Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

FocusMamba claims that event-camera motion signals can guide adaptive token sparsification for RGB-event object detection, cutting computation while improving accuracy over fixed-rate pruning and existing fusion methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 10:34 UTC pith:QDCBQGKW

load-bearing objection Event-driven token sparsification is a genuine new idea and the accuracy gains look real, but the paper overclaims efficiency by leaning on FLOPs while its own runtime numbers show FocusMamba is 22–26% slower than SFNet. the 3 major comments →

arxiv 2509.03872 v1 pith:QDCBQGKW submitted 2025-09-04 cs.CV

Focus Through Motion: RGB-Event Collaborative Token Sparsification for Efficient Object Detection

Nan Yang , Yang Wang , Zhanwen Liu , Yuchao Dai , Yang Liu , Xiangmo Zhao This is my paper
classification cs.CV
keywords RGB-event fusionobject detectiontoken sparsificationevent camerasMambavision state space modeladaptive computationmultimodal fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

FocusMamba tries to show that the motion signal from an event camera can drive where a multimodal object detector spends computation. Instead of pruning tokens by a fixed count or threshold, it uses the fraction of pixels that trigger events to adaptively sharpen token scores and set the keep threshold separately for each sample and each modality. The retained tokens then double as a guide for fusing RGB and event features, so fusion attends only to complementary regions instead of all background. On DSEC-Det and PKU-DAVIS-SOD, FocusMamba-B reports higher mAP than the previous best fusion method while using roughly a quarter to a third of its FLOPs. If the claim holds, event-guided sparsification is a generally applicable way to make RGB-event perception cheaper without sacrificing accuracy.

Core claim

On the paper's own terms, the central discovery is that the event spatial ratio r—the share of pixels that produced events in a frame interval—can serve as a per-sample control signal for multimodal token pruning. EGMS computes independent importance scores for image tokens (L2 activation) and event tokens (spatiotemporal continuity), then rescales those scores and the selection threshold with monotonic functions of r. The same sparsification maps are reused by the CMFF module, which enhances each modality in regions the other modality sees and fuses the union of kept tokens through bidirectional Mamba scans. The effect is that fixed-rate methods' failures—under-sparsifying simple scenes and

What carries the argument

Event-Guided Multimodal Sparsification (EGMS): statistical token scoring (L2 norm for RGB, spatiotemporal event continuity for events) plus an Event-Guided Control Mechanism in which scale factor = r^(1/ρ) sharpens softmax score differences and control factor = (1-r)^(1/ρ) sets threshold α = 1/(N·Control), producing binary sparsification maps M_I and M_E. Cross-Modality Focus Fusion (CMFF): Complementarity-Aware Enhancement boosts each modality in map-difference regions; Focused Interlaced Mamba gathers the union of kept tokens, interleaves them, and runs bidirectional selective scanning so long-range dependencies are modeled only among informative tokens.

Load-bearing premise

The load-bearing premise is that the event spatial ratio r—the fraction of pixels that fire events—is a trustworthy measure of how much and where object information exists; when r is inflated by ego-motion or noise, or made uninformative by a static scene, the adaptive scale and threshold will keep the wrong tokens.

What would settle it

A controlled experiment on a sequence where the camera moves but objects are stationary: measure r and the kept-token ratio; if the kept ratio rises while mAP on static objects falls below a fixed-rate baseline, the claim that r tracks object information is falsified. Similarly, a static scene with r≈0 should leave the control factor at 1 and degenerate to non-adaptive pruning, a failure the paper's own Limitation acknowledges.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Adaptive retention beats fixed-rate pruning: compared with applying EGMS's average kept rate to every sample, EGMS raises mAP50 by 2.3 points at the same FLOPs (53.0/32.8 vs 55.3/34.6).
  • Sparsification maps are not only a compute-saving device; they provide the fusion module with complementary-region cues, since the CAE module alone adds 1.2 mAP with no measurable FLOP increase.
  • The design transfers beyond Mamba: EGMS applied to a MaxViT backbone cuts FLOPs by about 45.5% while improving mAP by 0.5 on DSEC-Det.
  • Incorporating EGMS into the VSS baseline reduces FLOPs by roughly 30% and runtime by 7–10% on both datasets while improving mAP, so the savings come without accuracy loss.
  • Against the strongest fusion baseline (SFNet), FocusMamba-B reports +4.2 mAP on DSEC-Det and +0.8 mAP on PKU-DAVIS-SOD while using 29.1% and 22.2% of its FLOPs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: the same r-controlled threshold could be plugged into any token-pruning detector, not just Mamba backbones; the paper's MaxViT experiment already hints at this generality, but a direct comparison on a transformer detector with fixed-rate pruning would make the causal role of r explicit.
  • The paper treats r as a scalar proxy; an obvious refinement is a per-region or per-stage event density map, which would let the control mechanism adapt spatially instead of globally and might fix the static-scene failure mode the authors acknowledge.
  • Because scale and control are deterministic functions of r, one can precompute the exact mapping from r to kept ratio; a cheap follow-up is calibrating ρ per dataset or per scene category rather than as a single global hyperparameter.
  • The monotonic dependence on r also predicts a specific failure: ego-motion-only scenes with moving backgrounds will inflate r and retain too many background tokens, so detection of small static objects should degrade exactly in those regimes; that is directly checkable on current benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes FocusMamba, an RGB-event fusion detector that jointly sparsifies tokens in both modalities before fusion. An Event-Guided Multimodal Sparsification (EGMS) strategy computes statistical importance scores for each modality and then uses the event spatial ratio r to set a per-sample scale factor and threshold via Eqs. (4)-(6). The resulting sparsification maps guide sparse VSS layers and a Cross-Modality Focus Fusion (CMFF) module, which enhances complementary regions and fuses the retained tokens with a Mamba-based scan. Experiments on DSEC-Det and PKU-DAVIS-SOD report SOTA mAP and large FLOPs reductions relative to SFNet and other fusion baselines, with additional ablations over scoring methods, EGCM factors, fusion strategies, and a ViT-based extension.

