A Benchmark Environment for Neuromorphic Stereo Vision.

Area of Science:

• Computer Vision
• Robotics
• Neuromorphic Engineering

Background:

• Conventional artificial stereo vision faces high resource demands and processing times, hindering real-time performance due to redundant data from high frame rates.
• Neuromorphic visual sensors offer reduced data redundancy and improved sampling, but existing stereo vision techniques do not leverage their event-based nature.
• A lack of standardized evaluation hinders the comparison of novel event-based stereo vision algorithms.

Purpose of the Study:

• To propose a benchmark environment for evaluating and comparing different depth reconstruction algorithms for event-based stereo vision.
• To provide a framework for synchronized and calibrated data recording from event-based and depth sensors.
• To define metrics for assessing algorithm performance, precision, and applicability in event-based stereo vision.

Main Methods:

• Development of a benchmark environment with tools for comparing event-based stereo vision algorithms.
• Establishment of an experimental setup with synchronized, calibrated event-based and depth sensors.
• Implementation of a stereo matching algorithm as a test case and execution of experiments with varying parameters.

Main Results:

• A functional benchmark environment and experimental setup for event-based stereo vision data recording.
• Defined metrics for comprehensive algorithm evaluation, including performance, precision, and applicability.
• Demonstrated the benchmark's utility through experiments with a stereo matching algorithm.

Conclusions:

• The proposed benchmark provides a foundation for robust and flexible evaluation of emerging event-based stereo vision techniques.
• This work facilitates meaningful comparisons between different depth reconstruction algorithms, advancing the field.
• The benchmark addresses the need for a common data basis and standardized metrics in neuromorphic stereo vision research.