{"id":"6e37d86d-da70-4948-b212-30dc2ae02d05","arxiv_id":"2507.15150","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A spatiotemporal multigraph with B-spline spatial kernels and motion-vector attention outperforms prior graph-based event-based object detectors on Gen1 and eTraM.","lead":"This paper proposes a graph-based method for object detection from event cameras, using two separate graphs to capture spatial structure and motion. The method reports higher accuracy and lower compute than previous graph-based event detectors, tested on two driving and traffic datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 labels DAGr's values as 'not representative,' and Section 4.2 does not specify the protocol for either DAGr configuration; the headline gain over the SOTA graph baseline is therefore unverified.","rationale":"The architecture is coherent: decoupled spatial and temporal graphs, a 2D anisotropic spline spatial kernel, and motion-vector attention are well motivated, and the ablations in Tables 3, 4, 6, 7, and 8 support those design choices. The central quantitative claim, however, is anchored to the DAGr baseline, and the paper itself attaches a footnote to DAGr's row saying the values are 'not representative for the above comparison.' Section 4.2 describes two DAGr configurations without giving the evaluation protocol (window length, split, training details, variance). This matches the reader's weakest assumption, and I agree with that identification. The 'over 6%' figure is consistent with an absolute 6.7 percentage-point gain over the 0.304 entry, so the issue is not arithmetic; it is that the 0.304 value is disavowed by the footnote. The same flagged row underlies the MFLOPs comparison. Secondary issues include the '5x speedup' claim in Section 4.2, which is attributed to kernel asymptotics rather than a direct timing comparison against AEGNN, and the lack of a described per-event label-assignment protocol for the detection head; both are reproducibility gaps, but the DAGr baseline is the first thing to settle. The reader's CONDITIONAL verdict is appropriate: if the proposed re-run confirms a fair DAGr baseline below 0.311 mAP, the headline claim would be supported and acceptance could follow; if not, the central claim would need substantial revision. Since my analysis reinforces the existing verdict rather than changing it, I recommend UNCHANGED.","tokens_in":16767,"tokens_out":11520,"duration_ms":119927,"concrete_test":"Run DAGr's released code on Gen1 under eGSMV's protocol (100ms windows, same train/val/test split, same mAP computation) for both DAGr configurations, with and without early temporal aggregation to a single voxel. Using the paper's percentage-point convention, if the 'more optimal setting' mAP is at least 0.311, the claimed 'over 6%' advantage is not supported; if it is below 0.311, it is supported. Also report MFLOPs/ev, per-run variance, and whether DAGr uses frames in addition to events, to resolve the 'event+image fusion' caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim is the 'over 6%' gain over DAGr, the state-of-the-art graph baseline. Table 1 lists DAGr mAP as 0.212/0.304 with a footnote stating these values are 'not representative for the above comparison,' and Section 4.2 says only that 0.212 corresponds to no early temporal aggregation and 0.304 to the 'more optimal setting' (early aggregation to a single voxel). Even if 'over 6%' is read as 6.7 percentage points (0.371-0.304), the baseline itself is disavowed by the footnote, and the re-run protocol is not given: no window length, no test-split details, no DAGr training recipe, no variance. Because DAGr is the comparison target in the headline, an unreliable DAGr number can shrink or invert the claimed advantage. The same uncertainty affects the efficiency comparison, since DAGr's MFLOPs/ev (6.27/4.58) are part of the flagged row. Section 4.2 also notes DAGr is an event+image fusion technique, so it is unclear whether this is even a like-for-like graph comparison. This is the load-bearing weak point: if the DAGr baseline is not trustworthy, the quantitative claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes eGSMV, an event-based object detection method that represents an event stream as a spatiotemporal multigraph with separate spatial and temporal edges. Spatial structure is learned with an