A High Resolution Multi-exposure Stereo Event-Intensity-Depth Database of Natural Scenes

The emerging field of event-based vision delivers outstanding visual output with unique properties such as asynchronous data with a high temporal resolution, negligible motion blur, and polarity as a measure of change in brightness. These properties make event data a better alternative than frame-based data for applications such as optical flow estimation, visual odometry, etc. On the contrary, the application of event data is constrained by the lack of scene-specific pixel-intensity information and the absence of event generation for static regions. The absence of suitable experimental data for both event-centered and frame-oriented processes impedes primary research development that combines the strengths of these approaches. We propose a multi-exposure stereo event-intensity-depth dataset for natural scenes with a broad depth field, intricate depth structure, complex object movement, rich color dynamics, illumination variations, texture discrepancy, and substantial randomness introduced by moving camera and background motion. We have discussed the progress, challenges, potential use cases and future research opportunities for our proposed dataset concerning its application in object classification, depth estimation, HDR video reconstruction, and visual odometry.