SpikeTAD: Spiking Neural Networks for End-to-End Temporal Action Detection
Researchers have developed SpikeTAD, a novel Spiking Neural Network (SNN) architecture for end-to-end temporal action detection in videos. This approach aims to address the high power consumption and large model sizes of traditional Artificial Neural Networks, making it suitable for deployment on mobile devices and neuromorphic chips. SpikeTAD demonstrates competitive performance, achieving 67.2% mAP on THUMOS14 and 37.42% mAP on ActivityNet-1.3, while maintaining significantly lower power usage. AI
IMPACT Enables more power-efficient video understanding models for edge devices.