{"id":"1040f02c-a0d2-4498-81f3-c0383a7167ff","arxiv_id":"2501.11971","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SMamba prunes uninformative event tokens using a spatiotemporal continuity score, then scans the survivors with Mamba blocks to match or beat prior event detectors at lower compute.","lead":"This paper introduces SMamba, an event-camera object detector that drops empty and noisy image regions before processing while keeping global context through Mamba-style scanning. If the reported results hold, it offers a better speed-accuracy trade-off than current event-based detectors on three driving and traffic benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central accuracy-efficiency claim is undercut by the paper's own runtime table: SMamba is ~2x slower than RVT-B despite lower FLOPs, so 'superior efficiency' is not established.","rationale":"The reader's weakest assumption (STCA continuity prior) is plausible and partially tested by the ablations, but the more immediately load-bearing problem is the internal inconsistency between the FLOPs-based efficiency narrative and the reported runtimes. SMamba is consistently about 2x slower than RVT-B on all three datasets despite lower backbone FLOPs, and the sparsification module produces only marginal wall-clock savings. This attacks the 'efficiency' half of the central accuracy-efficiency claim without depending on any assumption about noise statistics. The conditional verdict remains appropriate: the accuracy numbers and ablations are valuable, but the paper should either provide a fair latency benchmark with total FLOPs or substantially soften the efficiency claim. No change to the verdict category is needed because the concern is addressable through additional experiments and reporting.","tokens_in":12149,"tokens_out":6456,"duration_ms":67137,"concrete_test":"Run a controlled end-to-end benchmark on the same GPU (e.g., A100 or RTX 3090) with identical batch size, resolution, and framework (PyTorch) for SMamba, RVT-B, GET-T, SAST-CB, and S5-ViT-B on Gen1/1Mpx/eTram. Measure wall-clock latency per sample, backbone-only FLOPs and total FLOPs. Also ablate SMamba's sparsification on/off to measure actual wall-clock savings. If SMamba's latency is not lower than RVT-B/SAST-CB, the efficiency claim should be downgraded and the abstract/conclusion revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is a better accuracy-efficiency trade-off, but Table 1's runtime column contradicts the efficiency half. On Gen1, SMamba achieves 50.4 mAP at 2.4G backbone FLOPs, but its end-to-end runtime is 24.0 ms, versus 10.2 ms for RVT-B (47.2 mAP, 3.5G FLOPs) and 9.4 ms for S5-ViT-B (47.7 mAP, >3.1G). On 1Mpx, SMamba runs 26.0 ms vs RVT-B's 11.9 ms; on eTram, 25.2 ms vs 11.9 ms. Sparsification itself saves 23-31% FLOPs but only 4-8% runtime (25.2 to 24.0 ms on Gen1; 27.5 to 26.0 on 1Mpx; 27.5 to 25.2 on eTram). If wall-clock latency is the relevant efficiency metric for event cameras, SMamba's sparse scanning and ConvLSTM overhead negate the FLOPs advantage, so the central trade-off claim is not supported by the reported data. The FLOPs numbers are backbone-only, while runtime includes the full detector; the paper does not report total FLOPs or hardware details. This is load-bearing independent of whether the STCA scoring prior holds: even granting that sparsification improves accuracy, the 'superior efficiency' conclusion requires a fair end-to-end latency comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SMamba, a Mamba-based backbone for event-based object detection that combines a Spatio-Temporal Continuity Assessment (STCA) module for token sparsification, an Information-Prioritized Local Scan (IPL-Scan) for spatial scanning, and a Global Channel Interaction (GCI) module for channel-wise modeling. The method is evaluated on Gen1, 1Mpx, and eTram, and the paper claims superior accuracy and efficiency over existing CNN, Transformer, and SSM baselines, with code released. The main evidence is the mAP/FLOPs/parameter/runtime tables plus ablations of the scoring module, scanning strategy, and GCI placement.","tokens_in":12454,"tokens_out":5813,"duration_ms":60244,"significance":"If the accuracy-efficiency trade-off claimed in Tables 1 and 2 is reliable, SMamba is a practically interesting design: it retains global spatial and channel modeling while using sparsification to reduce backbone FLOPs, and it ships code and evaluates on three datasets, including a less-studied traffic-monitoring dataset. The ablations are informative and the visualizations support the claim that the sparsification map is scene-adaptive. However, the central efficiency claim is currently weakened by the reported end-to-end latencies, and the numerical margins over the authors' own baseline are small enough that multi-seed statistics are needed.","major_comments":[{"comment":"The central accuracy-efficiency claim is not supported by the reported runtime. SMamba's end-to-end latency is 24.0 ms on Gen1, 26.0 ms on 1Mpx, and 25.2 ms on eTram, versus 10.2/11.9/11.9 ms for RVT-B and 9.4/10.9/10.9 ms for S5-ViT-B; the sparsification itself reduces latency by only 