Significance. If the accuracy and efficiency claims are taken as stated, the paper makes a useful contribution to efficient multimodal detection: it explicitly couples per-sample sparsification decisions to the event stream and designs a fusion module that only operates on retained tokens. The ablation structure is sensible and each added component gives an incremental gain, so the central accuracy claim is defensible. The extension to MaxViT also suggests generality. However, the headline efficiency claim is based on FLOPs and is contradicted by the reported wall-clock runtimes, and the scene-adaptivity evidence in Fig. 10 is largely circular because the same r is used to construct the control mechanism. These issues affect the main contribution and need to be resolved before the paper's central claims can be accepted.

major comments (3)
  1. [Section IV-B, Table I] The central efficiency claim is not supported by the measured runtimes. The text states FocusMamba-B requires only 29.1% (DSEC-Det) and 22.2% (PKU-DAVIS-SOD) of SFNet's FLOPs and concludes 'an excellent balance between accuracy and efficiency.' However, Table I reports runtime 54.7 ms vs 44.8 ms on DSEC-Det and 53.5 ms vs 42.3 ms on PKU-DAVIS-SOD, i.e., FocusMamba-B is about 22% and 26% slower. The sentence in Section IV-B conceding that inference speed is limited by Mamba's inefficient hardware utilization is not quantified and does not appear in the abstract or conclusion. The efficiency claim should be restricted to FLOPs, or supported by matched latency/energy measurements, before 'superior in both accuracy and efficiency' can be made.
  2. [Section III-B2, Eqs. (4)-(6), Fig. 10, Limitation] The claim that EGMS is scene-adaptive is partly circular. Scale = r^{1/ρ} and Control = (1-r)^{1/ρ} are both monotone in r, and r is also the quantity plotted in Fig. 10 against the token kept ratio; hence the observed trend is a direct consequence of the chosen functional forms, not an independent empirical validation. The paper's own Limitation section concedes that r is unreliable for static scenes, and Eqs. (4)-(5) also have an undefined behavior at r=0 (Scale=0 leads to division by zero). Please provide an r-perturbation ablation or an independent proxy for 'object information' to show that the retained tokens track scene content rather than the injected r, and specify the behavior for r=0 frames.
  3. [Tables III-VII, Section IV-D] All ablation numbers appear to come from a single run without error bars or seed counts. The incremental mAP gains that carry the component claims are 0.8, 1.0, and 1.2 points (Tables III and V), which are within plausible run-to-run variation for detection on these datasets. Please report mean ± std over at least three seeds (or confidence intervals) for the main comparison and ablations; otherwise the contribution of each module is not statistically distinguishable.
minor comments (5)
  1. [Contribution list vs Section IV-B] The percentage of SFNet's FLOPs is reported as 28.5% and 22.2% in the introduction, but Section IV-B says 29.1% and 22.2% for the same comparison. Please make the numbers consistent.
  2. [Section IV-E] The text states that in stage 1 the image modality uses the sparsification map from the event modality. This design choice is not described in the method section and is important for reproducibility; it should be part of the architecture description.
  3. [Figure 10] The plot shows an aggregate trend but reports no correlation coefficient, number of sequences, or per-sequence variability. Please add these statistics, and note explicitly that the trend is expected from the definition of the control mechanism.
  4. [Eq. (8)] The notation 'ME - MI' for binary maps is confusing. It should be defined as a set difference or as a logical operation with explicit precedence, because the equality with ME xor (ME and MI) is not immediately clear.
  5. [Figure 1(a)] The caption '15%15% 1.8%18.7%' is hard to parse. Please clarify which numbers correspond to which scenes and methods.

Circularity Check

1 steps flagged

One self-definitional illustration of scene-adaptivity; central accuracy/efficiency comparisons are independent.

specific steps
  1. self definitional [Section IV-E (Sparsification Visualizations), Fig. 10; Eqs. (4)-(7) in Section III-B2]
    "Additionally, we assess the scene sparsity by calculating the ratio of triggered event pixels. The relationship between the token kept ratio and the event spatial ratio across the DSEC-Det and PKU-DAVIS-SOD datasets is depicted in Figure 10. The overall trend depicted in the figure indicates that the token kept ratio increases with the event spatial ratio, demonstrating that our method can adaptively perform sparsification based on the scene's sparsity."

