Learning local depth regression from defocus blur by soft-assignment encoding.

Area of Science:

• Computer Vision
• Computational Imaging
• Machine Learning

Background:

• Depth from Defocus (DFD) is crucial for 3D scene understanding.
• Current DFD methods often rely on patch classification, which struggles with continuous depth variations.
• This leads to inaccuracies in depth estimation.

Purpose of the Study:

• To develop a novel patch-based approach for accurate depth regression from defocus blur.
• To overcome the limitations of classification-based DFD methods.
• To enable precise depth estimation without requiring blur or scene models.

Main Methods:

• A patch-based depth regression technique is proposed.
• A classification model is adapted using soft-assignment encoding for training.
• A regression scale predicts intermediate depth values, handling continuous depth variations.

Main Results:

• The proposed method outperforms traditional classification and direct regression approaches.
• Demonstrated superior performance on simulated datasets with varying textures.
• Validated effectiveness on real-world data with optical aberrations.

Conclusions:

• The novel patch-based depth regression method offers improved accuracy for DFD.
• Soft-assignment encoding effectively handles continuous depth variations.
• This approach provides a robust and model-free solution for depth estimation.