Time-Ordered Recent Event (TORE) Volumes for Event Cameras.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Technology

Background:

• Event cameras offer high-speed, low-latency, and wide dynamic range imaging.
• Current machine learning architectures struggle with sparse event camera data.
• Existing event representations often interpolate data, losing timing information and reducing performance.

Purpose of the Study:

• To introduce a novel event representation called Time-Ordered Recent Event (TORE) volumes.
• To design a compact and efficient representation for raw event camera spike timing.
• To improve the performance of machine learning algorithms using event camera data.

Main Methods:

• Developed Time-Ordered Recent Event (TORE) volumes for compact storage of raw spike timing.
• TORE volumes are bio-inspired, memory-efficient, and computationally fast.
• The representation avoids time-blocking and incorporates local memory from past data.

Main Results:

• TORE volumes demonstrated significant improvements in state-of-the-art performance on challenging tasks.
• Evaluated on event denoising, image reconstruction, classification, and human pose estimation.
• The representation achieved superior results compared to existing event data handling methods.

Conclusions:

• TORE volumes offer an effective solution for processing sparse event camera data.
• This representation preserves crucial timing information and enhances network performance.
• TORE volumes are a versatile and easily implementable replacement for current event representations.