Velocity Disambiguation for Video Frame Interpolation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing video frame interpolation (VFI) methods struggle with precise object movement prediction due to implicit time-to-location mapping.
• This often leads to blurry interpolated frames, especially with complex or long-range object trajectories.

Purpose of the Study:

• To develop a novel approach for VFI that improves frame sharpness and perceptual quality.
• To reduce the uncertainty in object speed prediction inherent in traditional VFI methods.

Main Methods:

• Introduced "distance indexing" to provide explicit object travel distance information to VFI models.
• Proposed an iterative reference-based estimation strategy to break down long-range predictions into short-range steps.
• Explored a continuous map estimator for pixel-wise dense distance indexing using multiple frames for enhanced motion disambiguation.

Main Results:

• Distance indexing significantly reduces blurriness and improves motion prediction accuracy in VFI.
• The iterative strategy and continuous map estimator enhance perceptual quality and quantitative performance.
• The proposed methods integrate seamlessly with existing VFI models without extra computational cost for uniform maps.

Conclusions:

• Distance indexing offers a clearer learning objective for VFI, leading to sharper and more accurate interpolated frames.
• The proposed strategies effectively handle complex motion and improve video quality.
• Manual distance indexing provides a novel tool for video editing and temporal manipulation.