anisotropic 2D spline convolution; temporal structure is learned with motion-vector features and GATv2 attention on a separate temporal neighborhood. The authors report 0.371 mAP on Gen1 and 0.431 mAP@50 on eTraM, claiming an over-6% accuracy gain over the graph-based DAGr baseline, lower MFLOPs per event than AEGNN and DAGr, and a 5x speedup from quadratic rather than cubic kernels. Ablations examine the SSL/MVL split, detection-head granularity, graph length, and graph density.","tokens_in":17107,"tokens_out":6405,"duration_ms":62938,"significance":"If the baseline comparisons were fully specified and reliable, the contribution would be useful: the decoupled spatial/temporal multigraph is a sensible way to preserve sparsity and asynchronous operation while avoiding cubic kernels, and the ablations (Tables 3, 4, 6, 8; Figure 7) give internal evidence that both modules contribute. The paper also candidly discusses storage and hardware limitations in Section 5. At present, however, the headline accuracy and speed claims rest on baselines that the paper itself disavows or does not measure, so the quantitative contribution is not yet established.","major_comments":[{"comment":"The headline 'over 6%' improvement over DAGr is not supported by the data as presented. Table 1 flags the DAGr row (0.212/0.304 mAP, 6.27/4.58 MFLOPs/ev, 34.6M params) as 'not representative for the above comparison,' and Section 4.2 gives no protocol for either re-run configuration: no graph or time-window length, no early-aggregation details, no training recipe, and no test-split or variance information. Because the DAGr numbers are the comparison target of the central accuracy claim, and the paper also notes DAGr is an event+image fusion technique, the 6% gain cannot be verified or interpreted as a like-for-like graph-based gain. Please provide a complete, reproducible DAGr re-evaluation (or use published numbers with explicit provenance) and restate the claim with the appropriate uncertainty.","section":"Section 4.2, Table 1"},{"comment":"The reported eTraM improvement over AEGNN is inconsistent. The abstract and introduction state a 25% improvement, while Section 4.2 states mAP 35% higher using a re-run AEGNN benchmark 'under the same evaluation settings as our method,' but Table 1 lists AEGNN at 0.180 mAP@50 and no separate re-run result or protocol is given. With the printed numbers, eGSMV's 0.431 exceeds 0.180 by 0.251 points (139% relative), so neither 25% nor 35% is self-consistent without additional context. Please report the exact AEGNN configuration, the resulting number, and state whether percentages are absolute points or relative gains.","section":"Section 4.2 and Abstract/Introduction"},{"comment":"The abstract's '5x speedup' is not measured. Table 2 compares a 'dense graph update' to the proposed serial and parallel asynchronous updates; it does not compare against AEGNN or DAGr. The claim that 'our quadratic kernels enable more than 5x speedup compared to AEGNN' appears in the text without any AEGNN timing experiment. Please either add direct wall-clock latency measurements against AEGNN (and DAGr) under the same hardware and graph sizes, or revise the speedup claim to a kernel-complexity bound that is explicitly labeled as such.","section":"Section 4.2, Timing Experiments and Table 2"}],"minor_comments":[{"comment":"The statement 'our method achieves a 21% improvement over AEGNN on Gen1' should be qualified: Table 1 gives AEGNN 0.163 and eGSMV 0.371, a 0.208 mAP-point gap (128% relative), so if the intended meaning is 21 percentage points, say so explicitly.","section":"Section 1 and Table 1"},{"comment":"The classification and confidence losses use inconsistent notation for predictions and ground truth (y_i vs ŷ_i, and s_i vs ŝ_i); please align with a standard convention.","section":"Section 8.1, Equations (9) and (12)"},{"comment":"It is unclear whether the coordinates in the neighbor condition are raw pixel coordinates, normalized coordinates, or a mix; please define the units and normalization for R_XY, R_t, and the edge features in Equations (5) and (6).","section":"Section 3.1, Equation (2)"},{"comment":"Table 1 uses '*' for estimated values and '^' for non-representative values, but the DAGr MFLOPs/ev entries are inside the flagged row; please state explicitly that the efficiency comparison is likewise affected.","section":"Table 