SpikeDet: Better Firing Patterns for Accurate and Energy-Efficient Object Detection with Spiking Neural Networks

TL;DR

SpikeDet tackles local firing saturation in spiking neural networks used for object detection by stabilizing firing patterns in the backbone with MDSNet and preserving them through multi-directional fusion in the neck with SMFM. The approach introduces the Local Firing Saturation Index (LFSI) to quantify saturation and demonstrates substantial gains in COCO AP (52.2) with roughly half the power, plus strong performance on event-based and challenging sub-tasks. Theoretical analyses show gradient stability via Block Dynamical Isometry, and extensive ablations confirm the benefits of deeper MDSNet, more fusion directions, and SF-Block design. Overall, SpikeDet narrows the accuracy–energy gap for SNN-based detectors and proves robust across diverse sensing modalities and difficult scenes.

Abstract

Spiking Neural Networks (SNNs) are the third generation of neural networks. They have gained widespread attention in object detection due to their low power consumption and biological interpretability. However, existing SNN-based object detection methods suffer from local firing saturation, where adjacent neurons concurrently reach maximum firing rates, especially in object-centric regions. This abnormal neuron firing pattern reduces the feature discrimination capability and detection accuracy, while also increasing the firing rates that prevent SNNs from achieving their potential energy efficiency. To address this problem, we propose SpikeDet, a novel spiking object detector that optimizes firing patterns for accurate and energy-efficient detection. Specifically, we design a spiking backbone network, MDSNet, which effectively adjusts the membrane synaptic input distribution at each layer, achieving better neuron firing patterns during spiking feature extraction. For the neck, to better utilize and preserve these high-quality backbone features, we introduce the Spiking Multi-direction Fusion Module (SMFM), which realizes multi-direction fusion of spiking features, enhancing the multi-scale detection capability of the model. Furthermore, we propose the Local Firing Saturation Index (LFSI) to quantitatively measure local firing saturation. Experimental results validate the effectiveness of our method, with SpikeDet achieving superior performance. On the COCO 2017 dataset, it achieves 52.2% AP, outperforming previous SNN-based methods by 3.3% AP while requiring only half the power consumption. On object detection sub-tasks, including event-based GEN1, underwater URPC 2019, low-light ExDARK, and dense scene CrowdHuman datasets, SpikeDet also achieves the best performance.

Figures (14)

• Figure 1: Visualization of local firing saturation problem in SNN-based object detector on COCO dataset. Each pixel represents the neuron firing rate. (a) and (b) show feature maps before the detection head, as these features determine both classification and regression. (a) averages the $4D$ spike tensor $([T, C, H, W])$ across both time and channel dimensions to show overall spatial firing distribution, while (b) selects a representative channel and averages across time steps to reveal individual neuron firing patterns. (c) shows neuron firing patterns of backbone features at $1/16$ resolution. (d) presents the final detection results.
• Figure 2: Comparisons with other state-of-the-art methods on COCO AP and Power consumption. Squares represent ANN-based object detectors, circles represent SNN-based object detectors, and triangles represent our methods. (a) Comparison results on the COCO 2017 dataset. (b) Comparison results on the GEN1 dataset.
• Figure 3: The architecture of SpikeDet. SpikeDet comprises MDSNet, SMFM, and the SpikeYOLO Detection Head luo_integervalued_2025. The model receives two types of inputs: event and static data, with the input coding and output represented in the figure. The core of SpikeDet is MDSNet, which consists of 5 stages. Each stage downsampling factor relative to the original input is marked in the figure, with MDS-Block1 and MDS-Block2 performing feature extraction without and with downsampling, respectively. By incorporating MDS, this architecture successfully stabilizes the firing patterns at each stage, alleviating the local firing saturation problem. Additionally, SMFM enables multi-direction feature fusion, allowing features to undergo multiple refinements through the model, thereby improving the model's capability to detect multi-scale objects and preserving neuron firing patterns.
• Figure 4: Firing rate distribution of I-LIF neurons for different presynaptic input $\mathbf{x}^{t,n}$ when $T=1$ and $D=4$. The input $\mathbf{x}$ is sampled from Gaussian distributions with different variances. Increased variance leads to higher firing saturation probability.
• Figure 5: Influence of multi-direction feature fusion on firing patterns of SNN-based detector. We visualize feature maps at the 1/16 downsampling stage, averaging across $T$ and $C$ dimensions to reveal overall firing patterns. Figures (a) to (e) show neuron firing patterns for feature maps with no fusion, one-way, two-way, three-way, and four-way fusion, respectively.