PerVFI<T>

Models & Types · AiDotNet.Video.FrameInterpolation

PerVFI: perception-oriented video frame interpolation.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → FrameInterpolationBase<T> → PerVFI<T>

For Beginners

PerVFI optimizes for what “looks good” rather than what’s mathematically closest, producing results that appear sharper and more natural.

Usage:

How It Works

PerVFI (2024) uses perceptual quality optimization for frame interpolation:

  • Perceptual loss hierarchy: uses a multi-scale perceptual loss computed from VGG/LPIPS

features, prioritizing visual quality over pixel-exact reconstruction

  • Asymmetric distortion: applies different loss weights to different frequency bands,

penalizing structural errors more heavily than texture errors

  • Perception-guided refinement: iteratively improves visual quality in regions where the

perceptual loss is highest using a refinement network with perceptual error maps

  • Motion-aware perceptual attention: uses estimated motion magnitude to weight the perceptual

loss, applying stronger constraints in high-motion regions

Reference: “PerVFI: Perception-Oriented Video Frame Interpolation” (2024)

Constructors

Constructor Summary
PerVFI(NeuralNetworkArchitecture<T> architecture, PerVFIOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a PerVFI model in native training mode.
PerVFI(NeuralNetworkArchitecture<T> architecture, string modelPath, PerVFIOptions options) Creates a PerVFI model in native training mode.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Interpolate(Tensor<T> frame0, Tensor<T> frame1, double t)