SAS: A General Framework Induced by Sequence Association for Shape From Focus.

Area of Science:

• Computer Vision
• Image Processing
• 3D Reconstruction

Background:

• Shape from focus (SFF) estimates scene depth using multifocus images.
• Traditional SFF methods use focus measure operators but ignore image sequence associations.
• Deep learning SFF methods often require difficult-to-obtain labeled datasets.

Purpose of the Study:

• To propose a novel sequence association-based (SAS) framework to enhance the generalizability of SFF methods.
• To address limitations of traditional and deep learning SFF approaches, particularly the need for labeled data.

Main Methods:

• The SAS framework treats image sequences as 3D data for multiview decomposition, selective fusion, and multiscale feature aggregation.
• A tighter multiview learning generalization error bound guides the selective fusion method.
• The selective fusion method utilizes isomorphisms among multiple views to reduce outlier noise impact.

Main Results:

• The SAS framework demonstrated effectiveness and generalizability across seven synthetic datasets and two real-world scenes.
• Experiments showed superior performance compared to state-of-the-art SFF methods.
• The framework successfully mitigated outlier noise effects in scene reconstruction.

Conclusions:

• The proposed SAS framework significantly improves the generalizability of shape from focus techniques.
• This approach offers a robust solution for depth estimation, especially in scenarios with limited or no labeled data.
• SAS provides a promising direction for advancing SFF methods in computer vision.