FloRNNOptions

Options & Configuration · AiDotNet.Video.Options

Configuration options for FloRNN (Li et al., ECCV 2022), a unidirectional video denoiser whose look-ahead recurrent module mimics a backward one.

Inheritance: ModelOptions → NeuralNetworkOptions → FloRNNOptions

For Beginners

FloRNN removes noise from video one frame at a time but peeks a few frames ahead. A forward recurrent module carries what it has learned from past frames; a look-ahead module runs ahead and its features are warped back along the motion so the current frame also benefits from the future.

How It Works

Defaults follow the authors’ sRGB training script (github.com/nagejacob/FloRNN, train_models/sRGB_train.py): 64 feature channels, 15 residual blocks per recurrent module, a look-ahead of 3 frames, a border of 0.1 of each side for the forward-warping, and Adam at 1e-4 (the script then divides it by 10 for the last of 12 epochs).

Constructors

Constructor Summary
FloRNNOptions() Initializes a new instance with default values.
FloRNNOptions(FloRNNOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
OnnxModelOptions OnnxOptions { get; set; } Gets or sets the ONNX runtime options.
double BorderRatio { get; set; } Gets or sets the zero border, as a fraction of each side, around a hidden state before it is forward-warped; features splatted past the frame edge land there and are restored when the state is warped back (0.1).
double LearningRate { get; set; } Gets or sets Adam’s learning rate (1e-4).
double? NoiseLevel { get; set; } Gets or sets the noise standard deviation (in the frames’ own units) of the noise-level map that inference feeds; null estimates it from the frames.
int FlowPyramidLevels { get; set; } Gets or sets the pyramid levels of the frozen SPyNet flow estimator that stands in for the paper’s pretrained PWC-Net.
int ForwardCount { get; set; } Gets or sets how many frames the look-ahead module runs ahead of the output frame (3 in the authors’ sRGB script).
int NumFeatures { get; set; } Gets or sets the feature channels of both recurrent modules’ hidden states (64).
int NumResBlocks { get; set; } Gets or sets the residual blocks in each recurrent module (15 in the authors’ sRGB script).
string ModelPath { get; set; } Gets or sets the path to the ONNX model file.