Test-Time Adaptation for Video Frame Interpolation via Meta-Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video frame interpolation (VFI) is complex due to varied motion, frame rates, and occlusions.
• Generalizing VFI across diverse scenes with a single fixed-parameter network is challenging.
• Developing unique networks for each scenario is computationally infeasible.

Purpose of the Study:

• To propose MetaVFI, an adaptive VFI algorithm leveraging readily available test-time information.
• To enhance VFI efficiency and performance through meta-learning and test-time adaptation.
• To create a model-agnostic VFI framework adaptable to existing networks.

Main Methods:

• Employing test-time adaptation via network fine-tuning.
• Integrating meta-learning to achieve efficient adaptation with a single gradient update.
• Developing MetaVFI as a model-agnostic framework compatible with various VFI networks.

Main Results:

• Achieved significant performance gains in video frame interpolation.
• Demonstrated improved efficiency through meta-learning without additional parameters.
• Validated MetaVFI's effectiveness on multiple benchmark datasets, enhancing baseline VFI networks.

Conclusions:

• MetaVFI offers an effective adaptive solution for video frame interpolation.
• The proposed meta-learning approach enables efficient, parameter-free adaptation.
• MetaVFI significantly boosts the performance of existing VFI models across various datasets.