    In Section III-B2, EGCM is defined so that r directly controls the score scaling and the sparsification threshold: Eq. 4 sets Scale = r^{1/ρ} and Eq. 6 sets α = 1/(N×(1−r)^{1/ρ}). Therefore the number of tokens surviving the threshold in Eq. 7 is a function of r by construction. Plotting the retained-token ratio against r in Fig. 10 and reading the trend as 'demonstrating' scene-adaptivity is a restatement of the control law, not an independent empirical confirmation. The figure is an illustration of the mechanism's intended behavior, not a test of it. The core accuracy and FLOPs results are unaffected, but this particular supporting claim is circular.

full rationale

The core derivation chain is experimental: EGMS and CMFF are evaluated against external benchmarks (DSEC-Det, PKU-DAVIS-SOD) and ablations, so the main accuracy/efficiency claim does not reduce to the method's definitions. The self-citation to SMamba [45] supplies a scoring heuristic whose equations are restated in the paper; it is reused, not load-bearing. The one genuinely circular step is the Fig. 10 'observed correlation' claim: because Eq. 4 and Eq. 6 make both the softmax scale and the threshold explicit functions of r, the dependence of retained-token ratio on r is a by-construction property of EGCM, not an empirical discovery. This is a supporting illustration rather than the central result. The Limitation passage explicitly concedes the r proxy fails in static scenes, which is a stated assumption, not a hidden circularity. Finally, Table I shows FocusMamba-B has lower FLOPs but ~22-26% higher runtime than SFNet; the 'excellent balance' claim is therefore FLOPs-only, a correctness/evaluation concern outside circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central adaptive mechanism rests on the event spatial ratio r, the reliability of two statistical scoring metrics, and two ad-hoc functional forms with hyperparameters. The main free parameters are rho, beta, and unspecified details of r and sigma; none are reported with fitted values. No new physical entities are postulated.

free parameters (4)
  • rho (ρ)
    Hyperparameter controlling the scale and control factors in EGCM (Eq. 4, 6). Its value is not reported and no sensitivity analysis is provided; it modulates how aggressively tokens are pruned.
  • beta (β)
    Enhancement coefficient in the CAE module (Eq. 9) that scales complementary regions before fusion. Value not reported; treated as hand-chosen.
  • Gaussian sigma (σ)
    Variance in the Gaussian kernel for event spatiotemporal continuity scoring (Eq. 3). Not specified; affects event scoring.
  • Event trigger threshold for r
    The definition of 'pixels that trigger events' used to compute the spatial ratio r is unspecified; a threshold on event count per pixel would be a free parameter.
axioms (4)
  • domain assumption The spatial ratio r of pixels that trigger events reflects the objects' information content and the scene's sparsity level.
    Central to EGCM: both scale (Eq. 4) and control (Eq. 6) factors are functions of r. Only valid for moving objects producing events; acknowledged in the Limitation for static scenes.
  • domain assumption L2 activation norm of image tokens is a reliable indicator of object relevance.
    Used in Eq. (1) for image scoring, following prior work [42], [46]. Assumes background tokens have low L2 norm.
  • domain assumption Spatiotemporal continuity of events distinguishes activity regions from noise.
    Event scoring uses the STCA-style metric from SMamba [45], a self-cited prior method; assumes noise events are temporally and spatially isolated.
  • ad hoc to paper The functional forms Scale = r^(1/ρ) and Control = (1-r)^(1/ρ) are chosen without derivation.
    No justification is given for these specific monotonic forms beyond desired qualitative behavior; they introduce the hyperparameter rho.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 16847 in / 13334 out tokens · 115377 ms · 2026-08-05T10:34:46.443400+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Focus Through Motion: RGB-Event Collaborative Token Sparsification for Efficient Object Detection." pith.science (2026). https://pith.science/paper/QDCBQGKW

@misc{pith2026250903872,
  author       = {Pith},
  title        = {Pith review of: Focus Through Motion: RGB-Event Collaborative Token Sparsification for Efficient Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDCBQGKW}},
  note         = {Machine review of arXiv:2509.03872}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing RGB-Event detection methods process the low-information regions of both modalities (background in images and non-event regions in event data) uniformly during feature extraction and fusion, resulting in high computational costs and suboptimal performance. To mitigate the computational redundancy during feature extraction, researchers have respectively proposed token sparsification methods for the image and event modalities. However, these methods employ a fixed number or threshold for token selection, hindering the retention of informative tokens for samples with varying complexity. To achieve a better balance between accuracy and efficiency, we propose FocusMamba, which performs adaptive collaborative sparsification of multimodal features and efficiently integrates complementary information. Specifically, an Event-Guided Multimodal Sparsification (EGMS) strategy is designed to identify and adaptively discard low-information regions within each modality by leveraging scene content changes perceived by the event camera. Based on the sparsification results, a Cross-Modality Focus Fusion (CMFF) module is proposed to effectively capture and integrate complementary features from both modalities. Experiments on the DSEC-Det and PKU-DAVIS-SOD datasets demonstrate that the proposed method achieves superior performance in both accuracy and efficiency compared to existing methods. The code will be available at https://github.com/Zizzzzzzz/FocusMamba.

