CIFAR10-DVS: An Event-Stream Dataset for Object Classification.

Area of Science:

• Computer Vision
• Neuromorphic Engineering
• Machine Learning

Background:

• High-quality event-stream datasets are crucial for advancing neuromorphic vision research.
• Existing event-stream datasets are limited, hindering algorithm development.
• Creating such datasets typically requires time-consuming recording with neuromorphic cameras.

Purpose of the Study:

• To address the scarcity of event-stream datasets.
• To introduce a novel dataset, CIFAR10-DVS, derived from a widely used image dataset.
• To provide a benchmark for event-driven object classification.

Main Methods:

• Converted 10,000 frame-based images from the CIFAR-10 dataset into event streams.
• Utilized a dynamic vision sensor (DVS) for event generation.
• Employed a repeated closed-loop smooth (RCLS) image movement technique for realistic event data generation.

Main Results:

• Generated CIFAR10-DVS, a dataset of 10,000 event streams across 10 classes.
• The RCLS method simulates realistic intensity changes captured by DVS.
• Established a performance benchmark for event-driven object classification using state-of-the-art algorithms.

Conclusions:

• The CIFAR10-DVS dataset offers a valuable resource for neuromorphic vision research.
• This work provides a foundation for developing and comparing event-driven pattern recognition algorithms.
• The dataset and benchmark are expected to accelerate progress in the field.