Noise2Image for event cameras

The Noise2Image Dataset is a groundbreaking collection of event camera recordings, designed to enable research into noise-based recovery of static scenes. This dataset demonstrates how noise events triggered by event cameras correlate with the absolute intensity of a scene, allowing the visualization of static images through noise analysis. Perfect for researchers and developers in computer vision, dynamic vision sensors, and event-based camera applications.

Key Features

• Event Camera Modality: Recorded using the Prophesee Metavision EVK3-HD, capturing pure noise events with default bias parameters and no additional denoising.
• Static Scene Recovery: Enables exploration of how event cameras ‘see’ static scenes by analyzing noise events correlated with scene intensity.
• High-Quality Collection: Includes noise recordings paired with corresponding ground truth intensity images displayed on a calibrated LCD monitor.

Dataset Structure

The dataset is divided into two main parts:

1. In-Distribution Set (indist): Human portrait images sourced from Unsplash for training and validation.
2. Out-of-Distribution Set (ood_DIV2K): 100 natural images from the DIV2K super-resolution dataset for final evaluation.

Each part contains:

• Event Recordings (*_events): 10-second-long noise recordings in raw format.
• Ground Truth Intensity Images (*_images): Static images displayed during recording.