New SNN Method Achieves 100% Recall for Visual Place Recognition

A team led by Katerina Maria Oikonomou and Antonios Gasteratos has published a new approach to visual place recognition (VPR) using spiking neural networks (SNNs) trained with unsupervised spike-timing-dependent plasticity (STDP). While previous STDP-based models offered high classification accuracy, they fell short on the Recall at 100% Precision (R@100P) metric critical for reliable autonomous navigation. The researchers present a discrete, tensor-native implementation of the SNN-VPR pipeline built entirely in PyTorch with the snnTorch library, enabling efficient training and inference on GPU hardware and paving the way for deployment on neuromorphic devices.

The study systematically evaluates three inference-stage design decisions: (1) replacing the standard argmax neuron assignment with a closed-form deterministic tensor pipeline, which significantly improved R@100P; (2) resetting neuron states after each query, which independently boosted recall; and (3) applying velocity-compensated sliding window aggregation over consecutive frames. On the 100-place Nordland dataset, these choices combined to yield a perfect R@100P = 100.00% at a window size of k=5 frames, with only 0.20 ms additional latency for constant-velocity traversal. The results highlight how careful low-level implementation decisions can bridge the gap between SNN-based VPR and the reliability required for real-world robotic loop closure.