NER-Net+: Seeing Motion at Nighttime With an Event Camera.

Area of Science:

• Computer Vision
• Computational Imaging
• Sensor Technology

Background:

• Conventional cameras struggle with low-light dynamic scene imaging due to exposure trade-offs.
• Event cameras offer high temporal resolution and dynamic range, ideal for dynamic scenes.
• Lack of real nighttime paired event-image data and tailored reconstruction methods limits current event camera applications.

Purpose of the Study:

• To address the challenges of nighttime dynamic scene imaging using event cameras.
• To create the first real paired low-light event dataset (RLED) for training and evaluation.
• To develop an advanced event reconstruction network specifically for nighttime conditions.

Main Methods:

• Constructed the RLED dataset with 80,400 aligned low-light event frames and ground truth images.
• Analyzed characteristics of nighttime events: noise, temporal trailing, and spatial non-uniformity.
• Proposed the Nighttime Event Reconstruction Network (NER-Net+) with LETC and NSIE modules.

Main Results:

• The RLED dataset provides a unified benchmark for nighttime event reconstruction.
• NER-Net+ effectively corrects temporal trailing and suppresses noise and non-uniformity.
• Experimental results show superior performance over state-of-the-art methods in visual quality and generalization.

Conclusions:

• The developed dataset and network significantly advance the field of nighttime dynamic scene imaging with event cameras.
• NER-Net+ demonstrates robust performance on real-world nighttime data, overcoming limitations of existing methods.