Research map / Detection and segmentation / CNN features
SFOD: Spiking Fusion Object Detector
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.
Proposes SFOD, an SNN-based event-camera object detector with a Spiking Fusion Module that concatenates multi-scale features (via transposed-conv upsampling) and refines them with SEW-residual pyramid blocks, plus a study showing spiking rate decoding with MSE loss works best.
Problem
Object detection with event cameras using spiking neural networks is limited, especially because multi-scale feature fusion has not been explored in SNNs and spiking output decoding and loss choices are not studied for event data.
Approach
A Spiking Fusion Module that fuses multi-scale spiking feature maps by concatenation (not element-wise sum, to keep binary activations) with transposed-convolution upsampling and a SEW-residual Spiking Pyramid Extraction Submodule, combined with Spiking DenseNet backbone and SSD head; plus analysis showing spiking rate decoding with MSE beats count decoding and cross-entropy.
Key results
- SFOD reaches 93.7% accuracy on NCAR with DenseNet121-16 and spiking rate decoding with MSE loss.
- SFOD achieves 32.1% mAP@0.5:0.95 on GEN1 with 11.9M parameters and 24.04% firing rate, versus 18.9% for DenseNet121-24+SSD (reported as 0.189 in the paper's table).
- Spiking rate decoding outperforms spiking count decoding by about 7 accuracy points under MSE on NCAR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Spiking Rate Decoding instead of Spiking Count Decoding (ablation) | DenseNet121-16 classifier · NCAR | accuracy: 0.869 -> 0.937 | Helped |
| Use MSE instead of CE loss with Spiking Rate Decoding (ablation) | DenseNet121-16 classifier · NCAR | accuracy: 0.930 -> 0.937 | Helped |
| Use Spiking Rate Decoding instead of Spiking Count Decoding (ablation) | SNN object detector · GEN1 | mAP@0.5:0.95: 0.235 -> 0.288 | Helped |
| Increase DenseNet121 growth rate from 16 to 24 (ablation) | DenseNet121-16 classifier · NCAR | accuracy: 0.937 -> 0.928 | Hurt |
| Increase DenseNet depth from 121 to 169 at growth rate 16 (ablation) | DenseNet121-16 classifier · NCAR | accuracy: 0.937 -> 0.921 | Hurt |
| Use DenseNet121-24 rather than DenseNet121-16 as the detection backbone (ablation) | SSD detector · GEN1 | mAP@0.5:0.95: 0.262 -> 0.288 | Helped |
| Fuse three feature-map layers rather than four (ablation) | SFOD-B · GEN1 | mAP@0.5:0.95: 0.294 -> 0.299 | Helped |
| Enhance SPES with a SEW Res Block rather than the basic SPES (ablation) | SFOD · GEN1 | mAP@0.5:0.95: 0.299 -> 0.321 | Helped |
| Enhance SPES with a Spiking Dense Block rather than the basic SPES (ablation) | SFOD · GEN1 | mAP@0.5:0.95: 0.299 -> 0.286 | Hurt |
| Use MSE instead of CE loss with Spiking Count Decoding (ablation) | DenseNet121-16 classifier · NCAR | accuracy: 0.920 -> 0.869 | Hurt |
| Use DenseNet121-24-SSD rather than the SFOD fusion module (ablation) | SFOD detector · GEN1 | mAP@0.5:0.95: 0.288 -> 0.321 | Hurt |
Surprising findings
- Cross-entropy loss gave better results than MSE when spiking count decoding was used, reversing the MSE advantage seen with rate decoding.
- Fusing the fourth feature map (4x5) reduced firing rate but did not improve mAP, so it was dropped.
Limitations
- Detection performance remains dependent on backbone pretraining and fusion layer choice; fusing 4th layer hurt mAP.
- Energy and runtime for some baselines are estimated locally rather than measured directly.
- Authors note data augmentation as a direction for improvement, implying current augmentation is limited (only horizontal flipping).
Experimental setup
Datasets: NCAR GEN1
Benchmarks: NCAR GEN1
Metrics: accuracy mAP@0.5:0.95 mAP@0.5 firing rate parameters energy runtime
Baselines: DenseNet121-24+SSD MobileNet-64+SSD VGG-11+SSD EMS-YOLO RED RVT Asynet HATS
Builds on: Spiking DenseNet SEW ResNet SSD PLIF
Closest papers by meaning
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.93)
- SpikingYOLOX: Improved YOLOX Object Detection with Fast Fourier Convolution and Spiking Neural Networks (AAAI 2025 · similarity 0.91)
- CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.91)
- Brain-Inspired Spiking Neural Networks for Energy-Efficient Object Detection (CVPR 2025 · similarity 0.91)
- Hybrid Spiking Vision Transformer for Object Detection with Event Cameras (ICML 2025 · similarity 0.90)
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.