SpikingMOT: A Spike-Driven Multi-Object Tracker
Abstract Overview
This paper studies multi-object tracking from the perspective of activation sparsity, arguing that dense neural responses may be unnecessary and sometimes counterproductive for trajectory prediction. The authors introduce SpikingMOT, a spike-driven MOT framework based on spiking neural networks that models motion through pseudo-trajectory basis decomposition and prediction-error calibration. The method predicts boxes from recent trajectory history, selects informative motion bases through spike-gated dynamics, and updates basis weights using residuals from matched detections. The paper evaluates the approach on DanceTrack, SportsMOT, SeaDroneSee, and MOT17, with additional ablations analyzing sparse alternatives, core modules, basis count, neuron types, and detector dependence.
Novelty
The paper frames an activation sparsity preference (ASP) phenomenon for MOT trajectory prediction and uses it to motivate a tracking architecture built around spiking neural networks. It also introduces a task-specific closed loop that combines pseudo-trajectory decomposition with prediction-error calibration, rather than simply substituting a generic SNN backbone into an existing tracker.
Results
Across benchmarks, SpikingMOT reports state-of-the-art HOTA of 74.9 on SportsMOT and 56.5 on DanceTrack, while also achieving strong results on SeaDroneSee and comparable performance on MOT17. The authors report that, relative to ANN baselines, the method reduces motion-predictor parameters by 72% and estimated per-frame energy by 86.7%. Ablation studies further show that the full design outperforms sparse ANN alternatives and that gains are consistent across different detectors on SportsMOT.
Key Points
- SpikingMOT uses SNN-based sparse spike gating to select among multiple pseudo-trajectory motion bases for next-frame prediction.
- A prediction-error calibration module feeds matched detection residuals back into the basis posterior, forming a closed-loop motion inference process across frames.
- Empirical results show strong association-sensitive tracking performance, especially on dynamic datasets such as SportsMOT and DanceTrack, alongside substantially lower estimated model complexity and energy.