Pith. sign in

REVIEW 3 major objections 5 minor 76 references

E-TraMamba is the first Mamba-based tracker that turns sparse event streams into long-lived 3D feature trajectories with linear-cost state-space modeling.

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 · grok-4.5

2026-07-12 06:29 UTC pith:XL6HXVFE

load-bearing objection First Mamba-based event 3D feature tracker with clear SOTA gains on real stereo data and a usable new synthetic set; the 2 imes lifetime claim holds on E-3DTrack. the 3 major comments →

arxiv 2607.02866 v1 pith:XL6HXVFE submitted 2026-07-03 cs.CV cs.RO

E-TraMamba: A New Paradigm for Efficient Long-Term 3D Feature Tracking with Event Cameras

Juwei Shen , Yujie Wu , Changwen Chen This is my paper
classification cs.CV cs.RO
keywords event cameras3D feature trackingMambastate-space modelslong-range temporal modelingsemi-synthetic event datasetreal-time SLAM
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.

Event cameras deliver microsecond-resolution brightness changes that should support high-speed 3D perception, yet existing CNN and Transformer trackers cannot keep long-range temporal memory without exploding compute. This paper shows that a selective state-space model (Mamba) can fuse local event patches, correlation maps and positional cues into track tokens, then refine them with linear complexity so that 3D points stay accurate far longer. A lightweight affine predictor further stabilizes tracks under blur and occlusion. The authors also release EvD-PointOdyssey, a large semi-synthetic set of synchronized events, depth and 3D trajectories, to train and test such models. On real stereo benchmarks the method more than doubles feature lifetime under a 0.1 m error budget while running at roughly 475 frames per second, making continuous low-latency 3D tracking practical for robotics and SLAM.

Core claim

A Mamba state-space backbone that scans track tokens in four space-time directions, combined with multi-scale cue fusion and an affine motion head, produces 3D feature tracks whose lifetimes exceed those of prior event trackers by more than a factor of two under strict accuracy thresholds, all at linear rather than quadratic cost.

What carries the argument

E-TraMamba: a linear state-space (Mamba) refinement module that processes multi-scale track tokens (local patches + correlation + position + optional depth) with four-directional scanning and residual updates, plus a lightweight affine-transform predictor.

Load-bearing premise

The aggressively filtered semi-synthetic trajectories and fixed-threshold event simulation used for training and part of the evaluation are realistic enough that the measured gains will hold on real asynchronous event cameras and real-world 3D motion.

What would settle it

Train and run the identical architecture on a large real-world stereo event dataset that has ground-truth 3D trajectories and measure whether feature age under the 0.1 m threshold still exceeds the previous best method by a factor of two; if the gap collapses, the claim fails.

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

If this is right

  • Real-time visual odometry and SLAM systems can maintain denser, longer-lived 3D landmarks from pure event input without quadratic attention cost.
  • High-speed interactive robots can track geometric features continuously even under motion blur and extreme lighting.
  • Future event trackers can replace Transformer blocks with linear state-space modules and still improve both accuracy and speed.
  • The EvD-PointOdyssey construction pipeline supplies a reusable template for generating large-scale 3D event training data from existing RGB-D point-tracking assets.

Where Pith is reading between the lines

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

  • The same four-directional Mamba scan could be applied to other sparse asynchronous sensors (e.g., spike cameras or lidar event streams) with only the token-construction front-end changed.
  • Because the model already predicts confidence and affine warps, an online re-detection module could be added with almost no extra latency, turning the tracker into a full long-term SLAM front-end.
  • If the linear-complexity advantage scales to hundreds of simultaneous tracks, multi-camera event arrays become feasible for wide-baseline 3D reconstruction at video rates.

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 introduces E-TraMamba, claimed as the first Mamba-based (linear state-space) framework for event-based 3D feature tracking. It constructs track tokens by fusing local CNN event patches, correlation features, affine-aligned reference templates, and positional encodings (Eqs. 5–8), refines them via multi-group four-directional Mamba scanning for long-range spatiotemporal modeling at linear cost, and predicts 2D positions, disparity/depth and visibility. A lightweight affine-transform predictor is integrated for stability under blur/occlusion. The authors also release EvD-PointOdyssey, a filtered semi-synthetic monocular+depth dataset derived from PointOdyssey with event simulation. Experiments on real stereo E-3DTrack and the new dataset report SOTA gains (e.g., FA@0.1 m of 0.6146 vs. 0.2601, RMSE 0.0914 m vs. 0.1181 m) at ~475 FPS, with ablations on backbone, window size, groups and encodings.

