Exposure-Aware Training for Low-Light Object Detection Without Target-Domain Data.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Low-light object detection is difficult due to poor illumination, creating a gap between training and testing data.
• Current methods often require complex detector changes, image processing, or specific low-light datasets.

Purpose of the Study:

• To introduce Exposure-Aware Training (EAT), a simple yet effective method to enhance low-light object detection.
• To demonstrate that EAT improves performance without modifying existing detector architectures or increasing inference costs.

Main Methods:

• EAT applies simulated illumination attenuation and Gaussian noise to normal-light images during training.
• Degradation parameters are derived from real low-light image pairs.
• The object detection model architecture remains unchanged.

Main Results:

• Moderate simulated degradation consistently boosts low-light detection accuracy.
• Overly strong degradation can harm semantic information, particularly for small objects.
• EAT provides stable improvements for YOLOv8 and Faster R-CNN, especially for illumination-sensitive objects.

Conclusions:

• Task-specific illumination degradation during training enhances low-light detection.
• EAT is a lightweight, degradation-based strategy that improves performance without inference overhead.
• The method offers a practical solution for improving object detection in challenging lighting conditions.