates per-window flows instead of using Eq. (4), to show that the difference signal does not degrade sharply when the linearity assumption is violated.","section":"3.3, Eq. (4)"},{"comment":"All quantitative claims rest on single-run evaluations. The DSEC improvement over TMA is 0.72 vs 0.74 EPE, and the ablation gains (W/o Diff 0.82, W/o Corr 0.83, W/o SE 0.74, W/o MSAttn 0.74, Ours 0.72) are of similar or smaller magnitude. Without error bars or multiple seeds, it is not possible to tell whether the reported ordering is significant. Since the public DSEC benchmark does not allow repeated submissions, please report mean and standard deviation over at least three seeds on a local held-out split (as in the supplementary error-map experiment) for the main DSEC numbers and for the core ablations, and state the evaluation protocol explicitly.","section":"Tables 1, 4-6"},{"comment":"The efficiency comparison is central to the claim of 'better performance with lower complexity,' but the paper does not state how MACs and runtime were obtained for the baselines. Differences in input resolution, implementation, or source (original paper vs re-measurement) can change MACs materially, and the table includes both measured and claimed numbers without distinction. Specify for each row whether the number is from the original paper or measured here, the exact input size used for MAC/runtime computation, and the runtime measurement protocol (GPU, batch size, number of runs).","section":"Table 1"}],"minor_comments":[{"comment":"The sentence 'We briefly review RFAT [35]' contains a typo; it should be 'RAFT'.","section":"3.1"},{"comment":"The sentence 'along with the reference event stream E_{t->t+1}' should read 'E_{t-dt->t}' to match the problem setup.","section":"3.2"},{"comment":"The loss exponent uses the index i, but the summation index is k; it should be 0.8^{K-k}.","section":"3.4, Eq. (7)"},{"comment":"The row for 'IDNet-4' cites [21] but should cite [39], since IDNet is [39].","section":"Table 3"},{"comment":"The 'Reduc.' column shows values such as -0.61, which are absolute EPE gaps rather than percentage reductions; please define this column clearly.","section":"Table 3"},{"comment":"The rows 'Ours (+in1/2/3)' are not directly comparable with baselines trained only on outdoor day2; if kept, mark them with a separate symbol and state that the day2-only rows are the ones used for direct comparison.","section":"Table 2"},{"comment":"Please state in the captions whether MACs and runtime are measured for all rows under identical conditions or taken from the original publications, and specify the input resolution used.","section":"Tables 1 and 3"},{"comment":"The abbreviations 'Add.', 'Concat', 'GRU', and 'DWConv3d' in the first column are not defined in the table caption or the main text; please define them.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"This is a competent engineering contribution that fits the journal's scope. The main risk is not internal inconsistency but an under-validated load-bearing assumption (Eq. (4)) and missing uncertainty quantification on small accuracy differences. I recommend major revision rather than rejection because both issues are fixable with additional experiments and reporting. The paper would also benefit from a clearer efficiency-measurement protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid engineering advance, not a big leap. EDCFlow shows you can replace the expensive temporally dense cost volumes in TMA with high-resolution temporal feature differences fused with a cheap low-res cost volume, and get better DSEC EPE (0.72 vs 0.74) at 2.5M params and 247 G MACs. The ablations are clean and each component earns its keep. The plug-and-play refinement mode for RAFT-like backbones is a useful extra.\n\nWhat's genuinely new is the combination: multi-scale temporal feature differences at 1/4 resolution with an attention-based fusion, rather than building multiple cost volumes. The paper is honest that TMA and IDNet already use pieces of this recipe; the contribution is the efficient fusion and the clear efficiency/accuracy benefit.\n\nI checked the stress-test worry about Eq. (4). It does not hold up. The flow scaling f_{0->i} = (i/g) f is between the centroids of the time windows. Under constant velocity, the centroid-to-centroid displacement is exactly proportional to the index, and the reference window's offset (starting at t-dt rather than t) cancels because both endpoints shift. So the linear-motion assumption is a genuine simplification for acceleration and occlusions, but it is not a systematic bias for constant-velocity motion. The authors acknowledge the limitation in the future-work note.\n\nSoft spots: no code or pretrained models released, so the tables are not independently verifiable. The numbers are single runs without error bars, which is typical for this benchmark but a referee should ask. Some hyperparameters (window count, scales, iterations) were selected using the public test set; that can inflate numbers, though the sim-to-real generalization test (train on Blinkflow, eval on DSEC) suggests the model is not just memorizing. The comparisons to IDNet-4 are fair; similar EPE but much lower MACs.