Significance. If the reported gains hold under broader real-world conditions, the work is a solid engineering contribution to event-based vision: it demonstrates that selective state-space models can replace quadratic Transformers for high-frequency sparse 3D feature tracking while preserving or improving accuracy and enabling real-time rates useful for VO/SLAM/robotics. The explicit multi-scale token design, four-directional scanning, and the new EvD-PointOdyssey benchmark (with released sequence lists promised) are concrete assets that the community can build on. The linear-complexity claim is supported by the FLOPs comparison in Table 2 and the architecture details in the supplement; these are genuine strengths relative to prior RNN/Transformer event trackers.

major comments (3)
  1. [Table 1 vs. Table 2 / §4.4] Table 1 (E-3DTrack block) reports TFR(0.1 m)=0.7274 / RMSE=0.0914 m for the final model, yet the corresponding ablation configuration in Table 2 (Mamba, WinSize=12, Grps=2) reports only TFR0.1=0.5950 / RMSE=0.0921. The manuscript never states whether Table 2 uses a different split, omits fusion/affine heads, or evaluates a stripped model; without this clarification the ablation cannot be used to justify the design choices that produce the main SOTA numbers.
  2. [§3.4 / Algorithm 1] Sec. 3.4 and Algorithm 1 impose aggressive filters (max inter-frame displacement ≤12 px, full visibility for the entire window, finite depth, non-overlapping windows, ≤30 points, fixed contrast θ=0.02). These choices produce clean trajectories but systematically exclude the very motion blur, occlusion and large-displacement regimes that the abstract and introduction claim the method handles. Consequently the “high-fidelity / realistic dynamics” claim for EvD-PointOdyssey, and any cross-domain narrative that relies on it, is overstated; at minimum a quantitative characterization of the discarded trajectories and a real-event monocular transfer experiment are required.
  3. [§3.2–3.3 / Eqs. 9–10 / Supplement §7.7] The abstract, introduction and Fig. 2 repeatedly credit a “lightweight affine-transform predictor” for stability under blur and occlusion, yet the loss (Eqs. 9–10) supervises only xy, disparity/depth and 3D coordinates; the supplement’s prediction-head description likewise omits any affine parameters or loss term. Either the affine head is unsupervised (raising the question of how it is trained) or its contribution is not isolated; an ablation removing it is needed to support the claim.
minor comments (5)
  1. [Fig. 1] Fig. 1 caption claims O(N T) versus O(N² T²) but does not define N and T consistently with the later token tensor shape B×T×N×d; a short clarification would help.
  2. [header] The arXiv identifier and year (2607.02866 / 2026) appear anomalous; confirm the correct preprint metadata.
  3. [§2] Several recent event-tracking and Mamba-tracking works (e.g., additional 2024–2025 frame-event fusion trackers) are missing from the related-work discussion; a short update would strengthen positioning.
  4. [§3.2] Notation for the sliding-window event set S_t and the patch Ω_i(t) is introduced twice with slightly different symbols; unify.
  5. [Table 3 / §4.4] Table 3 “TempEnc” ablation is interesting but the text does not explain why removing temporal encoding improves TFR; a one-sentence hypothesis would aid readability.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture + held-out metrics on external real data (E-3DTrack) and independently filterable synthetic data; no prediction reduces to a fitted free parameter or self-definition by construction.

full rationale

E-TraMamba is a standard engineering paper: it defines a Mamba-based token pipeline (Eqs. 5–8), four-directional SSM refinement, affine heads and multi-scale L1 loss (Eqs. 9–10), trains end-to-end, and reports FA/TFR/RMSE against ground-truth trajectories on the external real stereo benchmark E-3DTrack (Table 1 upper block) plus a newly constructed but fully specified monocular+depth set EvD-PointOdyssey. The 2 imes FA@0.1 m claim (0.6146 vs 0.2601) and RMSE reduction are direct numerical comparisons to published baselines under the same evaluation protocol; nothing is fitted on the test set and then re-labeled a “prediction.” Dataset filtering (Sec. 3.4 / Alg. 1) and fixed-threshold event simulation are design choices that affect realism, not circular reductions of the reported metrics. Ablations (Tables 2–3) vary architecture knobs and recompute the same external metrics; no uniqueness theorem, self-citation chain, or ansatz smuggled via prior author work is load-bearing for the central SOTA numbers. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The central empirical claims rest on standard Mamba discretization, conventional event-camera and stereo geometry assumptions, and a large set of hand-chosen architectural and training hyperparameters; the only invented entities are the specific tracker design and the filtered dataset itself.

