Dynamic Frame Interpolation in Wavelet Domain.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Video frame interpolation is crucial for enhancing visual fluency by increasing frame rates.
• Current methods often overlook spatial redundancy, leading to inefficient computations.
• Adaptive computation compression requires understanding spatial-temporal information.

Purpose of the Study:

• To introduce WaveletVFI, a two-stage framework for efficient video frame interpolation.
• To address computational inefficiency by exploring spatial redundancy and adaptive compression.
• To improve frame interpolation performance through a novel wavelet-based synthesis approach.

Main Methods:

• Estimating intermediate optical flow using a lightweight motion perception network.
• Employing a wavelet synthesis network with sparse convolutions for multi-scale wavelet coefficient prediction.
• Utilizing dynamic thresholding, learned by a classifier, to determine sparse valid masks for adaptive computation.

Main Results:

• WaveletVFI reduces computation by up to 40% on high-resolution and animation benchmarks.
• The proposed method maintains similar accuracy compared to existing state-of-the-art techniques.
• Dynamic thresholding significantly improves computational reduction over fixed threshold methods.

Conclusions:

• WaveletVFI offers a computationally efficient solution for video frame interpolation.
• The framework effectively balances accuracy and computational cost through adaptive sparse convolutions.
• This work advances low-level vision tasks by optimizing spatial redundancy in frame synthesis.