\n\nThis paper is for event-based vision researchers and people building efficient recurrent flow networks. The central claim holds, the experiments are thorough, and the limitation section is honest. It deserves a serious peer review. Recommendation: send to reviewers; ask for code and multiple runs.","headline":"Solid event-flow efficiency win with clean ablations; the temporal-offset worry in Eq. (4) cancels under constant velocity, so the main caveats are reproducibility, not correctness.","tokens_in":18341,"tokens_out":5988,"would_cite":true,"duration_ms":59573,"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":"EDCFlow shows that temporally dense feature differences between adjacent event frames, fused with a low-resolution cost volume, give high-accuracy event-based optical flow at 1/4 resolution with lower compute than cost-volume-only methods.","keywords":["event-based optical flow","feature difference maps","cost volume","multi-scale temporal aggregation","attention-based fusion","high-resolution flow","lightweight network","refinement module"],"falsifier":"Construct or record an event sequence with known strong rotation or acceleration within the 20 ms window and compare EDCFlow's end-point error against a variant whose intermediate warps use a nonlinear motion model; if the linear-warp version does not degrade on such sequences, the assumption survives, and if it degrades sharply, the central mechanism fails exactly where the assumption breaks.","tokens_in":17306,"feed_emoji":"⚡","tokens_out":4916,"duration_ms":52259,"temperature":0.7,"pith_summary":"This paper tries to establish that event-based optical flow can be estimated at 1/4 resolution, with better accuracy and lower compute than cost-volume-only models, by exploiting temporally dense feature differences between adjacent event frames. The authors argue that cost volumes capture global correspondences but are expensive and miss intermediate motion, while feature differences are cheap, detail-rich, and noise-prone, so the two should be fused. Their EDCFlow network computes multi-scale difference maps at high resolution, looks up a low-resolution cost volume, and adaptively combines both motion signals to iteratively refine flow. If the claim holds, it gives event-based flow a better accuracy/efficiency operating point and a plug-and-play refinement for existing recurrent flow networks.","feed_headline":"Dense temporal differences sharpen event-based optical flow","feed_subtitle":"Fusing high-res difference maps with a low-res cost volume lifts accuracy and cuts compute on driving events.","key_machinery":"The central object is the multi-scale temporal feature difference layer. After dividing the event stream into g+1 short time windows and extracting shared-weight features, the layer warps each intermediate feature toward the reference with flow scaled by i/g, reduces channels, and computes dense differences D^s_j = F^l_{(j+1)s} - F^f_{js} at sampling strides s=1,2,5. Depthwise separable 3D convolutions aggregate these maps in space and time, an attention module weighs the scales, and a channel-attention fusion combines the resulting difference motion feature with the upsampled correlation motion feature. This machinery is what lets the network encode continuous intermediate motion at high resolution without building multi-frame high-resolution cost volumes.","core_discovery":"On its own terms, the paper claims that the complementarity between temporally dense feature differences and a cost volume is enough to make the cost volume need only run at 1/8 resolution while flow is estimated at 1/4 resolution. EDCFlow warps each of the g+1 intermediate event-frame features toward the reference using the linear scaling f_{0->i}=(i/g)f, computes multi-scale difference maps, encodes them with depthwise separable 3D convolutions, and fuses the resulting high-resolution motion features with upsampled correlation features through channel attention. On DSEC this yields an end-point error of 0.72 pixels, angular error 2.65 degrees, and 1-pixel error 10.0, with 2.5M parameters and 247 G MACs, improving on the temporally dense cost-volume method TMA while using less compute, and matching the high-resolution IDNet-4 while being far lighter. The same module appended to RAFT-like event networks at 1/4 resolution improves their motion-boundary detail at modest extra cost.","pith_inferences":["The same difference-map-plus-cost-volume fusion could be transferred to frame-based optical flow, where high-resolution refinement currently relies on expensive upsampled cost volumes; the linear-warp difference layer would need no temporal event splitting.","The linear-motion assumption in Eq. (4) is the main fragility; a natural test is whether fitting a second-order (accelerated) motion model for the intermediate warps reduces error on fast rotational or oscillatory sequences.","The method suggests a cheaper alternative to multi-frame cost volumes in other event tasks such as depth or ego-motion estimation, where dense temporal differences can encode motion between