free parameters (6)
  • window size w = 12
    Chosen as 12 after ablation; controls temporal context and directly affects reported FA/TFR.
  • Mamba groups / directions = 2 groups × 4 dirs
    Two groups of four directional scans selected by ablation; changes FLOPs and accuracy.
  • hidden dim / state dim / expand = 256/32/2
    256 / 32 / 2 set by hand; define model capacity.
  • loss weights λ_xy, λ_disp, λ_3D and decay γ = 1.0 / 0.8
    All set to 1.0 and 0.8 respectively; balance multi-task regression.
  • event contrast threshold θ and depth clip [0.9,50] m = θ=0.02, [0.9,50]
    Fixed once for the entire synthetic dataset; control event density and valid depth range.
  • patch radius r=15 (31×31) and max displacement filter 12 px = 31 px / ≤12 px
    Hand-chosen for token extraction and trajectory cleaning.
axioms (4)
  • domain assumption Selective state-space (Mamba) discretization via zero-order hold yields a linear-time recurrent update that can model long-range dependencies in event tokens.
    Invoked throughout Sec. 3.1–3.2 and used as the sole long-range backbone; taken from Gu & Dao 2024 without re-derivation.
  • domain assumption Event streams can be binned into fixed multi-channel tensors (10 or 11 channels) without destroying the asynchronous advantage for tracking.
    Standard preprocessing assumption used for all CNN feature extractors (Sec. 3.2 and Supp. 7.1).
  • ad hoc to paper Cubic-spline initialization of the second half-window plus affine template alignment is a sufficient motion prior under the observed dynamics.
    Introduced in Sec. 3.2; no independent validation that the prior does not bias long-term metrics.
  • domain assumption Stereo disparity or monocular depth maps are available and accurate enough to lift 2D tracks to 3D for supervision and evaluation.
    Required for the 3D loss (Eq. 9) and for both dataset configurations.
invented entities (2)
  • E-TraMamba architecture (token fusion + 4-dir Mamba + affine head) no independent evidence
    purpose: End-to-end model that maps event patches to long-term 3D trajectories with linear complexity.
    The specific combination and scanning order are introduced here; no external independent evidence beyond the paper’s own ablations.
  • EvD-PointOdyssey dataset no independent evidence
    purpose: Provide large-scale synchronized event–depth–3D-trajectory triplets for training and 2D/3D evaluation.
    Constructed by the authors via monocular rendering + fixed-threshold simulation + filtering; independent evidence would require public release and third-party use.

pith-pipeline@v1.1.0-grok45 · 22616 in / 3365 out tokens · 32675 ms · 2026-07-12T06:29:51.798354+00:00 · methodology

0 comments
read the original abstract

Event-based 3D tracking enables low-latency and high-speed perception, while existing CNN- and Transformer-based trackers struggle to capture long-range spatiotemporal dependencies in sparse, noisy event streams, especially under real-time and efficiency constraints. To address these challenges, we present E-TraMamba, the first Mamba-based framework for 3D feature tracking on event data. This new framework adopts a linear state-space model for efficient long-range modeling and integrates a lightweight affine-transform predictor to maintain stable tracking under motion blur and occlusion. We also design an effective scheme to fuse multi-scale cues -- local spatiotemporal patches, correlation maps, and positional embeddings -- into a unified representation that enables stable and smooth 3D tracking. We construct a large-scale synthetic dataset, named EvD-PointOdyssey, which is generated with monocular rendering and provides synchronized event streams, depth maps, and accurate 3D trajectories for training and evaluating event-based 3D tracking models. Extensive experiments on event-based benchmarks demonstrate that E-TraMamba achieves state-of-the-art performance, delivering over $2\times$ longer feature lifetimes under strict accuracy thresholds (e.g., 0.1 m), with higher tracked-feature ratios and lower RMSE than all baselines. These results make E-TraMamba a strong candidate for low-latency visual odometry, real-time SLAM, and interactive robotics.

Figures

Figures reproduced from arXiv: 2607.02866 by Changwen Chen, Juwei Shen, Yujie Wu.

