DVS-OUTLAB A Neuromorphic Event-Based Long Time Monitoring Dataset for Real-World Outdoor Scenarios

Event: CVPR 2021 Workshop on Event-Based Vision · Duration: 3 min · ▶ Watch on YouTube

Abstract

The DVS-OUTLAB dataset is introduced as a neuromorphic event-based dataset designed for long-term monitoring in real-world outdoor scenarios. It leverages the technical advantages of Dynamic Vision Sensors (DVSs) such as low power consumption, low data redundancy, high temporal resolution, and high dynamic range, which are particularly useful in outdoor setups. Additionally, DVS technology offers social benefits by providing useful signals without direct grey/color values, addressing privacy concerns and legal issues associated with traditional CCTV systems. The dataset includes various environmental interferences and objects of interest, with a semi-automatic labeling pipeline for object classes and manual annotations for environmental factors. Experiments demonstrate the effectiveness of spatio-temporal filters, including deep learning and non-learning approaches, for denoising and event filtering within the dataset.

Speakers

  • Tobias Bolten — Hochschule Niederrhein University of Applied Sciences
  • Regina Pohle-Fröhlich — Hochschule Niederrhein University of Applied Sciences
  • Klaus D. Tönnies — Otto von Guericke Universität Magdeburg

Talks (1)

  • 00:00 — Tobias Bolten: DVS-OUTLAB A Neuromorphic Event-Based Long Time Monitoring Dataset for Real-World Outdoor Scenarios
    • Presentation of the DVS-OUTLAB dataset, an event-based dataset for long-term outdoor monitoring, highlighting its technical and social benefits, data generation process, and experimental results with spatio-temporal filters.

Key Takeaways

  • DVS-OUTLAB is a new event-based dataset for long-term outdoor monitoring, addressing real-world challenges.
  • Dynamic Vision Sensors (DVSs) offer significant technical and social advantages for outdoor surveillance, including privacy protection.
  • The dataset features diverse environmental interferences and objects of interest, labeled through a semi-automatic pipeline.
  • Simple non-learning spatio-temporal filters can achieve comparable denoising performance to deep learning approaches with lower computational cost for this use case.
  • The dataset and code will be publicly available to foster research in event-based vision for complex outdoor environments.

Methods / Models / Datasets Mentioned

  • DVS-OUTLAB
  • CeleX4
  • Mask-R-CNN
  • Block Super-Resolution
  • EDncCNN
  • Neighbor
  • SeqX
  • Time 3ms
  • Time 6.5ms
  • Time 10ms

Topics

Neuromorphic Event-Based Vision · Long-Time Monitoring · Outdoor Surveillance · Dataset Generation · Data Privacy · Spatio-Temporal Filtering · Deep Learning · Environmental Robustness · Object Detection · Event-Based Sensors

Notes

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