VPRTempo introduces a fast Spiking Neural Network (SNN) for Visual Place Recognition (VPR) that is trainable within minutes and queryable in milliseconds. The system employs a temporal code based on pixel intensity to determine spike timing, improving efficiency by over 100%. VPRTempo is trained using Spike-Timing Dependent Plasticity and a supervised delta learning rule. The system's accuracy is comparable to prior SNNs and NetVLAD while being significantly faster, suitable for real-time deployment. The network architecture consists of three layers: input layer, feature layer, and one-hot encoded output layer. The key contributions include novel temporal encoding, reduced training times, and high query speeds on both CPUs and GPUs.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Adam D. Hine... a las arxiv.org 03-04-2024
https://arxiv.org/pdf/2309.10225.pdfConsultas más profundas