The challenge on
Highspeed HDR Video Reconstruction From Event

Highspeed HDR Video Reconstruction from Events

Fig.1. Representative examples for Event-to-HDR dataset.. For each scene, a event stream captured by real event cameras are used as input data, and the HDR video frame that is merged by bracket exposures are used as ground truth.

Highspeed HDR Video Reconstruction from Events

Event cameras, differing from conventional cameras that capture scene intensities at a fixed frame rate, use a unique approach by detecting pixel-wise intensity changes asynchronously. This is triggered whenever a pixel's intensity change surpasses a certain contrast threshold. Unlike traditional frame-based cameras, event cameras have several advantages: low latency, low power consumption, high temporal resolution, and high dynamic range (HDR). These qualities make them particularly useful for a range of vision tasks, including real-time object tracking, high-speed motion estimation, face recognition, optical flow estimation, depth map prediction, ego motion analysis, and onboard robotics applications.

However, the distinct triggering mechanism of event cameras presents a challenge. The event data they capture, which lacks absolute intensity values and is represented as 4-tuples, is incompatible with standard frame-based vision algorithms. This discrepancy necessitates specialized processing pipelines, different from traditional image processing methods. Consequently, there is a growing interest in transforming event data into intensity images to leverage the high-speed and HDR capabilities of event cameras in practical applications.

To this end, we are launching a challenge focused on reconstructing high-speed HDR videos from event streams. We will utilize the high-quality Event-to-HDR dataset, captured by a co-axis system and developed by Prof. Fu's team as noted in [a]. This dataset includes aligned pairs of event streams and HDR videos in both spatial and temporal dimensions. The four-tuple event streams will serve as the input, while the ground truth will be HDR images merged from two high-speed cameras. The dataset features a high frame rate of 500fps. We will host the competition using open source online platform, e.g. CodaLab. All submissions are evaluated by our script running on the server and we will double check the results of top-rank methods manually before releasing the final test-set rating.

[a] Zou, Yunhao, et al. "Learning to reconstruct high speed and high dynamic range videos from events." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021