1 Introduction
For several decades, emotion recognition has remained one of the of the most important problems in the field of human computer interaction. A large portion of the community has focused on categorical models which try to group emotions into discrete categories. The most famous categories are the six basic emotions originally proposed by Ekman in [1, 2]: anger, disgust, fear, happiness, sadness, and surprise. These emotions were selected because they were all perceived similarly regardless of culture.
Several datasets have been constructed to evaluate automatic emotion recognition systems such as the extended CohnKanade (CK+) dataset [3] the MMI facial expression database [4] and the Toronto Face Dataset (TFD) [5]. In the last few years, several methods based on handcrafted and, later, learned features [6, 7, 8, 9] have performed quite well in recognizing the six basic emotions. Unfortunately, these six basic emotions do not cover the full range of emotions that a person can express.
An alternative way to model the space of possible emotions is to use a dimensional approach [10] where a person’s emotions can be described using a lowdimensional signal (typically 2 or 3 dimensions). The most common dimensions are (i) arousal and (ii) valence. Arousal measures how engaged or apathetic a subject appears while valence measures how positive or negative a subject appears.
Dimensional approaches have two advantages over categorical approaches. The first being that dimensional approaches can describe a larger set of emotions. Specifically, the arousal and valence scores define a two dimensional plane while the six basic emotions are represented as points in said plane. The second advantage is dimensional approaches can output timecontinuous labels which allows for more realistic modeling of emotion over time. This could be particularly useful for representing video data.
Given the success of deep neural networks on datasets with categorical labels [11, 9], one can ask the very natural question: is it possible to train a neural network to learn a representation that is useful for dimensional emotion recognition in video data?
In this paper, we will present two different frameworks for training an emotion recognition system using deep neural networks. The first is a single frame convolutional neural network (CNN) and the second is a combination of CNN and a recurrent neural network (RNN) where each input to the RNN is the fullyconnected features of a single frame CNN. While many works [12, 13, 14, 15] have trained recurrent neural networks (RNNs) on handcrafted features, few works [16, 17] have considered the effects of using a CNN as input to the RNN. Our work, developed concurrently with [16, 17], demonstrates the benefits of using a CNN+RNN model and also shows how much the CNN and the RNN individually contribute to the overall performance.
Thus, the main contributions of this work are as follows:

We train both a singleframe CNN and a CNN+RNN model and analyze their effectiveness on the dimensional emotion recognition task. We also conduct extensive analysis on the various hyperparameters of the CNN+RNN model to support our chosen architecture.