Figure 1
Figure 1. Figure 1: Overview of the proposed E-TraMamba framework. Our approach unifies real stereo event data (E-3DTrack) and a new semi-synthetic monocular+depth dataset (EvD-PointOdyssey) de￾rived from PointOdyssey. Compared with Transformers requir￾ing quadratic spatial–temporal attention O(N 2T 2 ), the Mamba￾based formulation models feature evolution with linear complexity O(NT), enabling efficient long-range and high-f… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed E-TraMamba framework. a) Architecture Overview: the overall pipeline takes event streams (monocular or stereo) and optionally depth or grayscale templates as input. Extracted point tokens are refined through the Spatial-temporal Mamba module to model long-range dependencies. The affine and displacement regression heads predict per-point motion and affine trans￾formations across tim… view at source ↗
Figure 3
Figure 3. Figure 3: Examples from our generated semi-synthetic dataset [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of 3D feature trajectories on the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of event-based 3D feature trajectories on challenging sequences. Red: ground-truth tra￾jectories; Green: E–TraMamba (ours); Cyan: ETAP [18]; Ma￾genta: DeepEvT [35]; Yellow ⋆: initial feature locations. a) Com￾bined visualization of all methods. b)–d) show individual results for each method, respectively. Our approach produces smoother, longer, and more stable trajectories even under … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

76 extracted references · 5 linked inside Pith

  1. [1]

    Ace: An efficient asynchronous corner tracker for event cameras

    Ignacio Alzugaray and Margarita Chli. Ace: An efficient asynchronous corner tracker for event cameras. In2018 In- ternational Conference on 3D Vision (3DV), pages 653–661,

  2. [2]

    Haste: multi- hypothesis asynchronous speeded-up tracking of events

    Ignacio Alzugaray and Margarita Chli. Haste: multi- hypothesis asynchronous speeded-up tracking of events. In 31st British Machine Vision Virtual Conference (BMVC 2020), page 744. ETH Zurich, Institute of Robotics and In- telligent Systems, 2020. 2

    2020

  3. [3]

    Surf: Speeded up robust features

    Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. InComputer Vision–ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006. Proceedings, Part I 9, pages 404–

    2006

  4. [4]

    A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor.IEEE Journal of Solid-State Circuits, 49(10):2333–2341, 2014

    Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor.IEEE Journal of Solid-State Circuits, 49(10):2333–2341, 2014. 1

    2014

  5. [5]

    emoe-tracker: Environmental moe-based transformer for robust event-guided object track- ing.IEEE Robotics and Automation Letters, 10:1393–1400,

    Yucheng Chen and Lin Wang. emoe-tracker: Environmental moe-based transformer for robust event-guided object track- ing.IEEE Robotics and Automation Letters, 10:1393–1400,

  6. [6]

    Utility and feasibility of a center surround event camera

    Tobi Delbruck, Chenghan Li, Rui Graca, and Brian Mcreynolds. Utility and feasibility of a center surround event camera. In2022 IEEE International Conference on Image Processing (ICIP), pages 381–385, 2022. 1

    2022

  7. [7]

    Had: Hierarchical asymmetric dis- tillation to bridge spatio-temporal gaps in event-based object tracking, 2025

    Yao Deng, Xian Zhong, Wenxuan Liu, Zhaofei Yu, Jingling Yuan, and Tiejun Huang. Had: Hierarchical asymmetric dis- tillation to bridge spatio-temporal gaps in event-based object tracking, 2025. 1

    2025

  8. [8]

    An event-based line sensor with configurable antagonistic center surround

    Arturo Di Girolamo, Christian Metzner, and ¨Ozcan Urhan. An event-based line sensor with configurable antagonistic center surround. In2025 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5, 2025. 1

    2025

  9. [9]

    Spatio-temporal recurrent networks for event-based optical flow estimation,

    Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation,

  10. [10]

    Spike cam- era and its coding methods, 2021

    Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike cam- era and its coding methods, 2021. 1

    2021

  11. [11]

    Peiqi Duan, Boyu Li, Yixin Yang, Hanyue Lou, Minggui Teng, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventaid: Bench- marking event-aided image/video enhancement algorithms with real-captured hybrid dataset.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 47(8):6959–6973,

  12. [12]

    Distractor-aware event-based tracking.IEEE Transactions on Image Processing, 32:6129–6141, 2023

    Yingkai Fu, Meng Li, Wenxi Liu, Yuanchen Wang, Jiqing Zhang, Baocai Yin, Xiaopeng Wei, and Xin Yang. Distractor-aware event-based tracking.IEEE Transactions on Image Processing, 32:6129–6141, 2023. 1

    2023

  13. [13]

    Davison, J ¨org Conradt, Kostas Daniilidis, and Davide Scaramuzza

    Guillermo Gallego, Tobi Delbr ¨uck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, J ¨org Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey.IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154–180, 2022. 2

    2022

  14. [14]

    Springer International Publishing, 2018

    Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza.Asynchronous, Photometric Feature Tracking Using Events and Frames, page 766–781. Springer International Publishing, 2018. 7

