Unsupervised Face Detection in the Dark.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Low-light face detection is crucial for applications like autonomous driving and surveillance.
• Existing models require extensive annotations and lack adaptability.
• Current methods struggle with the significant domain gap between normal and low-light conditions.

Purpose of the Study:

• To develop a method for learning face detectors without low-light annotations.
• To adapt existing face detectors trained on normal-light data to perform effectively in low-light environments.
• To address the challenges posed by the large brightness and complexity gap between normal and low-light images.

Main Methods:

• Proposed a joint High-Low Adaptation (HLA) framework.
• Implemented bidirectional low-level adaptation (enhancing dark images, degrading normal images).
• Utilized multitask high-level adaptation combining context-based and contrastive learning.

Main Results:

• The HLA-Face v2 model achieved superior low-light face detection performance.
• The model demonstrated effectiveness without relying on low-light specific annotations.
• The adaptation scheme proved versatile, applicable to other computer vision tasks.

Conclusions:

• The HLA framework successfully bridges the domain gap for low-light face detection.
• This approach offers a flexible and efficient solution for learning face detectors in challenging lighting.
• The method's adaptability extends its utility to broader computer vision applications, enhancing supervised learning and object detection.