UDVDOptions
Options & Configuration · AiDotNet.Video.Options
Configuration options for UDVD, unsupervised deep video denoising with a blind-spot network (Sheth et al., ICCV 2021).
Inheritance: ModelOptions → NeuralNetworkOptions → UDVDOptions
For Beginners
UDVD learns to denoise from noisy video alone. Its network is built so that it can never see the pixel it is predicting, so the only way to predict that noisy pixel well is to predict the clean signal from the surrounding pixels and frames; the noisy pixel is then folded back in with Bayes’ rule.
How It Works
Defaults follow the paper and the authors’ release (github.com/sreyas-mohan/udvd, models/udvd.py, “blind-video-net-4”): five frames, two blind-spot U-Nets with 48-channel encoders and 96-channel decoders, the first stage emitting 32 channels per frame triplet, Gaussian noise of σ = 30/255, and Adam at 1e-4 halved at epochs 20, 25 and 30.
Constructors
| Constructor | Summary |
|---|---|
UDVDOptions() |
Initializes a new instance with default values. |
UDVDOptions(UDVDOptions 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 LearningRate { get; set; } |
Gets or sets Adam’s starting learning rate (1e-4). |
double? NoiseLevel { get; set; } |
Gets or sets the standard deviation (in normalized [0, 1] units) of the Gaussian noise the likelihood and the posterior assume; null estimates it from the frames. |
int DecoderFeatures { get; set; } |
Gets or sets the channels of each blind-spot U-Net’s decoder and second stage’s output (96). |
int EncoderFeatures { get; set; } |
Gets or sets the channels of each blind-spot U-Net’s encoder (48). |
int FirstStageFeatures { get; set; } |
Gets or sets the channels the first-stage U-Net emits for each frame triplet (32). |
string ModelPath { get; set; } |
Gets or sets the path to the ONNX model file. |