LEARNING FROM EVENTS: ON THE FUTURE OF MACHINE LEARNING FOR EVENT-BASED CAMERAS

Event: CVPR 2019 Second International Workshop on Event-Based Vision and Smart Cameras · Duration: 14 min · ▶ Watch on YouTube

Abstract

Amos Sironi from Prophesee discusses the evolution of event-based cameras, highlighting the constant reduction in pixel size and increase in resolution. He presents current applications of event-based vision in optical flow estimation, feature point learning, and object detection, emphasizing the efficiency and low latency of these methods. Sironi then outlines the future of event-based AI, drawing parallels with the success of frame-based AI, and stressing the need for dedicated algorithms that exploit sparsity and temporal information, specialized neuromorphic hardware with integrated memory, and larger, more diverse datasets.

Speakers

  • Amos Sironi — Prophesee

Talks (1)

  • 00:00 — Amos Sironi: LEARNING FROM EVENTS: ON THE FUTURE OF MACHINE LEARNING FOR EVENT-BASED CAMERAS
    • A discussion on the evolution of event-based sensors, current machine learning applications, and future directions focusing on algorithms, hardware, and datasets.

Key Takeaways

  • Event-based sensor technology is rapidly advancing, with increasing resolution and integrated event signal processing.
  • Current machine learning applications leveraging event data demonstrate high performance and efficiency for tasks like optical flow and object detection.
  • Future event-based AI development should prioritize algorithms that exploit the unique sparsity and temporal advantages of event data.
  • Dedicated neuromorphic hardware with closely integrated memory and computation is crucial for unlocking the full potential of event-based vision.
  • The creation and release of larger, high-quality event-based datasets are essential to drive further innovation and benchmark progress in the field.

Methods / Models / Datasets Mentioned

  • EV-FLOWNET
  • MobileNet-V2
  • HOTS
  • HATS
  • SNN
  • EST
  • SCAMP
  • Spinnaker
  • TrueNorth
  • Dynap
  • Loihi
  • N-MNIST
  • N-Caltech
  • DVS Gestures
  • N-cars
  • DDD 2017
  • MVSEC Dataset
  • PROPHESEE N-CARS DATASETS
  • PROPHESEE HVGA CORNER DATASET
  • PROPHESEE DETECTION DATASET
  • ImageNet
  • AlexNet

Topics

event-based cameras · machine learning · optical flow · feature points · object detection · neuromorphic hardware · datasets · sparsity · temporal information · sensor evolution

Notes

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