HSEmotion

HSEmotion (High-Speed face Emotion recognition) is a model that performs facial emotion recognition, distinguishing between 7 different expressions.

It is implemented in TensorFlow. The official code can be found here.

Weights are available in the releases section of the repository:

• mobilenet_7.h5: data/models/emotions_hse/mobilenet_7.h5

Performance on AffectNet dataset (7 classes)

Backbone	Accuracy	Inference time (s/img) - CPU	Inference time (s/img) - GPU
mobilenet_7	64.71	0.0782	0.0742

^{CPU: Intel(R) Xeon(R) Silver 4116 || GPU: Quadro RTX 8000}

Usage example

import cv2
from toolbox.Models.emotions_hse import EmotionsClassifier

img = cv2.imread("data/samples/images/faces/emotions/ffhq_215.png")

model = EmotionsClassifier(
    model_path="data/models/emotions_hse/mobilenet_7.h5", 
    use_cuda=False
)
instances = model.predict(img)

print(len(instances))
# > 1

print(instances[0].fields)
# > ['emotion', 'confidence']

print(f"{instances[0].emotion} ({type(instances[0].emotion)}): {instances[0].confidence}")
# > HAPPINESS (<enum 'Emotion'>): 0.9740390181541443

Project configuration YAML example:

face_emotions:
  model_name: emotions_hse
  params:
    model_path: ../../../data/models/emotions_hse/mobilenet_7.h5
    use_cuda: False