We evaluate our models on the AV+EC 2015 dataset [13] and demonstrate that both of our techniques can achieve comparable or superior performance to the baseline model and other competing methods.
2 Dataset
The AV+EC 2015 [13] corpus uses data from the RECOLA dataset [18], a multimodal corpus designed to monitor subjects as they worked in pairs remotely to complete a collaborative task. The type of modalities include: audio, video, electrocardiogram (ECG) and electrodermal activity (EDA). These signals were recorded for 27 Frenchspeaking subjects. The dataset contains two types of dimensional labels (arousal and valence) which were annotated by 6 people. Each dimensional label ranges from . The dataset is partitioned into three sets: train, development, and test, each containing 9 subjects.
In our experiments, we focus on predicting the valence score using just the video modality. Also, since the test set labels were not readily available, we evaluate all of our experiments on the development set. We evaluate our techniques by computing three metrics: (i) Root Mean Square Error (RMSE) (ii) Pearson Correlation Coefficient (CC) and (iii) Concordance Correlation Coefficient (CCC). The Concordance Correlation Coefficient tries to measure the agreement between two variables using the following expression:
(1) 
where is the Pearson correlation coefficient, and
are the variance of the predicted and ground truth values respectively and
and are their means, respectively. The strongest method is selected based on whichever obtains the highest CCC value.3 Our Approach
3.1 Single Frame Regression CNN
The first model that we train is a single frame CNN. At each time point in a video, we pass the corresponding video frame through a CNN, shown visually in Figure 1
. The CNN has 3 convolutional layers consisting of 64, 128, and 256 filters respectively, each of size 5x5. The first two convolutional layers are followed by 2x2 max pooling while the third layer is followed by quadrant pooling. After the convolutional layers is a fullyconnected layer with 300 hidden units and a linear regression layer to estimate the valence label. We use the mean squared error (MSE) as our cost function.
All of the CNNs were trained using stochastic gradient descent with batch size of 128, momentum equal to 0.9, and weight decay of 1e5. We used a constant learning of 0.01 and did not use any form of annealing. All of our CNN models were trained using the anna software library
^{1}^{1}1https://github.com/ifpuiuc/anna.3.2 Adding Recurrent Neural Networks (RNNs)
Despite having the ability to learn useful features directly from the video data, the single frame regression CNN completely ignores temporal information. Similar to the model in [17], we propose to incorporate the temporal information by using a recurrent neural network (RNN) to propagate information from one time point to next.
We first model the CNN as a feature extractor by fixing all of its parameters and removing the regression layer. Now, when a frame is passed to the CNN, we extract a 300 dimensional vector from the fullyconnected layer. For a given time t, we take
frames from the past (i.e. ). We then pass each frame from time to to the CNN and extract W vectors in total, each length of 300. Each of these vectors is then passed as input to a node of the RNN. Each node in the RNN then regresses the output valence label. We visualize the model in Figure 2. Once again we use the mean squared error (MSE) as our cost function during optimization.We train our RNN models with stochastic gradient descent with a constant learning rate of 0.01, a batch size of 128 and momentum equal to 0.9. All of the RNNs in our experiments were trained using the Lasagne library ^{2}^{2}2https://github.com/Lasagne/Lasagne.
4 Experiments
4.1 Data Preprocessing
When preparing the video data, we first detect the face in each video frame using face and landmark detector in Dlibml [19]
. Frames where the face detector missed were dropped and their valence scores were later computed by linearly interpolating the scores from adjacent frames. We then map the detected landmark points to predefined pixel locations in order to ensure correspondence between frames. After normalizing the eye and nose coordinates, we apply mean subtraction and contrast normalization prior to passing each face image through the CNN.
4.2 Single Frame CNN vs. CNN+RNN
Table 1
shows how well our single frame regression CNN and our CNN+RNN architecture perform at predicting valence scores of subjects in the development set of the AV+EC 2015 dataset. When training our single frame CNN, we consider two forms of regularization: dropout (D) with probability 0.5 and data augmentation (A) in form of flips and color changes. For our CNN+RNN model, we use a single layer RNN with 100 units in the hidden layer and a temporal window of size 100 frames. We consider two types of nonlinearities: hyperbolic tangent (tanh) and rectified linear unit (ReLU).
From Table 1, we can see, not surprisingly, that adding regularization improves the performance of the CNN. Most notably, we see that our CNN model with dropout (CNN+D) outperforms the baseline LSTM model trained on LBPTOP features [13] (CCC = 0.326 vs. 0.273). Finally, when incorporating temporal information using the CNN+RNN model, we can achieve a significant performance gain over the single frame CNN.
In Figure 3, we plot the valence scores predicted by both our single frame CNN and the CNN+RNN model for one of the videos in the development set. From this chart, we can clearly see the advantages of using temporal information. The CNN+RNN model appears to model the ground truth more accurately and generate a smoother prediction than the single frame regression CNN.
Method  RMSE  CC  CCC 

Baseline [13]  0.117  0.358  0.273 
CNN  0.121  0.341  0.242 
CNN+D  0.113  0.426  0.326 
CNN+A  0.125  0.349  0.270 
CNN+AD  0.118  0.405  0.309 
CNN+RNN  tanh  0.111  0.518  0.492 
CNN+RNN  ReLU  0.108  0.544  0.506 
4.3 Hyperparameter Analysis
We study the effects of several hyperparameters in the CNN+RNN model, namely the number of hidden units, the length of the temporal window, and the number of hidden layers in the RNN. The results are shown in Tables 2, 3, and 4 respectively. Based on our results in Table 2, we conclude that it is best to have 100 hidden units given that both and resulted in decreases in performance. Similarly, for the temporal window length, we see that a window of length 100 frames appears to yield the highest CCC score, while reducing the window to 25 frames (1 second) and increasing it to 150 frames (6 seconds) both lead to significant decreases in performance. In Table 4, we see that increasing the number of hidden layers yields a small improvement in performance. Thus, based on our experiments, our best performing model had 3 hidden layers with a window length of W=100 frames, 100 hidden units in the first two recurrent layers and 50 in the third, and a ReLU as its nonlinearity.
Method  RMSE  CC  CCC 

CNN+RNN  h=50  0.110  0.519  0.485 
CNN+RNN  h=100  0.108  0.544  0.506 
CNN+RNN  h=150  0.112  0.529  0.494 
CNN+RNN  h=200  0.108  0.534  0.495 
Method  RMSE  CC  CCC 

