Unleashing SNNs in Object Detection with Time-Evolving Neuron and Dual-Stream Spiking Attention

ICLR 2026 Conference Submission17300 Authors

Keywords: spiking neural network, object detection

Abstract: Brain-inspired Spiking Neural Networks (SNNs) offer remarkable energy efficiency but still lag behind Artificial Neural Networks (ANNs) in fundamental tasks like object detection, primarily due to the precision bottleneck and limited spatial modeling. To narrow this gap, we propose \textit{SpikeDet}, a fully spiking object detector that redefines both the microscopic neuron model and macroscopic attention mechanism. At its core, the bio-inspired \textit{TE-LIF} neuron, with time-evolving membrane dynamics, enhances representational precision and achieves finer input pattern recognition, while maintaining computational efficiency. Building upon this, the proposed \textit{Dual-Stream Spiking Attention} employs a QV-only design that integrates GlobalMixer and LocalAmplifier modules, facilitating effective spatial semantic modeling with linear complexity. Together, these innovations empower SpikeDet to achieve the state-of-the-art performance across multiple object detection benchmarks with minimal energy consumption. On the widely used COCO dataset, SpikeDet achieves \textbf{68.3\% mAP@50} and \textbf{51.9\% mAP@50:95}, setting a new milestone in SNN-based detection and even surpassing several popular ANN models. Extensive ablation studies and evaluations across additional vision tasks further validate the effectiveness and generality of our approach.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 17300