Figures

Figures reproduced from arXiv: 2509.03872 by Nan Yang, Xiangmo Zhao, Yang Liu, Yang Wang, Yuchao Dai, Zhanwen Liu.

Figure 1
Figure 1. Figure 1: (a) Comparison of token retention (yellow regions) between the fixed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of FocusMamba. The image and event tensors are first tokenized and then processed through four stages for multi-scale feature extraction. Before each stage, EGMS generates independent sparsification maps for both modalities to guide the sparsification operation. The CMFF module is applied in the final three stages to fuse the features of two modalities. [59]–[66] have subsequently applied … view at source ↗
Figure 3
Figure 3. Figure 3: We select 10 tokens from sparse and dense object-information scenes [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-Modality Focus Fusion module. Guided by the sparsification maps MI and ME, the CMFF module accurately captures and fully leverages the complementary features, eliminating background interference and reducing redundant computation. captures and fully utilizes the complementary features of both modalities, while eliminating background interference. To enhance the degraded regions in each modality (e.g.… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison with SFNet, ConcatMamba, and CrossMamba on the DSEC-Det dataset. We utilize blue arrows to mark the failed cases. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with SFNet, ConcatMamba, and CrossMamba on the PKU-DAVIS-SOD dataset. We utilize blue arrows to mark the failed [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of features before and after CMFF module. (a) illustrates the kept regions of the image and event modalities. (c) represents the difference [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualizations of sparsification results at different stages on the DSEC [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The trend of the token kept ratio varies with scene sparsity across [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SkyEV: RGB-Event UAV detection and tracking dataset and baseline

    cs.CV 2026-07 conditional novelty 6.0

    The paper introduces SkyEV, a 2.17-hour RGB-event drone detection dataset with ego-motion and varied optics, plus a SAST+YOLOX fusion baseline.

Reference graph

Works this paper leans on

78 extracted references · 56 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Dynamic obstacle avoid- ance for quadrotors with event cameras,

    D. Falanga, K. Kleber, and D. Scaramuzza, “Dynamic obstacle avoid- ance for quadrotors with event cameras,”Science Robotics, vol. 5, no. 40, p. eaaz9712, 2020

    2020

  2. [2]

    Low-latency automotive vision with event cameras,

    D. Gehrig and D. Scaramuzza, “Low-latency automotive vision with event cameras,” Nature, vol. 629, no. 8014, pp. 1034–1040, 2024

    work page 2024

  3. [3]

    Mstf: Multiscale transformer for incomplete trajectory prediction,

    Z. Liu, C. Li, N. Yang, Y . Wang, J. Ma, G. Cheng, and X. Zhao, “Mstf: Multiscale transformer for incomplete trajectory prediction,” in 2024 IEEE Intelligent Vehicles Symposium (IV) . IEEE, 2024, pp. 573–580

    work page 2024

  4. [4]

    Multi-modal fusion based on depth adaptive mechanism for 3d object detection,

    Z. Liu, J. Cheng, J. Fan, S. Lin, Y . Wang, and X. Zhao, “Multi-modal fusion based on depth adaptive mechanism for 3d object detection,” IEEE Transactions on Multimedia , vol. 27, pp. 707–717, 2023

    work page 2023

  5. [5]

    Intention-convolution and hybrid-attention network for vehicle trajectory prediction,

    C. Li, Z. Liu, S. Lin, Y . Wang, and X. Zhao, “Intention-convolution and hybrid-attention network for vehicle trajectory prediction,” Expert Systems with Applications , vol. 236, p. 121412, 2024

    work page 2024

  6. [6]

    Regional attention network with data-driven modal representation for multimodal trajectory prediction,

    C. Li, Z. Liu, N. Yang, W. Li, and X. Zhao, “Regional attention network with data-driven modal representation for multimodal trajectory prediction,” Expert Systems with Applications, vol. 232, p. 120808, 2023

    work page 2023

  7. [7]

    Lightweight adaptive feature de-drifting for compressed image classification,

    L. Peng, Y . Cao, Y . Sun, and Y . Wang, “Lightweight adaptive feature de-drifting for compressed image classification,” IEEE Transactions on Multimedia, vol. 26, pp. 6424–6436, 2024

    2024

  8. [8]

    Cascade saccade machine learning network with hierarchical classes for traffic sign detection,

    Z. Liu, M. Qi, C. Shen, Y . Fang, and X. Zhao, “Cascade saccade machine learning network with hierarchical classes for traffic sign detection,” Sustainable Cities and Society , vol. 67, p. 102700, 2021

    work page 2021

  9. [9]

    End-to-end autonomous driving decision model joined by attention mechanism and spatiotemporal features,