reference states.","Because the difference layer's cost scales as O(T N C) and the cost volume is kept at 1/8, further resolution increases become feasible if the correlation upsampling noise noted in the ablation is addressed."],"forward_implications":["High-resolution (1/4) event flow is achievable with a low-resolution (1/8) cost volume when dense feature differences supply the local motion detail.","Temporally dense cost volumes are not necessary: O(T N C) difference maps can replace O(T N^2 C) correlation searches without losing accuracy.","The learned fusion generalizes from simulation (Blinkflow) to real driving data with the smallest accuracy drop among compared methods.","Appending EDCFlow as a refinement stage after E-RAFT or TMA reduces EPE and sharpens motion boundaries at 1/4 resolution with only a modest increase in parameters and MACs.","At equal accuracy, EDCFlow uses about one fifth the computation of IDNet-4 and runs faster, improving deployability on resource-constrained platforms."],"supporting_citations":[{"why":"Supplies the recurrent all-pairs cost-volume and GRU update framework that EDCFlow extends with high-resolution difference maps.","marker":"[35]"},{"why":"Establishes the event-based RAFT baseline and the training/evaluation protocol on DSEC and MVSEC.","marker":"[12]"},{"why":"Bases the temporally dense cost-volume approach that EDCFlow replaces with cheaper dense feature differences.","marker":"[21]"},{"why":"Shows the high-resolution lightweight alternative (IDNet-4) that EDCFlow matches in accuracy with lower compute.","marker":"[39]"},{"why":"Provides the DSEC driving benchmark used for the main accuracy and complexity comparisons.","marker":"[11]"},{"why":"Provides the Blinkflow simulated dataset used to test sim-to-real generalization.","marker":"[19]"},{"why":"Supplies the squeeze-and-excitation channel attention used to fuse difference and correlation motion features.","marker":"[14]"}],"fun_headline_variants":["Dense difference maps refine event flow cheaply","EDCFlow: dense differences beat cost volume alone","Event flow sharpened by dense temporal differences","Lightweight flow via high-res difference features","High-res event flow from dense difference maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that within a short event window (20 ms on DSEC, 10/40 ms on MVSEC) motion is linear, so each intermediate feature map can be warped by a fraction i/g of the final flow; under rotation, acceleration, or occlusion inside that window the warped features misalign and the difference maps encode spurious motion.","fun_headline_variants_meta":{"raw":{"variants":["Dense difference maps refine event flow cheaply","EDCFlow: dense differences beat cost volume alone","Event flow sharpened by dense temporal differences","Lightweight flow via high-res difference features","High-res event flow from dense difference maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1411,"prompt_tokens":915,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":531,"tokens_out":496,"duration_ms":6700,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:00:53.367241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or record an event sequence with known strong rotation or acceleration within the 20 ms window and compare EDCFlow's end-point error against a variant whose intermediate warps use a nonlinear motion model; if the linear-warp version does not degrade on such sequences, the assumption survives, and if it degrades sharply, the central mechanism fails exactly where the assumption breaks.","supporting_citations":[{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the recurrent all-pairs cost-volume and GRU update framework that EDCFlow extends with high-resolution difference maps."},{"cited_title":"E-raft: Dense optical flow from event cam- eras","cited_arxiv_id":null,"evidence_quote":"Establishes the event-based RAFT baseline and the training/evaluation protocol on DSEC and MVSEC."},{"cited_title":"Tma: Temporal motion aggregation for event-based optical flow","cited_arxiv_id":null,"evidence_quote":"Bases the temporally dense cost-volume approach that EDCFlow replaces with cheaper dense feature differences."},{"cited_title":"Lightweight event-based optical flow estimation via iterative deblurring","cited_arxiv_id":null,"evidence_quote":"Shows the high-resolution lightweight alternative (IDNet-4) that EDCFlow matches in accuracy with lower compute."},{"cited_title":"Dsec: A stereo event camera dataset for driv- ing scenarios.IEEE Robotics and Automation Letters, 6(3): 4947–4954, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the DSEC driving benchmark used for the main accuracy and complexity comparisons."},{"cited_title":"Blinkflow: A dataset to push the limits of event-based optical flow estimation","cited_arxiv_id":null,"evidence_quote":"Provides the Blinkflow simulated dataset used to test sim-to-real generalization."},{"cited_title":"Squeeze-and-excitation net- works","cited_arxiv_id":null,"evidence_quote":"Supplies the squeeze-and-excitation channel attention used to fuse difference and correlation motion features."}],"review_version":1}