Fast Piecewise-Affine Motion Estimation Without Segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Optimization Algorithms

Background:

• Traditional methods for piecewise affine motion estimation rely on sequential segmentation and estimation steps.
• These existing approaches often require initialization and can be computationally intensive, with costs dependent on motion field complexity.

Purpose of the Study:

• To develop a direct method for estimating piecewise affine motion fields that bypasses the need for intermediate segmentation.
• To improve computational efficiency and accuracy in motion estimation tasks.

Main Methods:

• Reformulated the motion estimation problem by enforcing piecewise constancy of the parameter field.
• Developed a proximal splitting optimization scheme tailored for this reformulated problem.
• Incorporated an efficient one-dimensional piecewise-affine estimator for vector-valued signals.

Main Results:

• The proposed method achieves competitive accuracy against existing piecewise-parametric techniques on benchmark datasets.
• Demonstrated superior performance compared to standard regularization methods like total variation and total generalized variation.
• The approach is initialization-free and exhibits computational costs independent of motion field complexity.

Conclusions:

• The direct estimation of piecewise affine motion fields offers significant advantages over segmentation-based methods.
• The novel optimization scheme provides a more efficient and accurate solution for motion estimation.
• This work advances the state-of-the-art in motion estimation by introducing a robust and computationally effective framework.