Use Case and High-Level Description

The anti-spoof-mn3 model is an anti-spoofing binary classificator based on the MobileNetv3, trained on the CelebA-Spoof dataset. It's a small, light model, trained to predict whether or not a spoof RGB image given to the input. A lot of advanced techniques have been tried and selected the best suit options for the task. For details see original repository.

Specification

Metric	Value
Type	Classification
GFlops	0.15
MParams	3.02
Source framework	PyTorch*

Accuracy

Metric	Original model	Converted model
ACER	3.81%	3.81%

Input

Original Model

Image, name: actual_input_1 , shape: [1x3x128x128], format: [BxCxHxW], where:

- B - batch size
- C - number of channels
- H - image height
- W - image width

Expected color order: RGB. Mean values: [151.2405,119.5950,107.8395], scale factor: [63.0105,56.4570,55.0035]

Converted Model

Image, name: actual_input_1 , shape: [1x3x128x128], format: [BxCxHxW], where:

- B - batch size
- C - number of channels
- H - image height
- W - image width

Expected color order: BGR.

Output

Original model

Probabilities for two classes (0 class is a real person, 1 - is a spoof image). Name: output1 Shape: [1,2], format: [BxC], where:

B - batch size
C - vector of probabilities.

Converted model

Probabilities for two classes (0 class is a real person, 1 - is a spoof image). Name: output1 Shape: [1,2], format: [BxC], where:

B - batch size
C - vector of probabilities.

Legal Information

The original model is distributed under the MIT License.