Supplementary Material: A Personalized Affective Memory...
4
Supplementary Material: A Personalized Affective Memory Model for Improving Emotion Recognition Pablo Barros 1 German I. Parisi 12 Stefan Wermter 1 1. Personalized Recognition (P): Supplementary Results The main parameter to be optimized in our P category exper- iments was the maximum number of frames that we would use the PK generated images during the affective memory training. Table 1 illustrates the summary of the final results, in terms of Concordance Correlation Coefficient (CCC), when experimenting with different maximum frames over the entire OMG-Emotion test set. The optimal value, which we choose in our final experiments, was set to be 1 second. Table 1. Arousal and valence Concordance Correlation Coefficient (CCC) regarding the maximum duration (in frames and seconds) in which the PK image generation was turned off during the affective memory training. Frames (Seconds) Arousal Valence 1 (0.04) 0.28 0.36 5 (0.25) 0.34 0.40 12 (0.5) 0.38 0.47 25 (1) 0.43 0.53 37 (1.5) 0.40 0.48 50 (2) 0.41 0.46 62 (2.5) 0.39 0.45 87 (3.5) 0.37 0.44 100 (4) 0.36 0.43 112 (4.5) 0.37 0.42 125 (5) 0.37 0.44 250 (10) 0.36 0.42 500 (20) 0.37 0.43 750 (30) 0.36 0.43 all frames 0.37 0.42 * Equal contribution 1 Knowledge Technology, Department of Informatics, University of Hamburg, Germany 2 Apprente Inc, Montain View, CA, USA. Correspondence to: Pablo Barros <[email protected]>. Proceedings of the 36 th International Conference on Machine Learning, Long Beach, California, PMLR 97, 2019. Copyright 2019 by the author(s). 2. Examples of Generated Faces Figures 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 and 11 illustrate differ- ent examples of edited faces by the PK. We observe that the model imposes the desired facial characteristics on the person’s face, but maintain the general facial structure. The examples we illustrate here demonstrate the capability of the PK to impose expressions on very different faces. The original images were collected from both AffectNet and OMG-Emotion datasets. Figure 1.
Transcript of Supplementary Material: A Personalized Affective Memory...
Supplementary Material: A Personalized Affective Memory Model forImproving Emotion Recognition
Pablo Barros 1 German I. Parisi 1 2 Stefan Wermter 1
1. Personalized Recognition (P):Supplementary Results
The main parameter to be optimized in our P category exper-iments was the maximum number of frames that we woulduse the PK generated images during the affective memorytraining. Table 1 illustrates the summary of the final results,in terms of Concordance Correlation Coefficient (CCC),when experimenting with different maximum frames overthe entire OMG-Emotion test set. The optimal value, whichwe choose in our final experiments, was set to be 1 second.
Table 1. Arousal and valence Concordance Correlation Coefficient(CCC) regarding the maximum duration (in frames and seconds) inwhich the PK image generation was turned off during the affectivememory training.
Frames (Seconds) Arousal Valence1 (0.04) 0.28 0.365 (0.25) 0.34 0.4012 (0.5) 0.38 0.4725 (1) 0.43 0.53
37 (1.5) 0.40 0.4850 (2) 0.41 0.46
62 (2.5) 0.39 0.4587 (3.5) 0.37 0.44100 (4) 0.36 0.43
112 (4.5) 0.37 0.42125 (5) 0.37 0.44
250 (10) 0.36 0.42500 (20) 0.37 0.43750 (30) 0.36 0.43
all frames 0.37 0.42
*Equal contribution 1Knowledge Technology, Department ofInformatics, University of Hamburg, Germany 2Apprente Inc,Montain View, CA, USA. Correspondence to: Pablo Barros<[email protected]>.
Proceedings of the 36 th International Conference on MachineLearning, Long Beach, California, PMLR 97, 2019. Copyright2019 by the author(s).
2. Examples of Generated FacesFigures 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 and 11 illustrate differ-ent examples of edited faces by the PK. We observe thatthe model imposes the desired facial characteristics on theperson’s face, but maintain the general facial structure. Theexamples we illustrate here demonstrate the capability ofthe PK to impose expressions on very different faces. Theoriginal images were collected from both AffectNet andOMG-Emotion datasets.
Figure 1.
Supplementary Material: A Personalized Affective Memory Neural Model for Improving Emotion Recognition
Figure 2.
Figure 3.
Figure 4.
Figure 5.
Supplementary Material: A Personalized Affective Memory Neural Model for Improving Emotion Recognition
Figure 6.
Figure 7.
Figure 8.
Figure 9.
Supplementary Material: A Personalized Affective Memory Neural Model for Improving Emotion Recognition
Figure 10.
Figure 11.
