A Cortically-inspired Architecture for Event-based Visual Motion Processing: From Design Principle to Real-world Applications

Event: CVPR 2025 Workshop · Duration: 3 min · ▶ Watch on YouTube

Abstract

This work presents a cortically-inspired, multilayer spiking neural network (SNN) architecture designed for event-based visual motion processing. The model aims to functionally mimic the cortical motion pathway by employing a compositional approach to build complex visual descriptors. It processes input from event-based sensors (DVS events) through Gabor-like receptive fields and spatio-temporal oriented filters, which act as motion energy units. The architecture culminates in a decoding stage that estimates local velocities and optical flow, striving for biological plausibility by adapting known rate-based mechanisms to a spike-based network.

Speakers

  • Francesca Peveri — Università degli Studi di Genova
  • Simone Testa — Università degli Studi di Genova
  • Silvio P. Sabatini — Università degli Studi di Genova

Talks (1)

  • 00:00:00 — Francesca Peveri: A Cortically-inspired Architecture for Event-based Visual Motion Processing: From Design Principle to Real-world Applications
    • Presentation of a bio-inspired, multilayer spiking neural network (SNN) architecture for event-based visual motion processing, mimicking cortical pathways to estimate optical flow.

Key Takeaways

  • The presented architecture is a multilayer Spiking Neural Network (SNN) designed for event-based visual motion processing.
  • It functionally mimics the cortical motion pathway, specifically drawing inspiration from V1 and MT cortex processing.
  • The architecture utilizes Gabor-like receptive fields and spatio-temporal oriented filters to extract early visual features and motion energy from DVS events.
  • Motion energy units are combined to estimate the magnitude and direction of local velocities, which are then decoded into optical flow.
  • The design emphasizes biological plausibility by translating principled firing-rate computational models into event-based SNNs.

Methods / Models / Datasets Mentioned

  • Eger and Simoncelli model
  • Gabor-like receptive fields
  • DVS events accumulator
  • Spatio-temporal oriented filters
  • Motion energy units
  • V1 (visual cortex area)
  • MT cortex (visual cortex area)

Topics

Spiking Neural Networks (SNN) · Bio-inspired architecture · Event-based vision · Motion estimation · Cortical motion pathway · Optical flow · Visual processing · Neuromorphic computing · Feature extraction

Notes

Open for commentary — connections to other work, critiques, follow-up reading.