    2018

  15. [15]

    E-raft: Dense optical flow from event cam- eras, 2021

    Mathias Gehrig, Mario Millh ¨ausler, Daniel Gehrig, and Da- vide Scaramuzza. E-raft: Dense optical flow from event cam- eras, 2021. 1, 7

    2021

  16. [16]

    Dense continuous-time optical flow from event cameras

    Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 46(7):4736–4746, 2024. 1

    2024

  17. [17]

    Mamba: Linear-time sequence mod- eling with selective state spaces, 2024

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces, 2024. 3

    2024

  18. [18]

    Etap: Event- based tracking of any point, 2025

    Friedhelm Hamann, Daniel Gehrig, Filbert Febryanto, Kostas Daniilidis, and Guillermo Gallego. Etap: Event- based tracking of any point, 2025. 1, 2, 3, 6, 7, 8

    2025

  19. [19]

    Event-based tracking any point with motion-augmented tem- poral consistency.arXiv preprint arXiv:2412.01300, 2024

    Han Han, Wei Zhai, Yang Cao, Bin Li, and Zheng-jun Zha. Event-based tracking any point with motion-augmented tem- poral consistency.arXiv preprint arXiv:2412.01300, 2024. 1, 2, 5

    Pith/arXiv arXiv 2024

  20. [20]

    Trackssm: A general motion predictor by state-space model,

    Bin Hu, Run Luo, Zelin Liu, Cheng Wang, and Wenyu Liu. Trackssm: A general motion predictor by state-space model,

  21. [21]

    ecdt: Event clustering for simultaneous fea- ture detection and tracking-, 2022

    Sumin Hu, Yeeun Kim, Hyungtae Lim, Alex Junho Lee, and Hyun Myung. ecdt: Event clustering for simultaneous fea- ture detection and tracking-, 2022. 2 9

    2022

  22. [22]

    Mamba-fetrack: Frame-event tracking via state space model, 2024

    Ju Huang, Shiao Wang, Shuai Wang, Zhe Wu, Xiao Wang, and Bo Jiang. Mamba-fetrack: Frame-event tracking via state space model, 2024. 3

    2024

  23. [23]

    Co- tracker: It is better to track together, 2024

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co- tracker: It is better to track together, 2024. 1, 3

    2024

  24. [24]

    Exploring temporally-aware features for point tracking, 2025

    In `es Hyeonsu Kim, Seokju Cho, Jiahui Huang, Jung Yi, Joon-Young Lee, and Seungryong Kim. Exploring temporally-aware features for point tracking, 2025. 1

    2025

  25. [25]

    Asynchronous event-based corner detection using adap- tive time threshold.IEEE Sensors Journal, 23(9):9512– 9522, 2023

    Jinjian Li, Li Su, Chuandong Guo, Xiangyu Wang, and Quan Hu. Asynchronous event-based corner detection using adap- tive time threshold.IEEE Sensors Journal, 23(9):9512– 9522, 2023. 2

    2023

  26. [26]

    3d feature tracking via event camera

    Siqi Li, Zhikuan Zhou, Zhou Xue, Yipeng Li, Shaoyi Du, and Yue Gao. 3d feature tracking via event camera. In 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 18974–18983, 2024. 2, 5, 6, 7

    2024

  27. [27]

    Mambalct: Boosting tracking via long-term context state space model, 2024

    Xiaohai Li, Bineng Zhong, Qihua Liang, Guorong Li, Zhiyi Mo, and Shuxiang Song. Mambalct: Boosting tracking via long-term context state space model, 2024. 3

    2024

  28. [28]

    Coherent event guided low-light video en- hancement

    Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video en- hancement. In2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10581–10591, 2023. 1

    2023

  29. [29]

    Eventups: Uncalibrated photometric stereo using an event camera

    Jinxiu Liang, Bohan Yu, Siqi Yang, Haotian Zhuang, Jieji Ren, Peiqi Duan, and Boxin Shi. Eventups: Uncalibrated photometric stereo using an event camera. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7516–7525, 2025. 1

    2025

  30. [30]

    A 128×128 120 db 15µs latency asynchronous temporal con- trast vision sensor.Solid-State Circuits, IEEE Journal of, 43: 566 – 576, 2008

    Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128×128 120 db 15µs latency asynchronous temporal con- trast vision sensor.Solid-State Circuits, IEEE Journal of, 43: 566 – 576, 2008. 1

    2008

  31. [31]

    Tracking any point with frame-event fusion network at high frame rate, 2024

    Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate, 2024. 1, 2

    2024

  32. [32]

