Aberration-Aware Depth-From-Focus.

Area of Science:

• Computer Vision
• Computational Imaging
• Machine Learning

Background:

• Traditional depth estimation methods rely on simplified camera models, neglecting optical imperfections.
• Simulated data for training deep learning models lacks realism due to unmodeled lens aberrations, impacting focus-sensitive tasks like Depth-from-Focus.
• Off-axis aberrations in real lenses introduce a domain gap, affecting the accurate identification of the best-focused frame in image sequences.

Purpose of the Study:

• To investigate the impact of lens aberrations on depth estimation accuracy in computer vision.
• To develop and evaluate a novel training strategy, aberration-aware training (AAT), to bridge the domain gap between simulated and real-world data.
• To enhance the robustness and generalizability of depth estimation models for focus-sensitive applications.

Main Methods:

• Developed a lightweight network to model lens aberrations across various positions and focus distances.
• Integrated the aberration-aware network into the standard deep network training pipeline.
• Evaluated model performance on both synthetic and real-world datasets.

Main Results:

• The proposed aberration-aware training (AAT) scheme significantly improves depth estimation accuracy.
• Models trained with AAT demonstrate improved generality across different datasets without requiring dataset-specific fine-tuning.
• The approach effectively addresses the domain gap caused by lens aberrations.

Conclusions:

• Aberration-aware training is a viable method for enhancing depth estimation in computer vision.
• The developed AAT approach offers a robust solution for real-world depth estimation challenges.
• The findings pave the way for more accurate and adaptable depth estimation systems.