Asynchronous Convolutional Networks for Object Detection in Neuromorphic Cameras

Event: CVPR 2020 · Duration: 3 min · ▶ Watch on YouTube

Abstract

This presentation introduces YOLE, an event-based object detection network adapted from YOLO, which utilizes a Leaky Surface Layer to reconstruct frames from asynchronous events. To enhance computational efficiency by leveraging the sparsity of event data, the authors developed fYOLE, a fully convolutional version incorporating novel asynchronous event-based layers. These stateful layers perform incremental computation by maintaining internal memory and selectively updating only the parts of feature maps affected by new events, leading to significant performance gains on neuromorphic camera datasets.

Speakers

  • Marco Cannici — Politecnico di Milano
  • Marco Ciccone — Politecnico di Milano
  • Andrea Romanoni — Politecnico di Milano
  • Matteo Matteucci — Politecnico di Milano

Talks (1)

  • 00:00:00 — Marco Cannici: Asynchronous Convolutional Networks for Object Detection in Neuromorphic Cameras
    • This talk introduces YOLE, an event-based object detection network, and fYOLE, an asynchronous version with stateful layers for incremental computation, demonstrating improved efficiency and performance on neuromorphic camera datasets.

Key Takeaways

  • YOLE adapts the YOLO architecture for object detection in event-based cameras by using a Leaky Surface Layer for frame reconstruction.
  • fYOLE introduces novel asynchronous, stateful convolutional and max-pooling layers to enable incremental computation.
  • These asynchronous layers maintain internal memory and only update regions affected by new events, significantly improving computational efficiency.
  • The fYOLE architecture demonstrates competitive object detection performance across various event-based datasets compared to its synchronous counterpart.

Methods / Models / Datasets Mentioned

  • YOLE
  • YOLO
  • Leaky Surface Layer
  • e-conv layer
  • e-max-pool layer
  • S-N-MNIST
  • Blackboard MNIST
  • OD-Poker-DVS
  • N-Caltech101

Topics

Event-based vision · Object detection · Neuromorphic cameras · Asynchronous neural networks · Incremental computation · Convolutional networks · Real-time processing · Sparse data processing

Notes

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