1 footnote"},{"comment":"Figure 5 is referenced in the text but not described; consider adding a sentence summarizing what it shows.","section":"Section 4.2 and Figure 5"},{"comment":"The text refers to 'Table??' when discussing multi-head attention; please fix the reference.","section":"Section 10.1, Table 8"}],"recommendation":"major_revision","confidential_remarks":"The architecture and ablations are interesting, but the quantitative claims need substantial work: the DAGr baseline is disavowed in Table 1, the eTraM gain is reported inconsistently, and the speedup is not directly measured. I would invite a revision with fully specified re-run baselines, consistent metrics, and either direct timing measurements or a clearly labeled complexity bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the architecture is genuinely new and the ablations are solid, but the headline accuracy gain over DAGr rests on a baseline the authors themselves disavow in a table footnote. That makes the central quantitative claim unverified, though the paper is still worth reading and refereeing.\n\nWhat's new and good: the decoupled spatial and temporal graphs with anisotropic 2D B-spline kernels and motion-vector attention are a clean way to avoid 3D convolutions, and the event-level detection head keeps the pipeline sparse end to end. The ablations back the design choices: the fused model beats each branch alone, 2D spline beats GCN, and the node-wise head beats fixed grids. These are real measurements, and the supplementary adds useful sweeps over graph length density and kernel choice. The authors also honestly list the storage bottleneck for graph methods in the limitations section.\n\nThe soft spots are quantitative and all in the comparison against prior work. Table 1 footnotes DAGr's numbers with \"not representative,\" and Section 4.2 reports two DAGr configurations without giving the protocol for either re-run. So the \"over 6%\" gain is against a baseline the authors themselves distrust. DAGr is also an event+image fusion method, so the graph-only comparison is muddled. The eTraM improvement is 25% in the introduction and 35% in Section 4.2, which is an inconsistency that needs fixing. And the 5x speedup over AEGNN is not directly measured—Table 2 compares against a dense graph update, so the 5x appears to be a kernel complexity argument, not a runtime comparison.\n\nThese are fixable problems, not fatal ones. The architecture is sound, and the paper's target audience—people working on event-based detection and graph learning—will get useful ideas from the ablations alone. But as it stands, the central claim is not established beyond reasonable doubt. I'd send it to a serious referee, with the expectation of heavy revision: the authors need to describe the DAGr re-run in full, report variance, reconcile the eTraM percentages, and provide a measured runtime comparison for the 5x speedup.\n\nRecommendation: send to peer review, but push for a revision that nails down the baseline protocol before you trust the numbers.","headline":"The architecture is new and the ablations are careful, but the headline DAGr gain rests on a baseline the authors themselves flag as not representative.","tokens_in":17630,"tokens_out":4595,"would_cite":true,"duration_ms":43975,"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":"Decoupled spatial and temporal event graphs improve graph-based event-camera object detection by over 6% with no added computation.","keywords":["event-based vision","event camera","graph neural networks","object detection","asynchronous inference","spatiotemporal graph","B-spline convolution","motion vector attention"],"falsifier":"Re-run DAGr, AEGNN, and eGSMV under identical protocols with the same 100 ms windows, same training budget, and same detection head, and publish both mAP and per-event FLOPs; if DAGr under that protocol reaches or exceeds 0.371 mAP on Gen1, the claimed over-6% advantage would not hold.","tokens_in":16587,"feed_emoji":"⚡","tokens_out":7900,"duration_ms":75676,"temperature":0.7,"pith_summary":"Event cameras emit asynchronous pixel-level brightness changes, and converting those events into dense image-like tensors for standard networks destroys the sparsity and low latency that make event cameras useful. This paper proposes representing an event stream as two decoupled graphs: one encoding