Hierarchical Motion Field Alignment for Robust Optical Flow Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Optical flow estimation is crucial for understanding motion in visual scenes.
• Accurate estimation of small and fast-moving objects remains a significant challenge.
• Existing methods struggle with large displacements and varying object sizes.

Purpose of the Study:

• To develop an improved optical flow estimation model.
• To specifically enhance performance in challenging scenarios with small and fast-moving objects.
• To ensure accurate motion field alignment and handle large displacements effectively.

Main Methods:

• Proposed a learning-based model with Hierarchical Motion Field Alignment module for accurate estimation across object sizes.
• Incorporated a Correlation Self-Attention module to effectively manage large displacements from fast-moving objects.
• Introduced a Multi-Scale Correlation Search layer to refine the four-dimensional cost volume for diverse motion types.

Main Results:

• The proposed model demonstrates superior generalization performance.
• Achieved significant improvements in estimating small, fast-moving objects.
• The model effectively addresses various types of motion through enhanced cost volume representation.

Conclusions:

• The developed model offers a robust solution for optical flow estimation, especially in complex scenarios.
• The integration of novel modules enhances accuracy and computational efficiency.
• This work advances the capabilities of computer vision in motion analysis.