Converting Static Image Datasets to Spiking Neuromorphic Datasets Using Saccades.

Area of Science:

• Computer Vision
• Neuromorphic Engineering
• Robotics

Background:

• Creating datasets for neuromorphic vision is challenging due to a scarcity of sensor recordings.
• Simultaneously capturing frame-based data for traditional computer vision comparison adds complexity.
• Existing methods for motion simulation can introduce timing artifacts.

Purpose of the Study:

• To propose and validate a method for converting static image datasets into neuromorphic vision datasets.
• To enable direct comparison between neuromorphic and traditional computer vision algorithms.
• To generate new datasets for the advancement of neuromorphic vision research.

Main Methods:

• Utilized an actuated pan-tilt camera platform to move the sensor, not the scene.
• Converted established computer vision datasets (MNIST, Caltech101) into neuromorphic formats.
• Evaluated performance using spike-based recognition algorithms on the generated datasets.

Main Results:

• Successfully converted MNIST and Caltech101 datasets into neuromorphic vision datasets.
• Provided performance metrics for spike-based algorithms on these converted datasets.
• Demonstrated a biologically realistic approach to data acquisition for neuromorphic sensors.

Conclusions:

• The proposed method offers a viable solution for generating neuromorphic vision datasets from existing static image collections.
• This approach facilitates more direct comparisons between neuromorphic and frame-based computer vision techniques.
• The generated datasets and performance metrics will support future research in neuromorphic vision.