    X. Zhao, M. Qi, Z. Liu, S. Fan, C. Li, and M. Dong, “End-to-end autonomous driving decision model joined by attention mechanism and spatiotemporal features,” IET intelligent transport systems, vol. 15, no. 9, pp. 1119–1130, 2021

    work page 2021

  10. [10]

    Multi- scale temporal fusion transformer for incomplete vehicle trajectory prediction,

    Z. Liu, C. Li, Y . Wang, N. Yang, X. Fan, J. Ma, and X. Zhao, “Multi- scale temporal fusion transformer for incomplete vehicle trajectory prediction,” IEEE Transactions on Intelligent Vehicles , 2024

    2024

  11. [11]

    Harnessing multimodal large language models for traffic knowledge graph generation and decision-making,

    S. Kuang, Y . Liu, X. Wang, X. Wu, and Y . Wei, “Harnessing multimodal large language models for traffic knowledge graph generation and decision-making,” p. 100146, 2024

    work page 2024

  12. [12]

    Envisioning the future of transportation: Inspiration of chatgpt and large models,

    X. Qu, H. Lin, and Y . Liu, “Envisioning the future of transportation: Inspiration of chatgpt and large models,” p. 100103, 2023

    work page 2023

  13. [13]

    Two-stream lstm network with hybrid attention for vehicle trajectory prediction,

    C. Li, Z. Liu, J. Zhang, Y . Wang, F. Ding, and X. Zhao, “Two-stream lstm network with hybrid attention for vehicle trajectory prediction,” in 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2022, pp. 1927–1934

    work page 2022

  14. [14]

    Boosting image de-raining via central-surrounding synergistic convolution,

    L. Peng, Y . Wang, X. Di, X. Fu, Y . Cao, Z.-J. Zha et al. , “Boosting image de-raining via central-surrounding synergistic convolution,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 6, 2025, pp. 6470–6478

    work page 2025

  15. [15]

    Pixel to gaussian: Ultra-fast continuous super-resolution with 2d gaussian modeling,

    L. Peng, A. Wu, W. Li, P. Xia, X. Dai, X. Zhang, X. Di, H. Sun, R. Pei, Y . Wanget al., “Pixel to gaussian: Ultra-fast continuous super-resolution with 2d gaussian modeling,” arXiv preprint arXiv:2503.06617 , 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Enhancing traffic object detection in variable illumination with rgb-event fusion,

    Z. Liu, N. Yang, Y . Wang, Y . Li, X. Zhao, and F.-Y . Wang, “Enhancing traffic object detection in variable illumination with rgb-event fusion,” IEEE Transactions on Intelligent Transportation Systems , 2024

    work page 2024

  17. [17]

    Boosting visual recognition in real-world degradations via unsupervised feature enhancement module with deep channel prior,

    Z. Liu, Y . Li, Y . Wang, B. Gao, Y . An, and X. Zhao, “Boosting visual recognition in real-world degradations via unsupervised feature enhancement module with deep channel prior,” IEEE Transactions on Intelligent Vehicles, 2024

    work page 2024

  18. [18]

    A deep cnn method for underwater image enhancement,

    Y . Wang, J. Zhang, Y . Cao, and Z. Wang, “A deep cnn method for underwater image enhancement,” in 2017 IEEE international conference on image processing (ICIP) . IEEE, 2017, pp. 1382–1386

    work page 2017

  19. [19]

    Decoupling-and- aggregating for image exposure correction,

    Y . Wang, L. Peng, L. Li, Y . Cao, and Z.-J. Zha, “Decoupling-and- aggregating for image exposure correction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 18 115–18 124

    work page 2023

  20. [20]

    Efficient real-world image super-resolution via adaptive directional gradient convolution,

    L. Peng, Y . Cao, R. Pei, W. Li, J. Guo, X. Fu, Y . Wang, and Z.-J. Zha, “Efficient real-world image super-resolution via adaptive directional gradient convolution,” arXiv preprint arXiv:2405.07023 , 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Brightness perceiving for recursive low-light image enhancement,

    H. Wang, L. Peng, Y . Sun, Z. Wan, Y . Wang, and Y . Cao, “Brightness perceiving for recursive low-light image enhancement,” IEEE Transac- tions on Artificial Intelligence , vol. 5, no. 6, pp. 3034–3045, 2023

    work page 2023

  22. [22]

    RobustGS: Unified Boosting of Feedforward 3D Gaussian Splatting under Low-Quality Conditions

    A. Wu, L. Peng, X. Di, X. Dai, C. Wu, Y . Wang, X. Fu, Y . Cao, and Z.-J. Zha, “Robustgs: Unified boosting of feedforward 3d gaussian splatting under low-quality conditions,” arXiv preprint arXiv:2508.03077 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Deep degrada- tion prior for low-quality image classification,

    Y . Wang, Y . Cao, Z.-J. Zha, J. Zhang, and Z. Xiong, “Deep degrada- tion prior for low-quality image classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 11 049–11 058

    work page 2020

  24. [24]

