Starting with DeepSight Toolkit 7.0, we are going to be introducing Emotion Recognition as a feature.
The detected emotions are:
- neutral
- anger
- disgust
- fear
- happiness
- sadness
- surprise
Each emotion can have a score between 1 and 0, with all emotions always adding up to 1 at any point in time.
To understand which emotion is most dominant for a given analysis, users should report the highest average per emotion given that the average exceeds 0.5. For example if the average emotion scores for ID 1 are the ones listed below, that ID / person can be reported as 'happy':
- neutral = 0.0 (avg)
- anger = 0.1 (avg)
- disgust = 0.1 (avg)
- fear = 0.1 (avg)
- happiness = 0.5 (avg)
- sadness = 0.1 (avg)
- surprise = 0.1 (avg)
dominant emotion: happiness = 0.5 (avg)
If for example all scores are low, or too close to each other, we recommend categorizing that ID / person as 'neutral':
- neutral = 0.2 (avg)
- anger = 0.1 (avg)
- disgust = 0.1 (avg)
- fear = 0.1 (avg)
- happiness = 0.2 (avg)
- sadness = 0.2 (avg)
- surprise = 0.1 (avg)
dominant emotion: none (neutral)
Head pose and lighting affect the accuracy of emotion estimations therefore we recommend only taking into account emotion estimations for faces detected within +/- 15 degrees of pitch and yaw (frontal faces).