4-8% relative to the non-sparse baseline (25.2->24.0, 27.5->26.0, 27.5->25.2). Because the FLOPs numbers are backbone-only, the statement that SMamba achieves 'superior performance and greater efficiency' conflates backbone complexity with end-to-end cost. Please report end-to-end FLOPs, hardware/software details, batch size, measurement repetitions, and latency under identical conditions, and soften the efficiency claim accordingly.","section":"Quantitative Results, Tables 1 and 2"},{"comment":"The score S_t in Eq. (4) sums absolute event timestamps, which is not a temporal continuity measure. It is not invariant under shifting the event stream in time and it grows monotonically with event time, so a single isolated noise event at a late timestamp can score higher than a dense train of early activity events. This undermines the premise of STCA. Please replace it with a genuine continuity statistic such as event count in a short sliding window, inverse inter-event interval, or local event rate, and validate the chosen statistic against the activity/noise distinction.","section":"Eq. (4), Spatio-Temporal Continuity Assessment"},{"comment":"The headline gains are small: 0.3-0.5 mAP over the authors' own baseline, 0.6 mAP over SAST-CB on 1Mpx, and 2.6 mAP on eTram, yet no error bars, number of seeds, or significance tests are reported. Without these, it is not possible to determine whether the claimed improvements are beyond run-to-run variation. Please report mean plus/minus standard deviation over at least three seeds for the main comparisons and ablations, and indicate statistical significance for the differences that drive the conclusions.","section":"Tables 1-6, Experimental Setup"},{"comment":"The threshold in Eq. (6) is alpha = mean(S_st)/beta, with beta manually adjusted, and the Gaussian variance sigma in Eq. (5) is also a free parameter; no sensitivity analysis is provided for either. Since the sparsification ratio and the accuracy-efficiency trade-off are direct functions of beta, the paper should report mAP, FLOPs, and latency at several beta and sigma values; otherwise the chosen operating point cannot be judged robust.","section":"Eqs. (5)-(6), Implementation Details"}],"minor_comments":[{"comment":"The phrase 'spare event data' should be 'sparse event data'.","section":"Introduction"},{"comment":"The definition 'FLOPs (Floating Point Operations Per Second)' is incorrect; FLOPs is the number of floating-point operations, not a rate. Please correct the definition.","section":"Experimental Setup, Metrics"},{"comment":"The notation is inconsistent: S_t is used for both the full-resolution map and the pooled map, and Eq. (5) writes S_t^q for neighbors after the text describes aggregation in pixel coordinates. Please clarify the indexing.","section":"Spatio-Temporal Continuity Assessment"},{"comment":"The limitation statement says the method assumes 'temporal and spatial independence of event noise,' but Eqs. (4)-(7) assume noise is spatially isolated and temporally discontinuous, not statistically independent. Please align the wording with the actual assumption.","section":"Limitation"}],"recommendation":"major_revision","confidential_remarks":"The efficiency discrepancy between backbone FLOPs and end-to-end latency is the main obstacle to acceptance; the authors should provide a complete accounting of total FLOPs and latency and multi-seed statistics before resubmission. The proposed architecture is reasonable and the ablations are useful, but the accuracy-efficiency trade-off as stated is not yet established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nSMamba is a serious attempt to bring Mamba-style SSMs to event-based detection with adaptive token sparsification. The new pieces are the spatio-temporal continuity score (STCA), the information-prioritized local scan (IPL-Scan), and the global channel interaction (GCI) module. The ablation study on eTram isolates each component's contribution, which is more than many architecture papers do. The FLOPs numbers do drop by 22-31% against their own baseline, and the mAP is competitive on three datasets. That part is solid and reproducible in principle.\n\nThe soft spot is the efficiency claim. The abstract says the method outperforms others in both performance and efficiency, but Table 1 shows end-to-end runtime is 24.0 ms on Gen1 versus 10.2 ms for RVT-B and 9.4 ms for S5-ViT-B. The FLOPs are backbone-only and lower, but for real-time event cameras wall-clock latency matters. The paper does report runtime honestly, and the text says the speed is 'comparable with SAST-CB', which is fair, but the abstract's 'superior efficiency' is an overstatement. Sparsification saves only 4-8% runtime in their own numbers; the ConvLSTM and scan overhead eat the FLOPs advantage. This does not falsify the accuracy comparisons, but it does mean the central trade-off claim is only half-supported.\n\nThe second issue is reproducibility. Eq. (4) defines the temporal continuity score as the sum of timestamps, which is more a measure of total event activity than continuity. It works in practice, but the naming oversells the concept. More importantly, the threshold beta in Eq. (6) is never reported, and there are no error bars or multiple seeds. The differences between SMamba and the next-best methods are 0.3-0.5 mAP in places, which could vanish with a different seed.