    Event- based vision: A survey,

    G. Gallego, T. Delbr ¨uck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidis et al., “Event- based vision: A survey,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 1, pp. 154–180, 2020

    work page 2020

  25. [25]

    E-mlb: Multilevel benchmark for event-based camera denoising,

    S. Ding, J. Chen, Y . Wang, Y . Kang, W. Song, J. Cheng, and Y . Cao, “E-mlb: Multilevel benchmark for event-based camera denoising,” IEEE Transactions on Multimedia , vol. 26, pp. 65–76, 2023

    work page 2023

  26. [26]

    Unveiling hidden details: A raw data-enhanced paradigm for real-world super-resolution,

    L. Peng, W. Li, J. Guo, X. Di, H. Sun, Y . Li, R. Pei, Y . Wang, Y . Cao, and Z.-J. Zha, “Unveiling hidden details: A raw data-enhanced paradigm for real-world super-resolution,” arXiv preprint arXiv:2411.10798, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Multi-grained spatio-temporal features perceived network for event-based lip-reading,

    G. Tan, Y . Wang, H. Han, Y . Cao, F. Wu, and Z.-J. Zha, “Multi-grained spatio-temporal features perceived network for event-based lip-reading,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20 094–20 103

    work page 2022

  28. [28]

    Event- based optical flow via transforming into motion-dependent view,

    Z. Wan, G. Tan, Y . Wang, W. Zhai, Y . Cao, and Z.-J. Zha, “Event- based optical flow via transforming into motion-dependent view,” IEEE Transactions on Image Processing , vol. 33, pp. 5327–5339, 2024

    work page 2024

  29. [29]

    Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation,

    J. Chen, Y . Wang, Y . Cao, F. Wu, and Z.-J. Zha, “Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 1, 2022, pp. 303–311

    work page 2022

  30. [30]

    Progressive retinex: Mutually reinforced illumination-noise perception network for low-light image enhancement,

    Y . Wang, Y . Cao, Z.-J. Zha, J. Zhang, Z. Xiong, W. Zhang, and F. Wu, “Progressive retinex: Mutually reinforced illumination-noise perception network for low-light image enhancement,” in Proceedings of the 27th ACM international conference on multimedia , 2019, pp. 2015–2023

    work page 2019

  31. [31]

    Event voxel set transformer for spatiotemporal representation learning on event streams,

    B. Xie, Y . Deng, Z. Shao, Q. Xu, and Y . Li, “Event voxel set transformer for spatiotemporal representation learning on event streams,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    work page 2024

  32. [32]

    Ecsnet: Spatio- temporal feature learning for event camera,

    Z. Chen, J. Wu, J. Hou, L. Li, W. Dong, and G. Shi, “Ecsnet: Spatio- temporal feature learning for event camera,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 2, pp. 701–712, 2022

    work page 2022

  33. [33]

    V oxel-based multi-scale transformer network for event stream processing,

    D. Liu, T. Wang, and C. Sun, “V oxel-based multi-scale transformer network for event stream processing,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2112–2124, 2023

    work page 2023

  34. [34]

    Sodformer: Streaming object detection with transformer using events and frames,

    D. Li, Y . Tian, and J. Li, “Sodformer: Streaming object detection with transformer using events and frames,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 11, pp. 14 020–14 037, 2023

    work page 2023

  35. [35]

    Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks,

    X. Long, X. Zhu, F. Guo, C. Chen, X. Zhu, F. Gu, S. Yuan, and C. Zhang, “Spike-brgnet: Efficient and accurate event-based semantic segmentation with boundary region-guided spiking neural networks,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  36. [36]

    Beyond conventional vision: Rgb-event fusion for robust object detection in dynamic traffic scenarios,

    Z. Liu, Y . Sun, Y . Wang, N. Yang, S. E. Li, and X. Zhao, “Beyond conventional vision: Rgb-event fusion for robust object detection in dynamic traffic scenarios,” Communications in Transportation Research, vol. 5, p. 100202, 2025

    work page 2025

  37. [37]

    Event-based monoc- ular depth estimation with recurrent transformers,

    X. Liu, J. Li, J. Shi, X. Fan, Y . Tian, and D. Zhao, “Event-based monoc- ular depth estimation with recurrent transformers,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 8, pp. 7417– 7429, 2024

    work page 2024

  38. [38]

    Fusing event-based and rgb camera for robust object detection in adverse conditions,

    A. Tomy, A. Paigwar, K. S. Mann, A. Renzaglia, and C. Laugier, “Fusing event-based and rgb camera for robust object detection in adverse conditions,” in 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 933–939

    work page 2022

  39. [39]

    Joint intensity and event framework for vehicle detection in degraded conditions,

    N. Yang, Z. Liu, S. Ma, Y . Sun, Y . He, and Y . Wang, “Joint intensity and event framework for vehicle detection in degraded conditions,” in 2023 7th International Conference on Transportation Information and Safety (ICTIS). IEEE, 2023, pp. 1568–1574

    work page 2023

  40. [40]

    Chasing day and night: Towards robust and efficient all-day object detection guided by an event camera,

