SpikingMOT achieves SOTA tracking with 86.7% less energy

Multi-object tracking (MOT) traditionally relies on densely activated artificial neural networks that consume significant energy, but new research questions whether such dense responses are necessary. A team led by Yiding Sun proposes SpikingMOT, a spike-driven tracker that leverages spiking neural networks (SNNs) to model trajectory dynamics with extreme efficiency. Instead of dense activations, SpikingMOT decomposes each trajectory state into pseudo-trajectory bases and uses the current prediction error to calibrate the posterior for next-frame prediction—a brain-inspired loop.

In extensive experiments, SpikingMOT achieves state-of-the-art results: 74.9 HOTA on SportsMOT and 56.5 HOTA on DanceTrack, while slashing parameters by 72% and energy use by 86.7% compared to conventional trackers. The theoretical foundation lies in a formal activation sparsity preference (ASP) analysis showing sparse gating is no worse than state-independent dropout at the same activation rate.

This work marks the first time spiking neural networks have been successfully applied to multi-object tracking, opening a promising direction for efficient, low-power visual perception systems. The code and data are available at the provided arXiv link.