    Mambavlt: Time-evolving multimodal state space model for vision-language tracking, 2024

    Xinqi Liu, Li Zhou, Zikun Zhou, Jianqiu Chen, and Zhenyu He. Mambavlt: Time-evolving multimodal state space model for vision-language tracking, 2024. 3

    2024

  33. [33]

    Distinctive image features from scale- invariant keypoints.International journal of computer vi- sion, 60:91–110, 2004

    David G Lowe. Distinctive image features from scale- invariant keypoints.International journal of computer vi- sion, 60:91–110, 2004. 1

    2004

  34. [34]

    An iterative image reg- istration technique with an application to stereo vision

    Bruce D Lucas and Takeo Kanade. An iterative image reg- istration technique with an application to stereo vision. In IJCAI’81: 7th international joint conference on Artificial in- telligence, pages 674–679, 1981. 1

    1981

  35. [35]

    Data-driven feature tracking for event cameras

    Nico Messikommer, Carter Fang, Mathias Gehrig, and Da- vide Scaramuzza. Data-driven feature tracking for event cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5642– 5651, 2023. 2, 3, 6, 7, 8

    2023

  36. [36]

    Dense optical tracking: Connecting the dots, 2024

    Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots, 2024. 1

    2024

  37. [37]

    Giraldo, Thierry Bouwmans, and Ananda S

    Anindya Mondal, Shashant R, Jhony H. Giraldo, Thierry Bouwmans, and Ananda S. Chowdhury. Moving object detection for event-based vision using graph spectral clus- tering. In2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), page 876–884. IEEE,

  38. [38]

    Fast event-based corner detection

    Elias Mueggler, Chiara Bartolozzi, and Davide Scaramuzza. Fast event-based corner detection. InBritish Machine Vision Conference (BMVC), 2017. 2

    2017

  39. [39]

    Stereo depth from events cameras: Concen- trate and focus on the future

    Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concen- trate and focus on the future. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6104–6113, 2022. 7

    2022

  40. [40]

    Delta: Dense efficient long-range 3d track- ing for any video, 2025

    Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evange- los Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d track- ing for any video, 2025. 1

    2025

  41. [41]

    A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds.IEEE Journal of Solid-State Circuits, 46(1):259–275,

    Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds.IEEE Journal of Solid-State Circuits, 46(1):259–275,

  42. [42]

    Teresa Serrano-Gotarredona and Bernab ´e Linares-Barranco. A 128×128 1.5% contrast sensitivity 0.9% fpn 3µs latency 4 mw asynchronous frame-free dynamic vision sensor using transimpedance preamplifiers.IEEE Journal of Solid-State Circuits, 48:827–838, 2013. 1

    2013

  43. [43]

    Improving object tracking performances with frequency learning for event cameras.IEEE Sensors Journal, 25:22419–22429, 2025

    Xilong Shang, Zhaoyuan Zeng, Xiaopeng Li, Cien Fan, and Weizheng Jin. Improving object tracking performances with frequency learning for event cameras.IEEE Sensors Journal, 25:22419–22429, 2025. 1

    2025

  44. [44]

    Blinktrack: Feature tracking over 100 fps via events and im- ages, 2024

    Yichen Shen, Yijin Li, Shuo Chen, Guanglin Li, Zhaoyang Huang, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinktrack: Feature tracking over 100 fps via events and im- ages, 2024. 2

    2024

  45. [45]

    Good features to track

    Jianbo Shi and Tomasi. Good features to track. In1994 Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 593–600, 1994. 1

    1994

  46. [46]

    Exploring his- torical information for rgbe visual tracking with mamba

    Chuanyu Sun, Jiqing Zhang, Yang Wang, Huilin Ge, Qianchen Xia, Baocai Yin, and Xin Yang. Exploring his- torical information for rgbe visual tracking with mamba. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6500–6509, 2025. 3

    2025

  47. [47]

    Revisiting color-event based tracking: A unified network, dataset, and metric, 2024

    Chuanming Tang, Xiao Wang, Ju Huang, Bo Jiang, Lin Zhu, Jianlin Zhang, Yaowei Wang, and Yonghong Tian. Revisiting color-event based tracking: A unified network, dataset, and metric, 2024. 1

    2024

  48. [48]

    Detection and tracking of point.Int J Comput Vis, 9(137-154):3, 1991

    Carlo Tomasi and Takeo Kanade. Detection and tracking of point.Int J Comput Vis, 9(137-154):3, 1991. 1

    1991

  49. [49]