\n\nThe paper also has a self-aware limitation section that admits the STCA prior may fail on large noise bursts. That is a good sign. The related work is thorough, and citation choices are appropriate, not self-citation heavy.\n\nWho reads this? Researchers working on efficient event-based detection or vision Mamba backbones will want to see the STCA and IPL-Scan ideas. The runtime issue should be fixed with a full-model FLOPs/latency table and a reported beta before anyone relies on the efficiency claim. I would give it a serious referee if I were the editor, but I would ask for a revised efficiency comparison and hyperparameter settings.\n\nRecommendation: engage, but treat the abstract's efficiency sentence with caution until the runtime numbers are explained.","headline":"Worth a look for the STCA and IPL-Scan ideas, but the paper's 'superior efficiency' claim is undercut by its own runtime column: SMamba is about 2x slower than RVT-B in end-to-end latency.","tokens_in":13004,"tokens_out":2721,"would_cite":false,"duration_ms":27383,"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":"SMamba shows event-camera detection can match transformer accuracy at a fraction of the compute by scanning only spatiotemporally continuous tokens.","keywords":["event-based object detection","state space models","selective scan","token sparsification","spatiotemporal continuity","event cameras","Mamba backbone"],"falsifier":"Inject spatially clustered, temporally continuous synthetic noise into eTram test events and compare SMamba's mAP against its no-sparsification baseline: if the accuracy gap widens as burst noise increases, or if a measurable share of ground-truth object tokens falls below the continuity threshold, the assumption that noise is isolated and discontinuous is refuted.","tokens_in":11947,"feed_emoji":"⚡","tokens_out":8493,"duration_ms":77650,"temperature":0.7,"pith_summary":"Event cameras record only brightness changes, so most of a scene is silent and much of the signal is noise; transformer detectors that process every token uniformly waste computation on these uninformative regions. This paper sets out to show that a Mamba-style detector can exploit that sparsity without giving up global modeling. Its SMamba architecture scores tokens by the spatiotemporal continuity of the events inside them, discards low-scoring tokens, reorders the survivors so informative tokens are scanned close together, and adds global channel interaction. On Gen1, 1Mpx, and eTram it reports the best or tied mAP among the compared methods with backbone FLOPs reduced by 22–31% against its own no-sparsification baseline. The result matters because it suggests event-camera detection can get transformer-level accuracy at a fraction of the compute.","feed_headline":"Sparse Mamba cuts event-detector cost while matching or beating accuracy","feed_subtitle":"It trims 22–31% of backbone FLOPs on three event-camera datasets and still edges out transformer baselines.","key_machinery":"The load-bearing mechanism is the Spatio-Temporal Continuity Assessment (STCA) module, which computes a per-token information score by accumulating event timestamps at each pixel, average-pooling to token resolution, and smoothing with a Gaussian neighborhood weight; the threshold $\\alpha = \\frac{\\sum S_{st}}{\\beta\\, H W / P^2}$ then converts scores into a binary sparsification map. This map drives two further components. The Information-Prioritized Local Scan (IPL-Scan) sorts token windows by their maximum continuity score and scans them with S6 blocks, so high-information tokens have short scan distances while local neighbors stay together; Global Channel Interaction (GCI) flattens the feature map along spatial dimensions and runs bidirectional channel scans through S6 plus a $1\\times 1$ convolution, letting each channel absorb global spatial context when interacting with other channels. Together these modules let a sparse set of tokens still receive global spatial and channel modeling.","core_discovery":"The central claim is that the spatiotemporal distribution of events alone is enough to identify informative tokens, and that an information-prioritized selective scan can keep global context while ignoring noise and empty regions. Concretely, SMamba reaches 50.4 mAP on Gen1 at 2.4G backbone FLOPs and 16.1M parameters, matching ERGO-12's mAP with about 5% of its FLOPs, and 49.3 on 1Mpx at 7.4G FLOPs. On eTram it reports 32.6 mAP at 6.6G FLOPs, beating the transformer baselines compared. In all three datasets the sparsified model improves on its own non-sparse baseline in both accuracy and cost, which the paper presents as evidence that discarding non-continuous event tokens is not just a compute saving but also a noise-suppression mechanism.","pith_inferences":["Editorial inference: the STCA scoring is attached to the raw event stream rather than to learned features, so the same module could be lifted into other event-based tasks such as segmentation or tracking, provided the continuity prior holds.","Editorial inference: the paper's stated limitation—noise bursts or clusters violate the independence assumption—points to a concrete extension: a learned or denoised continuity score that distinguishes semantically continuous noise from object motion, which would make the sparsification robust in crowded scenes.","Editorial inference: because FLOPs are reported for the backbone and runtime for one platform, the practical efficiency claim would be strengthened by measuring energy or latency on event-camera hardware, where sparse scanning may or may not translate to real speedups.","Editorial inference: the threshold formula ties the discarding rate to a manually tuned beta; a sensitivity analysis across beta and scene sparsity levels would reveal how much of the reported trade-off depends on this hyperparameter."],"forward_implications":["A compact 16–17M parameter detector can match or beat transformer baselines on event-camera detection, so global modeling does not require the full compute of self-attention.","Sparsification guided by a hand-crafted continuity score beats learned scoring (variance, entropy, and the learnable scoring module) in the reported comparisons, suggesting the prior itself carries much of the signal.","Combining Bidi-Scan with IPL-Scan improves mAP by 2.4 points over either alone on eTram, indicating that global scan order and information-prioritized reordering are complementary rather than redundant.","Placing global channel interaction only in the last two stages gives the best accuracy at the lowest cost, so 3D global modeling matters most when features are semantically rich."],"supporting_citations":[{"why":"Supplies the S6 selective state-space scan that forms the backbone of SMamba.","marker":"Gu and Dao 2023"},{"why":"Provides the RVT baseline and the preprocessing, augmentation, and evaluation protocols reused for fair comparison.","marker":"Gehrig and Scaramuzza 2023"},{"why":"SAST is the main sparse-transformer baseline and source of the FLOPs measurement protocol.","marker":"Peng et al. 2024"},{"why":"VMamba's VSS block is used to construct the SSM baseline and informs the scan designs.","marker":"Liu et al. 2024a"},{"why":"Gen1 is the primary automotive detection dataset for the main comparison.","marker":"De Tournemire et al. 2020"},{"why":"Supplies the 1Mpx dataset and the RED CNN baseline.","marker":"Perot et al. 2020"},{"why":"Provides the eTram traffic-monitoring dataset used for the sparsest evaluation.","marker":"Verma et al. 2024"},{"why":"Defines the voxel event representation used to tokenize the event stream.","marker":"Zhu et al. 2019"}],"fun_headline_variants":["SMamba: sparse scan trims FLOPs, beats transformers","Sparse Mamba event detector: less compute, better accuracy","Event-camera detection: SMamba sparsifies to win mAP and cost","SMamba drops noise tokens, cuts FLOPs, edges out baselines","Sparse Mamba nets higher mAP with fewer FLOPs on events"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that true objects produce events that are spatially close and temporally continuous while noise is spatially isolated and temporally discontinuous, so the continuity score can reliably tell them apart; as the paper's own limitation notes, large bursts or clusters of noise break this assumption and can cause informative tokens to be discarded.","fun_headline_variants_meta":{"raw":{"variants":["SMamba: sparse scan trims FLOPs, beats transformers","Sparse Mamba event detector: less compute, better accuracy","Event-camera detection: SMamba sparsifies to win mAP and cost","SMamba drops noise tokens, cuts FLOPs, edges out baselines","Sparse Mamba nets higher mAP with fewer FLOPs on events"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1297,"prompt_tokens":965,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":581,"tokens_out":332,"duration_ms":4553,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:38:59.761841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inject spatially clustered, temporally continuous synthetic noise into eTram test events and compare SMamba's mAP against its no-sparsification baseline: if the accuracy gap widens as burst noise increases, or if a measurable share of ground-truth object tokens falls below the continuity threshold, the assumption that noise is isolated and discontinuous is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RVT baseline and the preprocessing, augmentation, and evaluation protocols reused for fair comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SAST is the main sparse-transformer baseline and source of the FLOPs measurement protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 1Mpx dataset and the RED CNN baseline."},{"cited_title":"A.; Chakravarthi, B.; Vaghela, A.; Wei, H.; and Yang, Y","cited_arxiv_id":null,"evidence_quote":"Provides the eTram traffic-monitoring dataset used for the sparsest evaluation."},{"cited_title":"Z.; Yuan, L.; Chaney, K.; and Daniilidis, K","cited_arxiv_id":null,"evidence_quote":"Defines the voxel event representation used to tokenize the event stream."}],"review_version":1}