Research map / Detection and segmentation / CNN features
Brain-Inspired Spiking Neural Networks for Energy-Efficient Object Detection
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 Multi-scale Spiking Detector (MSD), a directly-trained SNN object detector built from a spiking convolutional neuron block (ONNB) and a multi-scale spiking detection framework, reporting 62.0% mAP@50 on COCO 2017 and 66.3% on Gen1 with 7.8M parameters.
Problem
Object detection with spiking neural networks is mostly based on ANN-to-SNN conversion, which needs many time steps, is bounded by the ANN baseline, and handles event data poorly; directly trained deep SNN detectors remain underexplored.
Approach
A spiking convolutional neuron (LIF dynamics with tdBN and SiLU-based spiking) is used as the core of an ONNB residual block that applies spiking on both residual and shortcut paths to avoid non-spiking MAC operations, combined with a spiking multi-scale feature fusion and spiking decoupled detection head.
Key results
- MSD reaches 62.0% mAP@0.5 and 45.3% mAP@0.5:0.95 on COCO 2017 with 7.8M parameters and 6.4 mJ, versus 59.2% mAP@0.5 for SpikeYOLO (13.2M params).
- On Gen1 with T×D 5×1, MSD achieves 66.3% mAP@0.5 and 38.9% mAP@0.5:0.95, +0.3 mAP@0.5 over SpikeYOLO while using about 40.9% of its parameters.
- Ablation on Gen1 shows ONNB and MSDF together reduce energy by 82.9% relative to the baseline (6.5 mJ vs 38.2 mJ) while raising mAP@0.5 from 56.8% to 66.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ONNB to the baseline (ablation) | Baseline detector · Gen1 | mAP@0.5: 56.8 -> 64.3 | Helped |
| Add ONNB to the baseline (ablation) | Baseline detector · Gen1 | mAP@0.5:0.95: 37.4 -> 34.2 | Hurt |
| Add MSDF to the baseline (ablation) | Baseline detector · Gen1 | mAP@0.5: 56.8 -> 59.1 | Helped |
| Add MSDF to the baseline (ablation) | Baseline detector · Gen1 | mAP@0.5:0.95: 37.4 -> 41.1 | Helped |
| Add ONNB and MSDF (ablation) | Baseline detector · Gen1 | mAP@0.5: 56.8 -> 66.3 | Helped |
| Add ONNB and MSDF (ablation) | Baseline detector · Gen1 | mAP@0.5:0.95: 37.4 -> 38.9 | Helped |
Surprising findings
- Table 3 shows MSDF alone lowers mAP@0.5 relative to baseline (59.1% vs 56.8% is actually higher, but ONNB alone yields lower mAP@0.5:0.95 than baseline), indicating the components trade off across metrics.
Limitations
- No ablations over network depth or parameter scale, as acknowledged by the authors.
- Effect of membrane potential reset mechanism on learning dynamics and convergence is unclear.
- Energy numbers are estimated from operation counts (AC/MAC) rather than measured on neuromorphic hardware.
Experimental setup
Datasets: COCO 2017 Gen1
Benchmarks: COCO 2017 Gen1
Metrics: mAP@0.5 mAP@0.5:0.95 parameters (M) energy (mJ)
Baselines: Spiking-Yolo Bayesian Optim Spike Calib EMS-YOLO Meta-SpikeFormer SpikeYOLO Tr-Spiking-Yolo
Builds on: YOLO MSResNet tdBN STBP SpikeYOLO
Closest papers by meaning
- SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.91)
- SpikingYOLOX: Improved YOLOX Object Detection with Fast Fourier Convolution and Spiking Neural Networks (AAAI 2025 · similarity 0.90)
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.89)
- CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.89)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.89)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.88)
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.