Research map / Detection and segmentation / CNN features

SpikingYOLOX: Improved YOLOX Object Detection with Fast Fourier Convolution and Spiking Neural Networks

AAAI 2025 · Detection and segmentation › CNN features · approach: Spiking event-camera detection · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds SpikingYOLOX, a directly trained spiking version of YOLOX that adds ternary signed spiking neurons and fast Fourier convolution (FFC) blocks in the deepest backbone stage, evaluated on COCO2017.

Problem

Spiking neural networks lag behind ANNs on object detection on complex datasets such as COCO, partly because SNN backbones lack a large effective receptive field and spikes are binary, limiting feature richness.

Approach

Introduces a ternary signed spiking neuron (outputs -1, 0, +1 with a secondary negative threshold gated by prior positive spikes) and places FFC residual blocks and spiking layers (CSP-FFC-SNN, SPP-SNN) in the dark5 stage of a YOLOX backbone to obtain global receptive field and richer spike features.

Key results

• SpikingYOLOX-L reaches 42.0 AP(0.5:0.95) on COCO2017val and SpikingYOLOX-S reaches 57.2 AP(0.5) per Table 1/3, outperforming SNN baselines such as Spike Calibration (45.4 AP50 at T=512) and EMS-YOLO-ResNet34 (50.1 AP50 at T=4).
• Among small YOLO models, SpikingYOLOX-tiny (5.06M params) reports 54.3 AP versus 52.3 for YOLOX-tiny, though the paper's tables are mixed in their metric definitions.
• Placing signed spiking neurons and FFC only in the dark5 (deepest) stage gave the best ablation results (57.21 mAP vs 54.50 for Dark4 placement).

Measured effects reported in the paper

Surprising findings

• The nano-sized SpikingYOLOX did not converge during training.
• The X-size model underperforms the L-size model (41.0 vs 42.0 AP), attributed to slower convergence at large scale.
• Adding FFC to both Dark4 and Dark5 degraded performance (54.54 mAP vs 57.67 with Dark5 only).

Limitations

• Nano-scale model fails to converge and very large models (X) underperform L, indicating scaling difficulties.
• Comparisons to other SNN detectors are limited to few methods, and the paper does not report energy/spike-rate measurements.
• Reported tables use inconsistent metrics and parameter counts, making comparisons across rows hard to verify.

Experimental setup

Datasets: MS-COCO 2017

Benchmarks: COCO2017val

Metrics: AP 0.5:0.95 AP 0.5 AP 0.75 AP small AP medium AP large params (M)

Baselines: Spiking-YOLO Bayesian Optimization Spike Calibration EMS-YOLO-ResNet34 YOLOX-tiny YOLOX-S YOLOv8-S YOLOv9-S

Builds on: YOLOX fastSNN Fast Fourier Convolution Spiking-YOLO EMS-YOLO

Closest papers by meaning

• SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.91)
• Brain-Inspired Spiking Neural Networks for Energy-Efficient Object Detection (CVPR 2025 · similarity 0.90)
• CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.89)
• Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.87)
• Spike2Former: Efficient Spiking Transformer for High-performance Image Segmentation (AAAI 2025 · similarity 0.86)
• Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · similarity 0.86)