robust-video-matting-mobilenetv3

Use Case and High-Level Description

The robust-video-matting-mobilenetv3 model is a robust high-resolution human video matting method that uses a recurrent architecture to exploit temporal information in videos and achieves significant improvements in temporal coherence and matting quality. This model is pre-trained in PyTorch* framework and converted to ONNX* format. More details provided in the paper. Backbone is MobileNetV3. For details see the repository. For details regarding export to ONNX see the instruction.

Specification

Metric

Value

Type

Background_matting

GFlops

9.3892

MParams

3.7363

Source framework

PyTorch*

Accuracy

Accuracy measured on a dataset composed with foregrounds from the HumanMatting dataset and backgrounds from the OpenImagesV5 one with input resolution 1280x720.

Metric

Original model

Converted model

Alpha MAD

20.79

20.82

Alpha MSE

15.1

15.11

Alpha GRAD

4.44

4.47

Foreground MSE

4.05

4.06

  • Alpha MAD - mean of absolute difference for alpha.

  • Alpha MSE - mean squared error for alpha.

  • Alpha GRAD - spatial-gradient metric for alpha.

  • Foreground MSE - mean squared error for foreground.

Input

Original Model

Image, name: src, shape: 1, 3, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Expected color order: RGB. scale factor: 255

Feature map, name: r1, shape: 1, 16, 144, 256, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r2, shape: 1, 20, 72, 128, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r3, shape: 1, 20, 36, 64, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r4, shape: 1, 20, 18, 32, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Converted Model

Image, name: src, shape: 1, 3, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Expected color order: BGR.

Feature map, name: r1, shape: 1, 16, 144, 256, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r2, shape: 1, 20, 72, 128, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r3, shape: 1, 20, 36, 64, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: r4, shape: 1, 20, 18, 32, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Output

Original model

Alpha matte. Name: pha Shape: 1, 1, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Foreground. Name: fgr Shape: 1, 3, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Feature map, name: rr1, shape: 1, 16, 144, 256, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr2, shape: 1, 20, 72, 128, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr3, shape: 1, 20, 36, 64, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr4, shape: 1, 20, 18, 32, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Converted model

Alpha matte. Name: pha Shape: 1, 1, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Foreground. Name: fgr Shape: 1, 3, 720, 1280, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - image height

  • W - image width

Feature map, name: rr1, shape: 1, 16, 144, 256, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr2, shape: 1, 20, 72, 128, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr3, shape: 1, 20, 36, 64, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Feature map, name: rr4, shape: 1, 20, 18, 32, format: B, C, H, W, where:

  • B - batch size

  • C - number of channels

  • H - feature map height

  • W - feature map width

Download a Model and Convert it into Inference Engine Format

You can download models and if necessary convert them into Inference Engine format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: