Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

SpikingMOT: A Spike-Driven Multi-Object Tracker

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A spike-driven trajectory predictor outperforms dense ANN motion models on association-heavy benchmarks at a fraction of the estimated energy.

desk verdict A genuinely new spike-driven motion predictor with solid empirical claims; the efficiency headline is softer than it looks because it is operation-count-based and does not cover the full pipeline. read the letter →

arxiv 2607.19875 v1 pith:7CAIVW6C submitted 2026-07-22 cs.NE

Yiding Sun , Xiangyang Yang , Dongxu Zhang , Qirui Wang , Zijie Xu , Wenxuan Liu , Shuiwang Li , Jihua Zhu
show 2 more authors
Zhaofei Yu Tiejun Huang
This is my paper · ORCID
classification cs.NE
keywords multi-objecttrackingspikingneuralnetworksmotionpredictionactivationsparsitypseudo-trajectorybasisprediction-errorcalibrationenergy-efficient
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that sparse, spike-driven motion prediction is not merely a way to save computation in multi-object tracking, but actively improves trajectory association. It formulates Activation Sparsity Preference (ASP), the observation that moderately suppressing activations in a dense motion predictor improves tracking quality, and proves that state-conditioned sparse gating is no worse than state-independent dropout at the same activation rate. To exploit this, the authors build SpikingMOT, the first RGB spike-driven tracker, which represents each trajectory state as a sparse combination of K learnable pseudo-trajectory bases selected by LIF neuron firing, and recalibrates the basis posterior from the prediction error of matched detections. Across SportsMOT, DanceTrack, SeaDroneSee, and MOT17, it reports competitive or state-of-the-art HOTA, including 74.9 on SportsMOT and 56.5 on DanceTrack, with 72% fewer motion-predictor parameters and an estimated 86.7% lower energy than the dense ANN baseline. If true, this means tracking accuracy and sparsity-based efficiency can go together, opening SNNs as a practical direction for association-heavy perception.

What carries the argument

The load-bearing mechanism is a two-step closed loop. First, motion-basis decomposition: the recent trajectory is encoded into a state vector, and each of K learnable basis functions proposes a one-step displacement; LIF (leaky integrate-and-fire) neurons transform the dense basis posterior into sparse binary spikes, and the aggregated spikes gate a weighted reconstruction of the predicted box. Second, prediction-error calibration: after Hungarian association, the scale-normalized residual between prediction and matched detection is fed back through a calibration head to adjust the basis posterior (via a logit correction in softmax space) for the next prediction. The paper also contributes t

What would settle it

Measure the energy of SpikingMOT and a matched dense ANN tracker (e.g., MambaTrack) on a neuromorphic chip running the same MOT benchmark; if the measured per-frame energy ratio is close to 1 or worse than the dense baseline, the central efficiency claim collapses even though the HOTA results might stand. Alternatively, find a trajectory distribution where state-independent dropout outperforms state-conditioned gating under identical activation rate, which would contradict the ASP theorem.

Watch

Extended reading notes

Core claim

The central discovery is that trajectory prediction in MOT exhibits an activation sparsity preference, and that a spiking neural network can exploit it. Concretely, the paper proves that state-conditioned sparse gating is no worse than state-independent dropout under the same activation rate, giving theoretical license to replace denser ANN motion predictors with spike-gated ones. SpikingMOT realizes this by decomposing each encoded motion state into K=8 learnable basis-conditioned displacement proposals, gating them with LIF-neuron spikes so only a few bases contribute, and closing the loop with a calibration head that turns the residual between the predicted box and the matched detection i

Load-bearing premise

The load-bearing assumption is that the claimed 86.7% energy saving holds on real hardware; the paper itself states that energy is estimated from operation counts, not measured on neuromorphic chips, so the efficiency result rests on the per-operation energy constants and sparsity assumptions being accurate on actual hardware.

Editorial extensions

If this is right

  • On association-sensitive benchmarks (DanceTrack, SportsMOT), a sparse spike-driven motion predictor can match or beat dense ANN predictors such as MambaTrack and TrackSSM.
  • The tracker's gains hold across three different detectors, indicating the benefit comes from the motion model itself, not from a particular detector.
  • Under an operation-count energy model, replacing the dense predictor with the spiking one reduces motion-predictor parameters by roughly 72% and per-frame energy by about 86.7%.
  • The reported success depends on two complementary components—basis decomposition and error calibration—and ablations show both are needed for the full HOTA gain.
  • MOT trackers need no longer assume dense activation is required for motion reasoning; sparsity can be treated as a learnable, beneficial prior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the operation-count energy model transfers to real neuromorphic hardware, the same design could make low-power edge tracking feasible for drones and robots—but that transfer is exactly what remains unmeasured.
  • The ASP principle is stated generally; it could be tested in other dense prediction tasks (e.g., optical flow, single-object tracking) where dropout-style sparsity has been studied.
  • A natural follow-up is to replace the LIF gates with non-spiking sparse gates at the same activation rate; if performance drops, the spiking dynamics themselves, rather than sparsity alone, carry the benefit.
  • Combining SpikingMOT with an event-camera detector would yield a fully spike-driven tracking pipeline, which the paper hints at but does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes SpikingMOT, a spike-driven motion predictor for multi-object tracking. The core idea is to replace dense ANN motion predictors with a spiking module that decomposes a trajectory state into K learnable pseudo-trajectory bases, gates those bases through LIF neurons, and recalibrates the basis posterior using the current prediction error. The authors motivate this design with an empirically observed 'activation sparsity preference' (ASP) and a theoretical claim, deferred to Supp. A, that state-conditioned sparse gating is no worse than state-independent dropout at the same activation rate. On association-heavy benchmarks, SpikingMOT-YL104M (with YOLOX detections) reports 74.9 HOTA on SportsMOT and 56.5 HOTA on DanceTrack, ahead of the MambaTrack baseline, while the motion predictor uses 3.6M parameters and an estimated 3.6 mJ/frame versus 13.3M and 26.9 mJ/frame. The paper also includes SpikeYOLO-based fully spiking variants, which reach lower absolute accuracy (e.g., 55.4 and 46.6 HOTA on SportsMOT/DanceTrack for SY69M). Extensive ablations study sparse-ANN alternatives, core components, basis count, neuron type, and detector-wise generalization.

Significance. If the efficiency numbers were measured rather than modeled, SpikingMOT would be the first spike-driven motion predictor to match or beat dense ANN predictors on association-heavy MOT benchmarks while using far fewer parameters and much lower per-frame energy. The performance claim is supported by controlled detector-wise comparisons (Table 4, Fig. 5) and by a clear ablation table (Table 3) showing that sparsity alone, dropout, ℓ1 regularization, or top-k pruning does not reproduce the gain. The paper is also commendable for stating in the Conclusion that energy is estimated from operation counts, not measured on neuromorphic hardware. However, the abstract's unqualified 'reducing the parameters and energy by 72% and 86.7%' overstates what is established: the reduction applies only to the motion predictor, not the full detection–tracking pipeline, and the energy figure rests on an operation-count model. Since efficiency is half of the central claim, this is a load-bearing limitation rather than a presentation detail.

major comments (5)
  1. [Abstract and Table 1 footnote] The abstract states that SpikingMOT reduces parameters and energy by 72% and 86.7%, without qualification. Table 1's footnote says 'Para. counts only the learnable parameters of the motion predictor,' and the Conclusion says energy is estimated from operation counts. The headline numbers therefore refer to the motion-predictor module, not to an end-to-end spike-driven tracker. Given that the best reported HOTA numbers use the ANN YOLOX detector, the efficiency comparison cannot be read as a property of a complete SpikingMOT system. Please rescope the abstract and Section 4 claims, or provide end-to-end parameter and energy estimates that include the detector. This is not merely wording; it changes what the reader can conclude about the system's efficiency.
  2. [Conclusion and Supp. C] The 86.7% energy reduction is based on an operation-count energy model with per-operation constants, as stated in Supp. C and acknowledged in the Conclusion. Operation-count models typically assign fixed per-MAC or per-spike costs and omit memory access, spike routing, idle power, and control overhead, which can dominate on neuromorphic or edge hardware. The efficiency claim is central to the paper's contribution, so the current estimate is insufficient. Please include at least a sensitivity analysis over the assumed energy constants, or clearly present the energy numbers as theoretical operation-count estimates in every place they appear, including the abstract and Figure 2.
  3. [Method (first paragraph) and Supp. A] The main text says 'Theoretical analysis shows that sparse gating is no worse than state-independent dropout under the same activation rate' and cites Supp. A, but the theorem, its assumptions, and its proof are not stated in the main text. This is a foundational motivation for the spiking design. Please state the theorem and its conditions in the main text (or at least a precise statement with a proof sketch); otherwise readers cannot verify that the architecture is actually derived from the stated principle rather than merely loosely inspired by it.
  4. [Related Works and Tables 1–2] The Related Works section says SpikingMOT 'together with SpikeYOLO, enables fully spike-driven tracking under the TBD paradigm.' The fully spike-driven variants (SpikingMOT-SY23M/SY69M) reach 55.4/46.6 HOTA on SportsMOT/DanceTrack with SY69M, while the headline 74.9/56.5 numbers use YOLOX, an ANN detector. The paper should clearly separate 'spike-driven motion predictor with an ANN detector' from 'fully spike-driven pipeline.' The current abstract and Figure 2 can be read as claiming that the state-of-the-art results are achieved by a fully spike-driven tracker, which is not the case. Please clarify this distinction in the abstract and in the benchmark discussion.
  5. [Table 1 and Table 2] It is unclear whether the energy entries for competing ANN trackers (e.g., MambaTrack 26.9 mJ) include the same operation-count assumptions as the authors' own entries. The paper says 'Eng. reports per-frame energy under the corresponding detector,' but the detector model and per-operation constants are not described in the main text. To make the comparison trustworthy, please specify in the caption or implementation details that the energy numbers are computed with the same constants and same detector for all rows, and state the detector used for each energy row.
minor comments (5)
  1. [Table 1] The entries '3.65.4' and '3.6 3.61×1' appear to be formatting errors (missing space) for SpikingMOT-YL104M. Please correct the Para/Eng columns.
  2. [Table 2] The same formatting issue appears in Table 2 for SpikingMOT-YL104M. Also, the table lists 'Para.' and 'Eng.' columns for the spiking rows only; a sentence in the caption describing how to read the N/A entries would improve clarity.
  3. [Fig. 6] The lower heatmaps' color scale and normalization are not defined. It is also unclear whether the historical frame index is the absolute video frame or a relative index. Please clarify.
  4. [Method, Eq. (11)] The definition of D as 'maximum integer value emitted during training' is ambiguous: it is not stated whether D is a per-neuron, per-basis, or global maximum, nor how it is computed during training (fixed after training or updated online). Please define precisely.
  5. [Implementation Details] The paper mentions surrogate gradient details in Supp. D.1 but does not state the surrogate function in the main text. Since the LIF dynamics in Eq. (10) are central, a one-sentence specification (e.g., 'we use a rectangular surrogate') would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical benchmarks and controlled ablations make the core claims self-contained.

full rationale

The paper's central empirical claims are self-contained. HOTA numbers are measured on external benchmarks (DanceTrack, SportsMOT, SeaDroneSee, MOT17) using the MambaTrack/TrackSSM training and association protocols, and Table 4 further holds the detector fixed while comparing motion predictors, showing the gains are not an artifact of a fitted target. The prediction-error calibration is a trained recurrent module (Eqs. 13-16) with a clear train/inference distinction, not a quantity derived from the reported result. The only theoretical element, the ASP claim that state-conditioned sparse gating is no worse than state-independent dropout at the same activation rate, is deferred to Supp A; as stated it is a near-tautological special-case argument and is used as motivation rather than as a derivation of the empirical HOTA gains, so it does not make the empirical claims circular. The efficiency claims are explicitly scoped by the Table 1 footnote ('Para. counts only the learnable parameters of the motion predictor') and by the Conclusion's statement that 'energy consumption is estimated from operation counts rather than measured on neuromorphic hardware'; this is a limitation, not a circularity. No load-bearing self-citations or imported uniqueness theorems appear.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central model is a trained system, so most of its content is learned rather than postulated. The listed free parameters are manually chosen hyperparameters and energy-model constants; the axioms are the deferred ASP theorem, the generalization of Fig. 1, and the two brain-inspired representational assumptions. No new physical entity is introduced; pseudo-trajectory bases are a representational device.

free parameters (5)
  • K (number of pseudo-trajectory bases) = 8
    Selected via ablation in Table 3c; balances motion diversity and firing rate. Hyperparameter, not derived from theory.
  • H (history length) = 5
    Set in Implementation Details; no ablation shown in main text.
  • Loss weights α, β, µ and ε = α=1.0, β=1e-3, µ=1e-2, ε=1e-6
    Hand-tuned (Supp D.3); balance prediction, calibration, spike sparsity, and basis diversity terms in Eq. (3).
  • LIF neuron hyperparameters (λ, ϑ, T, D) = T×D = 4×1 or 1×1; λ, ϑ in Supp D
    Leakage and threshold control firing; D is max spike count during training used to clip gates Eq. (11); T is inference timestep choice in Table 1.
  • Per-operation energy constants for Eng. = Not listed in main text (Supp C)
    86.7% energy reduction is computed from operation-count energy model, not hardware measurement; conclusion admits estimation.
assumptions (5)
  • ad hoc to paper State-conditioned sparse gating is no worse than state-independent dropout at the same activation rate (Supp A).
    Stated in Method as the theoretical basis for ASP; proof deferred to unexplained Supp A, so its assumptions cannot be audited in the main text.
  • domain assumption Moderate activation suppression improves MOT trajectory prediction (ASP), as observed in a MambaTrack dropout study (Fig. 1).
    Single-model, single-figure observation without error bars; generalized to a design principle for the proposed architecture.
  • domain assumption Continuous motion can be represented by sparse selection among K pseudo-trajectory bases (population coding analogy).
    Underlies motion-basis decomposition; supported only by the paper's ablations, not independent biological evidence.
  • domain assumption Matched-detection residual is a reliable corrective signal for the next frame's basis posterior (predictive coding analogy).
    Calibration head is trained on this residual; at inference the residual comes from Hungarian matching, whose errors are not modeled.
  • standard math LIF neuron dynamics with surrogate gradient provide a valid differentiable approximation for end-to-end training.
    Standard SNN training assumption cited to Neftci et al. 2019; needed for backpropagation through Eq. (10).
invented entities (2)
  • Pseudo-trajectory bases b_k
    purpose: K learnable motion hypotheses; each proposes a candidate displacement for the next box (Eq. 9).
    A representational construct without external evidence; falsifiable only via the paper's ablations/benchmarks.
  • Activation Sparsity Preference (ASP)
    purpose: Named principle that sparse activation benefits trajectory prediction; motivates switching from dense ANNs to SNNs.
    Empirical phenomenon observed in one baseline; not independently established outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SpikingMOT: A Spike-Driven Multi-Object Tracker." pith.science (2026). https://pith.science/paper/7CAIVW6C

@misc{pith2026260719875,
  author       = {Pith},
  title        = {Pith review of: SpikingMOT: A Spike-Driven Multi-Object Tracker},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7CAIVW6C}},
  note         = {Machine review of arXiv:2607.19875}
}
read the original abstract

Multi-object tracking (MOT) plays a fundamental role in visual perception, where accurate trajectory prediction is essential for reliable target association under complex motion patterns. Recent trackers have improved motion modeling with densely activated artificial neural networks, yet they largely overlook whether such dense responses are necessary for trajectory prediction. In this paper, we formulate activation sparsity preference (ASP) by tackling two key questions: 1. How can we identify a model architecture that appropriately and formally explains ASP, and 2. How can we translate this explanation into competitive tracking performance. Theoretical analysis shows that sparse gating is no worse than state-independent dropout under the same activation rate. Based on this insight, SpikingMOT is proposed as a spike-driven tracker that adaptively models sparse trajectory dynamics with spiking neural networks (SNNs). Specifically, SpikingMOT decomposes each trajectory state into pseudo-trajectory bases and uses the current prediction error to calibrate the posterior for next-frame prediction. With this brain-inspired loop, SpikingMOT achieves state-of-the-art performance in extensive experiments, 74.9 HOTA on SportsMOT and 56.5 HOTA on DanceTrack, while reducing the parameters and energy by 72% and 86.7%, respectively. These results bring SNNs into MOT, opening a promising direction for efficient tracking.

Figures

Figures reproduced from arXiv: 2607.19875 by the authors.

Figure 1
Figure 1. Empirical evidence for ASP. (a) Tracking per￾formance under progressively reduced activation rates in MambaTrack with all protocols fixed. Moderate suppression consistently improves HOTA within a nontrivial range. (b) Tracking cases showing that uniformly dense responses can amplify redundant motion cues and cause inaccurate pre￾diction or identity switches, whereas moderate suppression preserves informative motion … view at source ↗
Figure 2
Figure 2. Inspired by ASP, SpikingMOT replaces dense [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of SpikingMOT. Given historical boxes of each active track, SpikingMOT encodes the recent trajectory, decomposes the motion state into pseudo-trajectory bases, selects informative bases through LIF spike firing, reconstructs a motion prior for association, and uses the matched detection residual to calibrate future basis responses. into K basis-conditioned motion responses. The calibrated posterior from the… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of SpikingMOT in DanceTrack and SportsMOT dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Detector￾wise gain over Mam￾baTrack. ness, whereas too many introduce overlapping hypotheses and unnecessary firing. A moderate basis set therefore pro￾vides sufficient motion diversity while preserving selective and energy-efficient activation. Spiking neuron types. T…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Parameter-Efficient Fine-Tuning for Spiking Point Cloud Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    SpikePEFT matches or beats full fine-tuning on spiking point-cloud benchmarks while updating only ~5% of parameters via neuron-dynamics and silent-state adapters.

Reference graph

Works this paper leans on

75 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2112.10048 , year=

    Predictive coding theories of cortical function , author=. arXiv preprint arXiv:2112.10048 , year=

    arXiv

  2. [2]

    and Mostafa, Hesham and Zenke, Friedemann , title =

    Neftci, Emre O. and Mostafa, Hesham and Zenke, Friedemann , title =. IEEE Signal Processing Magazine , volume =. 2019 , doi =

    2019

  3. [3]

    arXiv preprint arXiv:1904.07850 , year=

    Objects as points , author=. arXiv preprint arXiv:1904.07850 , year=

    arXiv 1904

  4. [4]

    Nature Neuroscience , volume =

    Predictive Coding in the Visual Cortex: A Functional Interpretation of Some Extra-Classical Receptive-Field Effects , author =. Nature Neuroscience , volume =. 1999 , publisher =

    1999

  5. [5]

    Neural Computation , volume =

    Long Short-Term Memory , author =. Neural Computation , volume =

  6. [6]

    ISSCC , pages =

    1.1 Computing's Energy Problem (and What We Can Do About It) , author =. ISSCC , pages =. 2014 , publisher =

    2014

  7. [7]

    NeurIPS , pages =

    Attention Is All You Need , author =. NeurIPS , pages =

  8. [8]

    Nature Reviews Neuroscience , volume =

    Information Processing with Population Codes , author =. Nature Reviews Neuroscience , volume =. 2000 , publisher =

    2000

Show all 75 references
  1. [9]

    ICCV , pages =

    Cui, Yutao and Zeng, Chenkai and Zhao, Xiaoyu and Yang, Yichun and Wu, Gangshan and Wang, Limin , title =. ICCV , pages =

  2. [10]

    ICCV , month =

    Duan, Kaiwen and Bai, Song and Xie, Lingxi and Qi, Honggang and Huang, Qingming and Tian, Qi , title =. ICCV , month =

  3. [11]

    COLM , month =

    Gu, Albert and Dao, Tri , title =. COLM , month =

  4. [12]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. ECCV , pages =. 2014 , doi =

    2014

  5. [13]

    ACM MM , year =

    Xiao, Changcheng and Cao, Qiong and Luo, Zhigang and Lan, Long , title =. ACM MM , year =

  6. [14]

    Tracking Without Bells and Whistles , booktitle =

    Bergmann, Philipp and Meinhardt, Tim and Leal-Taix. Tracking Without Bells and Whistles , booktitle =

  7. [15]

    NeurIPS , pages =

    Xu, Zijie and Bu, Tong and Hao, Zecheng and Ding, Jianhao and Yu, Zhaofei , title =. NeurIPS , pages =

  8. [16]

    ICLR , year =

    Xu, Zijie and Shi, Xinyu and Dong, Yiting and Huang, Zihan and Yu, Zhaofei , title =. ICLR , year =

  9. [17]

    TNNLS , volume =

    Spiking Neural Network for Ultralow-Latency and High-Accurate Object Detection , author =. TNNLS , volume =

  10. [18]

    TNNLS , volume =

    Spiking Deep Residual Networks , author =. TNNLS , volume =

  11. [19]

    IJCAI , pages =

    Event-Based Action Recognition Using Motion Information and Spiking Neural Networks , author =. IJCAI , pages =

  12. [20]

    IEEE Access , volume =

    Wang, Song and Wang, Zhu and Li, Can and Qi, Xiaojuan and So, Hayden Kwok-Hay , title =. IEEE Access , volume =

  13. [21]

    IEEE Internet of Things Journal , volume =

    Zhong, Pengzhi and Wang, Xinzhe and Zeng, Dan and Zhou, Qihua and He, Feixiang and Li, Shuiwang , title =. IEEE Internet of Things Journal , volume =

  14. [22]

    NeurIPS , pages =

    Ren, Shaoqing and He, Kaiming and Girshick, Ross and Sun, Jian , title =. NeurIPS , pages =

  15. [23]

    ECCV , pages =

    Wang, Zhongdao and Zheng, Liang and Liu, Yixuan and Li, Yali and Wang, Shengjin , title =. ECCV , pages =

  16. [24]

    CVPR , pages =

    Redmon, Joseph and Divvala, Santosh and Girshick, Ross and Farhadi, Ali , title =. CVPR , pages =

  17. [25]

    IJCV , volume =

    Zhang, Yifu and Wang, Chunyu and Wang, Xinggang and Zeng, Wenjun and Liu, Wenyu , title =. IJCV , volume =. 2021 , doi =

    2021

  18. [26]

    ECCV , pages =

    Carion, Nicolas and Massa, Francisco and Synnaeve, Gabriel and Usunier, Nicolas and Kirillov, Alexander and Zagoruyko, Sergey , title =. ECCV , pages =

  19. [27]

    CVPR , pages =

    Meinhardt, Tim and Kirillov, Alexander and Leal-Taix. CVPR , pages =

  20. [28]

    ECCV , pages =

    Zeng, Fangao and Dong, Bin and Zhang, Yuang and Wang, Tiancai and Zhang, Xiangyu and Wei, Yichen , title =. ECCV , pages =

  21. [29]

    CoRR , volume =

    Sun, Peize and Cao, Jinkun and Jiang, Yi and Zhang, Rufeng and Xie, Enze and Yuan, Zehuan and Wang, Changhu and Luo, Ping , title =. CoRR , volume =. 2020 , eprint =

    2020

  22. [30]

    CVPR , pages =

    Cai, Jiarui and Xu, Mingze and Li, Wei and Xiong, Yuanjun and Xia, Wei and Tu, Zhuowen and Soatto, Stefano , title =. CVPR , pages =

  23. [31]

    Recherches Quantitatives sur l'Excitation

    Lapicque, Louis , journal =. Recherches Quantitatives sur l'Excitation

  24. [32]

    Engineering , volume =

    1000 Faster Camera and Machine Vision with Ordinary Devices , author =. Engineering , volume =

  25. [33]

    ICCV , pages =

    Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks , author =. ICCV , pages =

  26. [34]

    NeurIPS , volume =

    Long Short-Term Memory and Learning-to-Learn in Networks of Spiking Neurons , author =. NeurIPS , volume =

  27. [35]

    ICIP , pages =

    Bewley, Alex and Ge, Zongyuan and Ott, Lionel and Ramos, Fabio and Upcroft, Ben , title =. ICIP , pages =. 2016 , doi =

    2016

  28. [36]

    ICIP , pages =

    Wojke, Nicolai and Bewley, Alex and Paulus, Dietrich , title =. ICIP , pages =. 2017 , doi =

    2017

  29. [37]

    , title =

    Kalman, Rudolf E. , title =. Journal of Basic Engineering , volume =. 1960 , doi =

    1960

  30. [38]

    CVPR , pages =

    Cao, Jinkun and Pang, Jiangmiao and Weng, Xinshuo and Khirodkar, Rawal and Kitani, Kris , title =. CVPR , pages =

  31. [39]

    ICIP , pages =

    Maggiolino, Gerard and Ahmad, Adnan and Cao, Jinkun and Kitani, Kris , title =. ICIP , pages =

  32. [40]

    Neural Networks , volume =

    Xiao, Changcheng and Cao, Qiong and Zhong, Yujie and Lan, Long and Zhang, Xiang and Luo, Zhigang and Tao, Dacheng , title =. Neural Networks , volume =. 2024 , doi =

    2024

  33. [41]

    CVPR , pages =

    Lv, Weiyi and Huang, Yuhang and Zhang, Ning and Lin, Ruei-Sung and Han, Mei and Zeng, Dan , title =. CVPR , pages =

  34. [42]

    AAAI , volume =

    Luo, Run and Song, Zikai and Ma, Lintao and Wei, Jinlin and Yang, Wei and Yang, Min , title =. AAAI , volume =

  35. [43]

    ICASSP , year =

    Huang, Hsiang-Wei and Yang, Cheng-Yen and Chai, Wenhao and Jiang, Zhongyu and Hwang, Jenq-Neng , title =. ICASSP , year =

  36. [44]

    ECCV , series =

    Luo, Xinhao and Yao, Man and Chou, Yuhong and Xu, Bo and Li, Guoqi , title =. ECCV , series =. 2024 , doi =

    2024

  37. [45]

    NeurIPS , volume =

    Yao, Man and Hu, Jiakui and Zhou, Zhaokun and Yuan, Li and Tian, Yonghong and Xu, Bo and Li, Guoqi , title =. NeurIPS , volume =. 2023 , doi =

    2023

  38. [46]

    TPAMI , volume =

    Yao, Man and Qiu, Xuerui and Hu, Tianxiang and Hu, Jiakui and Chou, Yuhong and Tian, Keyu and Liao, Jianxing and Leng, Luziwei and Xu, Bo and Li, Guoqi , title =. TPAMI , volume =. 2025 , doi =

    2025

  39. [47]

    AAAI , volume =

    Lei, Zhenxin and Yao, Man and Hu, Jiakui and Luo, Xinhao and Lu, Yanye and Xu, Bo and Li, Guoqi , title =. AAAI , volume =. 2025 , doi =

    2025

  40. [48]

    ICML , volume =

    Zou, Shihao and Li, Qingfeng and Ji, Wei and Li, Jingjing and Yang, Yongkui and Li, Guoqi and Dong, Chao , title =. ICML , volume =

  41. [49]

    ECCV Workshops , pages =

    Performance Measures and a Data Set for Multi-Target, Multi-Camera Tracking , author =. ECCV Workshops , pages =

  42. [50]

    Evaluating Multiple Object Tracking Performance: The

    Bernardin, Keni and Stiefelhagen, Rainer , journal =. Evaluating Multiple Object Tracking Performance: The

  43. [51]

    NeurIPS , volume =

    Zhu, Rui-Jie and Wang, Ziqing and Gilpin, Leilani and Eshraghian, Jason , title =. NeurIPS , volume =. 2024 , doi =

    2024

  44. [52]

    ICASSP , pages =

    Jiang, Xiaoyang and Zhang, Qiang and Sun, Jingkai and Cao, Jiahang and Ma, Jingtong and Xu, Renjing , title =. ICASSP , pages =. 2025 , doi =

    2025

  45. [53]

    CoRR , volume =

    Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian , title =. CoRR , volume =. 2021 , eprint =

    2021

  46. [54]

    ECCV , pages =

    Zhang, Yifu and Sun, Peize and Jiang, Yi and Yu, Dongdong and Weng, Fucheng and Yuan, Zehuan and Luo, Ping and Liu, Wenyu and Wang, Xinggang , title =. ECCV , pages =

  47. [55]

    CVPR , pages =

    Girshick, Ross and Donahue, Jeff and Darrell, Trevor and Malik, Jitendra , title =. CVPR , pages =

  48. [56]

    CVPR , pages=

    Sdtrack: A baseline for event-based tracking via spiking neural networks , author=. CVPR , pages=

  49. [57]

    IJCV , volume=

    Hota: A higher order metric for evaluating multi-object tracking , author=. IJCV , volume=. 2021 , publisher=

    2021

  50. [58]

    CVPR , pages=

    SpikeTrack: A Spike-driven Framework for Efficient Visual Tracking , author=. CVPR , pages=

  51. [59]

    Nature , volume=

    Towards artificial general intelligence with hybrid Tianjic chip architecture , author=. Nature , volume=. 2019 , publisher=

    2019

  52. [60]

    CVPR , pages =

    Rezatofighi, Hamid and Tsoi, Nathan and Gwak, JunYoung and Sadeghian, Amir and Reid, Ian and Savarese, Silvio , title =. CVPR , pages =

  53. [61]

    CVPR , pages =

    Pang, Jiangmiao and Qiu, Linlu and Li, Xia and Chen, Haofeng and Li, Qi and Darrell, Trevor and Yu, Fisher , title =. CVPR , pages =

  54. [62]

    Global Tracking Transformers , booktitle =

    Zhou, Xingyi and Yin, Tianwei and Koltun, Vladlen and Kr. Global Tracking Transformers , booktitle =

  55. [63]

    , title =

    Khanna, Dheeraj and Bright, Jerrin and Chen, Yuhao and Zelek, John S. , title =. CVPRW , pages =

  56. [64]

    Science China Information Sciences , volume =

    Hu, Bin and Luo, Run and Liu, Zelin and Wang, Cheng and Liu, Wenyu , title =. Science China Information Sciences , volume =. 2026 , doi =

    2026

  57. [65]

    CVPR , pages =

    Sun, Peize and Cao, Jinkun and Jiang, Yi and Yuan, Zehuan and Bai, Song and Kitani, Kris and Luo, Ping , title =. CVPR , pages =

  58. [66]

    WACV , pages =

    Varga, Leon Amadeus and Kiefer, Benjamin and Messmer, Martin and Zell, Andreas , title =. WACV , pages =

  59. [67]

    IJCV , volume =

    Dendorfer, Patrick and O. IJCV , volume =. 2021 , doi =

    2021

  60. [68]

    Align then Adapt: Rethinking Parameter-Efficient Transfer Learning in

    Sun, Yiding and Zhu, Jihua and Cheng, Haozhe and Lu, Chaoyi and Yang, Zhichuan and Chen, Lin and Wang, Yaonan , journal =. Align then Adapt: Rethinking Parameter-Efficient Transfer Learning in

  61. [69]

    Wang, Yankai and Sun, Yiding and Wang, Qirui and Li, Pengbo and Lu, Chaoyi and Zhang, Dongxu , booktitle =

  62. [70]

    Guo, Zihao and Zhu, Jihua and Liu, Jian and Mian, Ajmal Saeed , journal =

  63. [71]

    2025 , publisher =

    Li, Chenyang and Liu, Wenxuan and Gong, Guoqiang and Ding, Xiaobo and Zhong, Xian , journal =. 2025 , publisher =

    2025

  64. [72]

    Naval Research Logistics Quarterly , volume =

    The Hungarian Method for the Assignment Problem , author =. Naval Research Logistics Quarterly , volume =

  65. [73]

    Liu, Wenxuan and Deng, Yao and Chen, Kang and Zhong, Xian and Yu, Zhaofei and Huang, Tiejun , booktitle =

  66. [74]

    Dong, Yiting and Ding, Jianhao and Xu, Zijie and Bu, Tong and Yu, Zhaofei and Huang, Tiejun , booktitle =

  67. [75]

    ECCV , year =

    Tri-Efficient Transfer Learning for Point Cloud Videos , author =. ECCV , year =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.