Use Case and High-Level Description

Fully convolutional network for recognition of five emotions ('neutral', 'happy', 'sad', 'surprise', 'anger').

Validation Dataset

For the metrics evaluation, the validation part of the AffectNet dataset is used. A subset with only the images containing five aforementioned emotions is chosen. The total amount of the images used in validation is 2,500.

Example

Input Image	Result
	Happiness

Specification

Metric	Value
Input face orientation	Frontal
Rotation in-plane	±15˚
Rotation out-of-plane	Yaw: ±15˚ / Pitch: ±15˚
Min object width	64 pixels
GFlops	0.126
MParams	2.483
Source framework	Caffe

Accuracy

Metric	Value
Accuracy	70.20%

Performance

Inputs

Name: input, shape: [1x3x64x64] - An input image in [1xCxHxW] format. Expected color order is BGR.

Outputs

name: "prob", shape: [1, 5, 1, 1] - Softmax output across five emotions ('neutral', 'happy', 'sad', 'surprise', 'anger').

Legal Information

[*] Other names and brands may be claimed as the property of others.