DIFRINT<T>

Models & Types · AiDotNet.Video.Stabilization

DIFRINT: Deep Iterative Frame Interpolation for Full-frame Video Stabilization.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → VideoStabilizationBase<T> → DIFRINT<T>

For Beginners

DIFRINT stabilizes shaky video by generating smooth intermediate frames. Unlike traditional stabilization that crops the frame, DIFRINT synthesizes full frames without losing any content from the edges.

Key advantages:

  • Full-frame stabilization (no cropping)
  • Handles large camera motions
  • Synthesizes missing content from warping
  • Real-time performance possible

Example usage:

How It Works

Technical Details:

  • Iterative refinement of stabilized frames
  • Flow-based motion estimation
  • Content synthesis for occluded regions
  • Temporal consistency enforcement

Reference: “DIFRINT: A Framework for Full-Frame Video Stabilization” https://arxiv.org/abs/2005.07055

Constructors

Constructor Summary
DIFRINT()
DIFRINT(NeuralNetworkArchitecture<T> architecture, DIFRINTOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction)
DIFRINT(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, DIFRINTOptions options)

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
List<Tensor<T>> Stabilize(List<Tensor<T>> frames)
List<ValueTuple<double, double, double>> EstimateMotionPath(List<Tensor<T>> frames) Estimates camera motion between frames.
ModelMetadata<T> GetModelMetadata()
Tensor<T> Stabilize(Tensor<T> unstableFrames)
Tensor<T> StabilizeFrame(Tensor<T> prevFrame, Tensor<T> currentFrame, Tensor<T> nextFrame) Synthesizes a stabilized frame by INTERPOLATING its two neighbours.