Reliable Event Generation With Invertible Conditional Normalizing Flow.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Event streams capture visual scene dynamics via intensity variations and noise.
• Current event generation methods use limited one-way mappings, hindering adaptability to diverse scenarios and event cameras.

Purpose of the Study:

• To develop a novel approach for learning bidirectional mappings between event stream features and their parameters.
• To enhance the generalization capabilities of event stream generation for diverse applications.

Main Methods:

• Synthesized massive event streams using an event simulator with randomly generated parameters.
• Proposed an event-based normalizing flow network to learn invertible mappings between event stream representations and parameters.
• Incorporated an intensity-guided conditional affine simulation mechanism and novel event losses for event priors.

Main Results:

• The proposed framework achieved superior performance in video reconstruction, optical flow estimation, and parameter estimation.
• Demonstrated excellent generalization across various synthetic and real-world datasets, scenes, and cameras.
• Outperformed state-of-the-art methods in key computer vision tasks.

Conclusions:

• The bidirectional mapping approach significantly improves event stream generation reliability and generalization.
• The method effectively integrates event priors through learned mappings and specialized loss functions.
• This work advances event-based vision by providing a more robust and adaptable event stream generation framework.