    J. Cao, X. Zheng, Y . Lyu, J. Wang, R. Xu, and L. Wang, “Chasing day and night: Towards robust and efficient all-day object detection guided by an event camera,” in2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9026–9032. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    work page 2024

  41. [41]

    Sparse detr: Efficient end-to-end object detection with learnable sparsity,

    B. Roh, J. Shin, W. Shin, and S. Kim, “Sparse detr: Efficient end-to-end object detection with learnable sparsity,” arXiv preprint arXiv:2111.14330, 2021

    Pith/arXiv arXiv 2021

  42. [42]

    Adaptive Sparse ViT: Towards Learnable Adaptive Token Pruning by Fully Exploiting Self-Attention

    X. Liu, T. Wu, and G. Guo, “Adaptive sparse vit: Towards learnable adaptive token pruning by fully exploiting self-attention,” arXiv preprint arXiv:2209.13802, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  43. [43]

    Revisiting token pruning for object detection and instance segmentation,

    Y . Liu, M. Gehrig, N. Messikommer, M. Cannici, and D. Scara- muzza, “Revisiting token pruning for object detection and instance segmentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 2658–2668

    work page 2024

  44. [44]

    Scene adaptive sparse transformer for event-based object detection,

    Y . Peng, H. Li, Y . Zhang, X. Sun, and F. Wu, “Scene adaptive sparse transformer for event-based object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 16 794–16 804

    work page 2024

  45. [45]

    SMamba: Sparse Mamba for Event-based Object Detection

    N. Yang, Y . Wang, Z. Liu, M. Li, Y . An, and X. Zhao, “Smamba: Sparse mamba for event-based object detection,” arXiv preprint arXiv:2501.11971, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  46. [46]

    Sparsevit: Revis- iting activation sparsity for efficient high-resolution vision transformer,

    X. Chen, Z. Liu, H. Tang, L. Yi, H. Zhao, and S. Han, “Sparsevit: Revis- iting activation sparsity for efficient high-resolution vision transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2061–2070

    work page 2023

  47. [47]

    Rgb-event fusion for moving object detection in autonomous driving,

    Z. Zhou, Z. Wu, R. Boutteau, F. Yang, C. Demonceaux, and D. Ginhac, “Rgb-event fusion for moving object detection in autonomous driving,” arXiv preprint arXiv:2209.08323 , 2022

    Pith/arXiv arXiv 2022

  48. [48]

    Embracing events and frames with hierarchical feature refinement network for object detection,

    H. Cao, Z. Zhang, Y . Xia, X. Li, J. Xia, G. Chen, and A. Knoll, “Embracing events and frames with hierarchical feature refinement network for object detection,” in European Conference on Computer Vision. Springer, 2025, pp. 161–177

    work page 2025

  49. [49]

    Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction,

    L. Xing, Q. Huang, X. Dong, J. Lu, P. Zhang, Y . Zang, Y . Cao, C. He, J. Wang, F. Wu et al. , “Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction,” arXiv preprint arXiv:2410.17247, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models,

    L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang, “An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models,” in European Conference on Computer Vision . Springer, 2024, pp. 19–35

    work page 2024

  51. [51]

    Fit and prune: Fast and training- free visual token pruning for multi-modal large language models,

    W. Ye, Q. Wu, W. Lin, and Y . Zhou, “Fit and prune: Fast and training- free visual token pruning for multi-modal large language models,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 21, 2025, pp. 22 128–22 136

    work page 2025

  52. [52]

    PACT: Pruning and Clustering-Based Token Reduction for Faster Visual Language Models

    M. Dhouib, D. Buscaldi, S. Vanier, and A. Shabou, “Pact: Pruning and clustering-based token reduction for faster visual language models,” arXiv preprint arXiv:2504.08966 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  53. [53]

    Boosting multimodal large language models with visual tokens withdrawal for rapid inference,

    Z. Lin, M. Lin, L. Lin, and R. Ji, “Boosting multimodal large language models with visual tokens withdrawal for rapid inference,” in Proceed- ings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 5, 2025, pp. 5334–5342

    work page 2025

  54. [54]

    Dynam- icvit: Efficient vision transformers with dynamic token sparsification,

    Y . Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “Dynam- icvit: Efficient vision transformers with dynamic token sparsification,” Advances in neural information processing systems, vol. 34, pp. 13 937– 13 949, 2021

    work page 2021

  55. [55]

    Spvit: Enabling faster vision transformers via latency-aware soft token pruning,

    Z. Kong, P. Dong, X. Ma, X. Meng, W. Niu, M. Sun, X. Shen, G. Yuan, B. Ren, H. Tang et al., “Spvit: Enabling faster vision transformers via latency-aware soft token pruning,” in European conference on computer vision. Springer, 2022, pp. 620–640

    work page 2022

  56. [56]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    Pith/arXiv arXiv 2023

  57. [57]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” arXiv preprint arXiv:2401.09417 , 2024