spatial structure with B-spline kernels and one encoding motion through attention over velocity-like edge features, so the network can stay in the sparse, asynchronous domain. The central claim is that this multigraph design, called eGSMV, detects objects better than previous graph-based event detectors while using fewer parameters and no more computation, reporting 0.371 mAP on Gen1, 0.431 mAP@50 on eTraM, and over 6% higher accuracy than DAGr, the prior graph-based state of the art. If correct, this shows that explicit modeling of spatial and temporal structure, rather than deeper dense backbones, is the main lever for making event-camera perception accurate and efficient.","feed_headline":"Two event graphs lift event-camera detection by 6%","feed_subtitle":"Separating spatial and motion graphs keeps detection sparse, asynchronous, and 5x faster than prior graph methods","key_machinery":"The load-bearing object is the decoupled spatiotemporal multigraph $G = \\{V, E_s, R_s, E_t, R_t\\}$: each event is a node in $\\mathbb{R}^3$ at $(x_i, y_i, t_i)$, with spatial edges defined by an ellipsoid whose semi-major axis lies in the XY-plane (4% of input size, 5 ms temporal radius) and temporal edges by an inverted ellipsoid (1% spatial, 40 ms temporal). The spatial view is learned with an anisotropic B-spline kernel of size $k \\times k \\times 1$, which is a 2D operation, and the temporal view is learned by GATv2-style multi-head attention over motion-vector edge features. This separation is what lets the network capture both global spatial structure and local motion while keeping all computation in the sparse, asynchronous regime; fusing the two views is deliberately a cheap concatenation plus MLP.","core_discovery":"The paper's central claim is that an event stream is better modeled as a spatiotemporal multigraph with two independent neighbor structures than as a single spatiotemporal graph or a dense tensor. The spatial subgraph connects nearby events within a tight ellipsoid and is processed by a 2D anisotropic B-spline convolution that captures local structure, while the temporal subgraph connects events that are close in space but separated in time and is processed by multi-head attention over edge features that include displacement, velocity ($\\Delta x/\\Delta t$, $\\Delta y/\\Delta t$), and polarity change. Because the two views are decoupled, message passing computes quadratic rather than cubic relations, and 2D kernels replace 3D kernels, cutting per-message-passing cost by up to 87.5%. With an event-level detection head that pools only proximal events before non-maximum suppression, the framework reports 0.371 mAP on Gen1 and 0.431 mAP@50 on eTraM at 4.5 and 26.1 MFLOPs per event, with 5.6M parameters, outperforming AEGNN by 21% and DAGr by over 6% without additional computation.","pith_inferences":["A direct test of the paper's central mechanism would swap the 2D spatial kernel for the 3D isotropic kernel at a matched parameter budget; the paper's own ablation shows 3D gains only 0.02 mAP at about six times the parameters, but that ablation is on a validation set with shorter sequences.","The authors leave open whether the same decoupling helps other event-level tasks; since the temporal edge features are essentially velocity and displacement, optical flow and motion segmentation are natural next targets.","The claim of no additional computational requirement is relative to graph baselines; comparing against a dense method at equal per-event FLOPs would test whether the sparse representation itself, rather than the efficiency trick, is the source of the accuracy gain."],"forward_implications":["Graph-based event detectors can match or narrow the gap to dense transformer-based detectors while using roughly three orders of magnitude fewer FLOPs per event, e.g., 4.5 versus about 2400 on Gen1.","Because the spatial kernel is 2D and the two graph views are independent, the spatial and temporal learning blocks can run in parallel, giving close-to-constant update time as the graph grows (21.4 ms at 25,000 nodes) and a 5x speedup over AEGNN's cubic kernels.","Retaining event-level granularity in the detection head matters: coarsening predictions to voxel grids of size 8x8 and above drops mAP from about 0.36 to 0.197 on the Gen1 validation set.","The design generalizes beyond ego-motion driving scenes to static