When Event-Based Shape from Polarization Meets Spiking Neural Networks

Event-based shape from polarization offers a transformative approach to balancing speed and accuracy in capturing surface geometries. This paper explores event-based shape from polarization using Spiking Neural Networks (SNNs) for efficient surface normal estimation. We introduce Single-Timestep and Multi-Timestep Spiking UNets, where the former minimizes computation by updating neuron membrane potentials once, and the latter leverages temporal dynamics for better accuracy. Evaluations on the ESfP-Real dataset show our models match state-of-the-art Artificial Neural Networks (ANNs) while significantly reducing energy consumption.