    Pith/arXiv arXiv 2024

  58. [58]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, and Y . Liu, “Vmamba: Visual state space model,” arXiv preprint arXiv:2401.10166, 2024

    Pith/arXiv arXiv 2024

  59. [59]

    A survey on vision mamba: Models, applications and challenges,

    R. Xu, S. Yang, Y . Wang, B. Du, and H. Chen, “A survey on vision mamba: Models, applications and challenges,” arXiv preprint arXiv:2404.18861, 2024

    Pith/arXiv arXiv 2024

  60. [60]

    Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges,

    B. N. Patro and V . S. Agneeswaran, “Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges,” arXiv preprint arXiv:2404.16112 , 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Sigma: Siamese mamba network for multi-modal semantic segmentation,

    Z. Wan, P. Zhang, Y . Wang, S. Yong, S. Stepputtis, K. Sycara, and Y . Xie, “Sigma: Siamese mamba network for multi-modal semantic segmentation,” arXiv preprint arXiv:2404.04256 , 2024

    Pith/arXiv arXiv 2024

  62. [62]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” in ECCV, 2024

    work page 2024

  63. [63]

    Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,

    W. Zou, H. Gao, W. Yang, and T. Liu, “Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,” in ACM Multimedia 2024 , 2024

    work page 2024

  64. [64]

    Mamba- fetrack: Frame-event tracking via state space model,

    J. Huang, S. Wang, S. Wang, Z. Wu, X. Wang, and B. Jiang, “Mamba- fetrack: Frame-event tracking via state space model,” in Chinese Confer- ence on Pattern Recognition and Computer Vision (PRCV) . Springer, 2024, pp. 3–18

    work page 2024

  65. [65]

    Directing mamba to complex textures: An efficient texture-aware state space model for image restoration,

    L. Peng, X. Di, Z. Feng, W. Li, R. Pei, Y . Wang, X. Fu, Y . Cao, and Z.-J. Zha, “Directing mamba to complex textures: An efficient texture-aware state space model for image restoration,” arXiv preprint arXiv:2501.16583, 2025

    Pith/arXiv arXiv 2025

  66. [66]

    Qmambabsr: Burst image super-resolution with query state space model,

    X. Di, L. Peng, P. Xia, W. Li, R. Pei, Y . Cao, Y . Wang, and Z.-J. Zha, “Qmambabsr: Burst image super-resolution with query state space model,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 23 080–23 090

    work page 2025

  67. [67]

    Unsupervised event- based learning of optical flow, depth, and egomotion,

    A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “Unsupervised event- based learning of optical flow, depth, and egomotion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 989–997

    work page 2019

  68. [68]

    Transformer interpretability beyond attention visualization,

    H. Chefer, S. Gur, and L. Wolf, “Transformer interpretability beyond attention visualization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 782–791

    work page 2021

  69. [69]

    Visual transformers: Token- based image representation and processing for computer vision,

    B. Wu, C. Xu, X. Dai, A. Wan, P. Zhang, Z. Yan, M. Tomizuka, J. Gonzalez, K. Keutzer, and P. Vajda, “Visual transformers: Token- based image representation and processing for computer vision,” arXiv preprint arXiv:2006.03677, 2020

    Pith/arXiv arXiv 2006

  70. [70]

    Soft filter pruning for accelerating deep convolutional neural networks,

    Y . He, G. Kang, X. Dong, Y . Fu, and Y . Yang, “Soft filter pruning for accelerating deep convolutional neural networks,” in International Joint Conference on Artificial Intelligence (IJCAI) , 2018, pp. 2234–2240

    work page 2018

  71. [71]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 . Springer, 2014, pp. 740–755

    2014

  72. [72]

    State space models for event cameras,

    N. Zubic, M. Gehrig, and D. Scaramuzza, “State space models for event cameras,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 5819–5828

    work page 2024

  73. [73]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430 , 2021

    Pith/arXiv arXiv 2021

  74. [74]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,” arXiv preprint arXiv:2410.17725 , 2024

    Pith/arXiv arXiv 2024

  75. [75]

    Mamba yolo: Ssms-based yolo for object detection,

    Z. Wang, C. Li, H. Xu, and X. Zhu, “Mamba yolo: Ssms-based yolo for object detection,” arXiv preprint arXiv:2406.05835 , 2024

    Pith/arXiv arXiv 2024

  76. [76]

    Asynchronous collaborative graph representation for frames and events,

    D. Li, J. Li, X. Liu, X. Fan, and Y . Tian, “Asynchronous collaborative graph representation for frames and events,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 1655– 1666

    work page 2025

  77. [77]

    Divprune: Diversity- based visual token pruning for large multimodal models,

    S. R. Alvar, G. Singh, M. Akbari, and Y . Zhang, “Divprune: Diversity- based visual token pruning for large multimodal models,” arXiv preprint arXiv:2503.02175, 2025

    Pith/arXiv arXiv 2025

  78. [78]

    Maxvit: Multi-axis vision transformer,

    Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y . Li, “Maxvit: Multi-axis vision transformer,” in European conference on computer vision. Springer, 2022, pp. 459–479

    work page 2022