    Spikepingpong: High-frequency spike vision-based robot learning for precise striking in ta- ble tennis game, 2025

    Hao Wang, Chengkai Hou, Xianglong Li, Yankai Fu, Chenx- uan Li, Ning Chen, Gaole Dai, Jiaming Liu, Tiejun Huang, and Shanghang Zhang. Spikepingpong: High-frequency spike vision-based robot learning for precise striking in ta- ble tennis game, 2025. 1

    2025

  50. [50]

    Tracking everything everywhere all at once, 2023

    Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once, 2023. 1

    2023

  51. [51]

    Mamba-fetrack v2: Revisiting state space model for frame-event based vi- sual object tracking, 2025

    Shiao Wang, Ju Huang, Qingchuan Ma, Jinfeng Gao, Chunyi Xu, Xiao Wang, Lan Chen, and Bo Jiang. Mamba-fetrack v2: Revisiting state space model for frame-event based vi- sual object tracking, 2025. 3 10

    2025

  52. [52]

    Towards low- latency event stream-based visual object tracking: A slow- fast approach.ArXiv, abs/2505.12903:null, 2025

    Shiao Wang, Xiao Wang, Liye Jin, Bowei Jiang, Lin Zhu, Lan Chen, Yonghong Tian, and Bin Luo. Towards low- latency event stream-based visual object tracking: A slow- fast approach.ArXiv, abs/2505.12903:null, 2025. 1

    Pith/arXiv arXiv 2025

  53. [53]

    Event stream- based visual object tracking: Hdetrack v2 and a high- definition benchmark.ArXiv, abs/2502.05574:null, 2025

    Shiao Wang, Xiao Wang, Chao Wang, Liye Jin, Lin Zhu, Bowei Jiang, Yonghong Tian, and Jin Tang. Event stream- based visual object tracking: Hdetrack v2 and a high- definition benchmark.ArXiv, abs/2502.05574:null, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    Visev- ent: Reliable object tracking via collaboration of frame and event flows.IEEE Transactions on Cybernetics, 54:1997– 2010, 2021

    Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visev- ent: Reliable object tracking via collaboration of frame and event flows.IEEE Transactions on Cybernetics, 54:1997– 2010, 2021

    1997

  55. [55]

    Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bowei Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline.2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), null:19248– 19257, 2023. 1

    2024

  56. [56]

    Fe-detr: Keypoint detection and tracking in low-quality image frames with events, 2024

    Xiangyuan Wang, Kuangyi Chen, Wen Yang, Lei Yu, Yan- nan Xing, and Huai Yu. Fe-detr: Keypoint detection and tracking in low-quality image frames with events, 2024. 2

    2024

  57. [57]

    Long-term visual object tracking with event cameras: An associative memory augmented tracker and a benchmark dataset.arXiv preprint arXiv:2403.05839,

    Xiao Wang, Xufeng Lou, Shiao Wang, Ju Huang, Lan Chen, and Bowei Jiang. Long-term visual object tracking with event cameras: An associative memory augmented tracker and a benchmark dataset.arXiv preprint arXiv:2403.05839,

    Pith/arXiv arXiv

  58. [58]

    Toward robust keypoint detection and tracking: A fusion approach with event-aligned image features.IEEE Robotics and Automation Letters, 9(9):8059–8066, 2024

    Xiangyuan Wang, Huai Yu, Lei Yu, Wen Yang, and Gui- Song Xia. Toward robust keypoint detection and tracking: A fusion approach with event-aligned image features.IEEE Robotics and Automation Letters, 9(9):8059–8066, 2024. 2

    2024

  59. [59]

    Mambatrack: A simple baseline for multiple object tracking with state space model, 2024

    Changcheng Xiao, Qiong Cao, Zhigang Luo, and Long Lan. Mambatrack: A simple baseline for multiple object tracking with state space model, 2024. 3

    2024

  60. [60]

    Spatialtracker: Tracking any 2d pixels in 3d space, 2024

    Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space, 2024. 1, 3

    2024

  61. [61]

    Robust tracking via mamba-based context-aware token learning, 2024

    Jinxia Xie, Bineng Zhong, Qihua Liang, Ning Li, Zhiyi Mo, and Shuxiang Song. Robust tracking via mamba-based context-aware token learning, 2024. 3

    2024

  62. [62]

    Mm-tracker: Motion mamba with margin loss for uav- platform multiple object tracking, 2024

    Mufeng Yao, Jinlong Peng, Qingdong He, Bo Peng, Hao Chen, Mingmin Chi, Chao Liu, and Jon Atli Benedikts- son. Mm-tracker: Motion mamba with margin loss for uav- platform multiple object tracking, 2024. 1, 3

