Universally Generalizable Foundation Model for Optical Flow

Develop a highly accurate and universally generalizable foundation model for optical flow that maintains robust performance across diverse domains and motion regimes, particularly for large-displacement motions.

Background

The paper highlights two core limitations of existing optical flow methods: dependence on task-specific features and domain-specific fine-tuning, which restricts out-of-distribution generalization, and architectural weaknesses in handling large displacements due to reliance on local iterative search.

These limitations motivate the need for a broadly applicable, accurate foundation model for optical flow that can generalize without per-domain adaptation and handle extreme motions.