traffic monitoring, where eGSMV reaches 0.431 mAP@50 on eTraM with a fraction of dense methods' parameters."],"supporting_citations":[{"why":"AEGNN, the asynchronous graph baseline the paper extends; supplies the event-graph update paradigm and the main accuracy/compute comparison point.","marker":"[49]"},{"why":"DAGr, the prior state-of-the-art graph-based detector the paper claims to beat by over 6%; supplies the deeper lookup-table message-passing baseline.","marker":"[21]"},{"why":"Gen1 Automotive Detection dataset, the primary benchmark; supplies the ego-motion event data and object labels for evaluation.","marker":"[13]"},{"why":"eTraM dataset, the second benchmark; supplies static-perspective traffic monitoring events and 8-class labels used for generalization.","marker":"[53]"},{"why":"RVT, the strongest dense transformer+RNN comparator; establishes the accuracy level and the much higher per-event FLOP cost that eGSMV contrasts against.","marker":"[22]"},{"why":"GATv2 attention; supplies the multi-head attention mechanism used in the motion vector learning block.","marker":"[8]"},{"why":"YOLOX; supplies the detection-head design, including classification, regression, and IoU confidence branches, which eGSMV adapts to event-level predictions.","marker":"[20]"}],"fun_headline_variants":["Dual event graphs lift detection 6%, run 5x faster","Spatial + motion event graphs beat prior methods by 6%","Event multigraph: 6% accuracy gain, 5x speedup","Decoupled event graphs outdo dense tensors for async detection","Two-graph event model improves detection by 6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline improvement over DAGr rests on the assumption that the DAGr numbers in Table 1 fairly represent that method under comparable settings, but the table itself footnotes that those values are 'not representative for the above comparison'.","fun_headline_variants_meta":{"raw":{"variants":["Dual event graphs lift detection 6%, run 5x faster","Spatial + motion event graphs beat prior methods by 6%","Event multigraph: 6% accuracy gain, 5x speedup","Decoupled event graphs outdo dense tensors for async detection","Two-graph event model improves detection by 6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1393,"prompt_tokens":994,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":610,"tokens_out":399,"duration_ms":4565,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:38:31.534835+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run DAGr, AEGNN, and eGSMV under identical protocols with the same 100 ms windows, same training budget, and same detection head, and publish both mAP and per-event FLOPs; if DAGr under that protocol reaches or exceeds 0.371 mAP on Gen1, the claimed over-6% advantage would not hold.","supporting_citations":[{"cited_title":"Aegnn: Asynchronous event-based graph neural networks","cited_arxiv_id":null,"evidence_quote":"AEGNN, the asynchronous graph baseline the paper extends; supplies the event-graph update paradigm and the main accuracy/compute comparison point."},{"cited_title":"Low-latency auto- motive vision with event cameras.Nature, 629(8014):1034– 1040, 2024","cited_arxiv_id":null,"evidence_quote":"DAGr, the prior state-of-the-art graph-based detector the paper claims to beat by over 6%; supplies the deeper lookup-table message-passing baseline."},{"cited_title":"A large scale event-based detec- tion dataset for automotive, 2020","cited_arxiv_id":null,"evidence_quote":"Gen1 Automotive Detection dataset, the primary benchmark; supplies the ego-motion event data and object labels for evaluation."},{"cited_title":"etram: Event-based traffic monitoring dataset","cited_arxiv_id":null,"evidence_quote":"eTraM dataset, the second benchmark; supplies static-perspective traffic monitoring events and 8-class labels used for generalization."},{"cited_title":"Recurrent vision transformers for object detection with event cameras","cited_arxiv_id":null,"evidence_quote":"RVT, the strongest dense transformer+RNN comparator; establishes the accuracy level and the much higher per-event FLOP cost that eGSMV contrasts against."},{"cited_title":"How attentive are graph attention networks? InInternational Conference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"GATv2 attention; supplies the multi-head attention mechanism used in the motion vector learning block."}],"review_version":1}