Dynamic Vision Sensor - The Road to Market

Event: ICRA’17 Workshop on “Event-based Vision” · Duration: 20 min · ▶ Watch on YouTube

Abstract

This presentation from Samsung outlines their roadmap for bringing Dynamic Vision Sensors (DVS) to mass production by 2017-2018. It addresses key market-driven requirements such as reducing cost and module size through pixel shrink and stacked technology, optimizing power consumption via advanced processes and dynamic wake-up modes, and enhancing event quality by suppressing redundant and handling missing events. The talk also covers strategies for data throughput reduction through efficient event encoding and discusses the need for specialized hardware acceleration for event processing.

Speakers

  • Yoel Yaffe — Samsung Israel Research Center, Israel; System LSI, South Korea; Samsung Electronics

Talks (1)

  • 00:00:00 — Yoel Yaffe: Dynamic Vision Sensor - The Road to Market
    • A presentation by Samsung on their Dynamic Vision Sensor (DVS) product roadmap, focusing on addressing market-driven requirements for mass production, including cost, size, power, event quality, data throughput, and processing acceleration.

Key Takeaways

  • Samsung is actively developing Dynamic Vision Sensors (DVS) for mass production, with a roadmap extending to 2018, focusing on consumer product integration.
  • Key technical advancements include significant pixel size reduction (from 18.5µm to 6µm) and optical format optimization using stacked Cu-Cu technology.
  • Power consumption is minimized through advanced manufacturing processes, subsampling modes, and dynamic wake-up mechanisms based on spatial histogram data.
  • Event quality is improved by on-chip hardware blocks for suppressing redundant events like noise, bad pixels, multiple detections from strong contrast changes, and light source flicker.
  • Data throughput is reduced by employing efficient event encoding schemes like Group Address Event Representation (GAER), achieving an average of ~8 bits per event, and specialized hardware accelerators are being explored for efficient event processing.

Methods / Models / Datasets Mentioned

  • BSI process
  • MIMCAP
  • Stacked Cu-Cu Technology
  • Spatial Histogram
  • Blocking Area
  • AER (Address Event Representation)
  • MIPI I/F
  • USB I/F
  • FIFO
  • Parallel I/F
  • MUX
  • PAD Out
  • SNE
  • LogAMP of DVS Gen2 (ISSCC'17)
  • Flicker Suppression Algorithm
  • GAER (Group Address Event Representation)
  • Histogram of Gradients (HOG)

Topics

Dynamic Vision Sensor (DVS) · Event-based vision · Product roadmap · Pixel technology · Power reduction · Event quality · Data throughput · Hardware acceleration · Market requirements · Sensor design

Notes

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