    2024

  63. [63]

    Jiqing Zhang, Xin Yang, Yingkai Fu, Xiaopeng Wei, Bao- cai Yin, and B. Dong. Object tracking by jointly exploit- ing frame and event domain.2021 IEEE/CVF International Conference on Computer Vision (ICCV), null:13023–13032, 2021

    2021

  64. [64]

    Dong, Haiwei Zhang, Jianchuan Ding, Fe- lix Heide, Baocai Yin, and Xin Yang

    Jiqing Zhang, B. Dong, Haiwei Zhang, Jianchuan Ding, Fe- lix Heide, Baocai Yin, and Xin Yang. Spiking transform- ers for event-based single object tracking.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), null:8791–8800, 2022

    2022

  65. [65]

    Frame-event alignment and fusion network for high frame rate tracking.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), null:9781–9790, 2023

    Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), null:9781–9790, 2023. 1

    2023

  66. [66]

    Smmt: Siamese motion mamba with self-attention for thermal infrared target tracking, 2025

    Shang Zhang, Huanbin Zhang, Dali Feng, Yujie Cui, Ruoyan Xiong, and Cen He. Smmt: Siamese motion mamba with self-attention for thermal infrared target tracking, 2025. 3

    2025

  67. [67]

    Debattista, Qiang Zhang, Guiguang Ding, and Jungong Han

    Tianlu Zhang, K. Debattista, Qiang Zhang, Guiguang Ding, and Jungong Han. Revisiting motion information for rgb- event tracking with mot philosophy.Advances in Neural In- formation Processing Systems, 2024. 1

    2024

  68. [68]

    Harley, Bokui Shen, Gordon Wet- zstein, and Leonidas J

    Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wet- zstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking, 2023. 2, 5

    2023

  69. [69]

    Snntracker: Online high-speed multi-object tracking with spike camera.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–15, 2025

    Yajing Zheng, Chengen Li, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Snntracker: Online high-speed multi-object tracking with spike camera.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–15, 2025. 1

    2025

  70. [70]

    Event-based feature tracking with probabilistic data associ- ation

    Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based feature tracking with probabilistic data associ- ation. In2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4465–4470, 2017. 2

    2017

  71. [71]

    Realtime time synchronized event-based stereo, 2018

    Alex Zihao Zhu, Yibo Chen, and Kostas Daniilidis. Realtime time synchronized event-based stereo, 2018. 7

    2018

  72. [72]

    Cr- sot: Cross-resolution object tracking using unaligned frame and event cameras.ArXiv, abs/2401.02826:null, 2024

    Yabin Zhu, Xiao Wang, Chenglong Li, Bowei Jiang, Lin Zhu, Zhixiang Huang, Yonghong Tian, and Jin Tang. Cr- sot: Cross-resolution object tracking using unaligned frame and event cameras.ArXiv, abs/2401.02826:null, 2024. 1

    Pith/arXiv arXiv 2024

  73. [73]

    Learning graph- embedded key-event back-tracing for object tracking in event clouds.Advances in Neural Information Processing Systems, 2022

    Zhiyu Zhu, Junhui Hou, and Xianqiang Lyu. Learning graph- embedded key-event back-tracing for object tracking in event clouds.Advances in Neural Information Processing Systems, 2022

    2022

  74. [74]

    Cross- modal orthogonal high-rank augmentation for rgb-event transformer-trackers.2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV), null:21988–21998, 2023

    Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross- modal orthogonal high-rank augmentation for rgb-event transformer-trackers.2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV), null:21988–21998, 2023. 1 1 E-TraMamba: A New Paradigm for Efficient Long-Term 3D Feature Tracking with Event Cameras Supplementary Material

    2023

  75. [75]

    Our model supports bothstereoandmonocular+depth configurations, which we describe separately below

    Architecture Details This section provides complete implementation details for theE-TraMambaframework, including precise network architectures, channel dimensions, and hyperparameters. Our model supports bothstereoandmonocular+depth configurations, which we describe separately below. 7.1. Input Representations Stereo Configuration.The input consists of sy...

  76. [76]

    Overview As an implementation supplement to Sec

    Dataset Construction Details 8.1. Overview As an implementation supplement to Sec. 3.4, we describe how long PointOdyssey sequences are segmented into train- ing windows forEvD-PointOdyssey. Each scene is rendered into high-frame-rate RGB videos (240 FPS), converted to events using asingle, fixedcon- trast thresholdθ= 0.02(calibrated once for the dataset)...