CNN+RNN  W=25  0.111  0.501  0.474 
CNN+RNN  W=50  0.112  0.526  0.492 
CNN+RNN  W=75  0.111  0.528  0.498 
CNN+RNN  W=100  0.108  0.544  0.506 
CNN+RNN  W=150  0.110  0.521  0.485 
Method  RMSE  CC  CCC 

CNN+RNN  W=100  1 layer  0.108  0.544  0.506 
CNN+RNN  W=100  2 layers  0.112  0.519  0.479 
CNN+RNN  W=100  3 layers  0.107  0.554  0.507 
4.4 Comparison with Other Techniques
Table 5 shows how our best performing CNN+RNN model compares to other techniques evaluated on the AV+EC 2015 dataset. Both our single frame CNN model with dropout and our CNN+RNN model achieve comparable or superior performance compared to the stateoftheart techniques. Our single frame CNN model achieves a higher CCC value than the baseline[13] and is comprable with two other techniques [14, 15], all of which use temporal information. While our CNN+RNN model’s performance is not quite as strong as the CNN+LSTM model of Chao et al. [16], in terms of CCC value, we would like to point out that the authors used a larger CNN on a larger external dataset. Specifically, the authors trained an AlexNet[20] on 110,000 images from 1032 people in the Celebrity Faces in the Wild (CFW) [21] and FaceScrub datasets [22].
Method  RMSE  CC  CCC 
Baseline [13]  0.117  0.358  0.273 
LGBPTOP + LSTM [14]  0.114  0.430  0.354 
LGBPTOP+ Deep BiDir. LSTM [15]  0.105  0.501  0.346 
LGBPTOP+LSTM+loss [16]  0.121  0.488  0.463 
CNN+LSTM+loss [16]  0.116  0.561  0.538 
Single Frame CNN+D  ours  0.113  0.426  0.326 
CNN+RNN  W=100  3 layers  ours  0.107  0.554  0.507 
5 Conclusions
In this work, we presented two systems for doing dimensional emotion recognition: a single frame CNN model and a multiframe CNN+RNN model. We showed that our simple learned representation (single frame CNN) can outperform the baseline temporal model trained on handcrafted features. With the CNN+RNN model, we showed how incorporating temporal information can yield smoother and more accurate predictions. Lastly, we conducted an extensive hyperparameter analysis and selected a CNN+RNN model that achieved comparable or superior performance to other stateoftheart emotion recognition techniques on the AV+EC 2015 dataset.
References
 [1] Paul Ekman, Wallace V Friesen, Maureen O’Sullivan, Anthony Chan, Irene DiacoyanniTarlatzis, Karl Heider, Rainer Krause, William Ayhan LeCompte, Tom Pitcairn, Pio E RicciBitti, et al., “Universals and cultural differences in the judgments of facial expressions of emotion.,” Journal of personality and social psychology, vol. 53, no. 4, pp. 712, 1987.
 [2] Paul Ekman, “Strong evidence for universals in facial expressions: a reply to russell’s mistaken critique.,” 1994.
 [3] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews, “The extended cohnkanade dataset (ck+): A complete dataset for action unit and emotionspecified expression,” in CVPRW, 2010, pp. 94–101.
 [4] Michel Valstar and Maja Pantic, “Induced disgust, happiness and surprise: an addition to the mmi facial expression database,” in Proc. 3rd Intern. Workshop on EMOTION (satellite of LREC): Corpora for Research on Emotion and Affect, 2010, p. 65.
 [5] Josh M Susskind, Adam K Anderson, and Geoffrey E Hinton, “The toronto face database,” Department of Computer Science, University of Toronto, Toronto, ON, Canada, Tech. Rep, 2010.
 [6] Caifeng Shan, Shaogang Gong, and Peter W McOwan, “Facial expression recognition based on local binary patterns: A comprehensive study,” Image and Vision Computing, vol. 27, no. 6, pp. 803–816, 2009.
 [7] Mengyi Liu, Shaoxin Li, Shiguang Shan, and Xilin Chen, “Auaware deep networks for facial expression recognition,” in FG, 2013, pp. 1–6.

[8]
Ping Liu, Shizhong Han, Zibo Meng, and Yan Tong,
“Facial expression recognition via a boosted deep belief network,”
in CVPR, 2014, pp. 1805–1812. 
[9]
Pooya Khorrami, Tom Le Paine, and Thomas S. Huang,
“Do deep neural networks learn facial action units when doing
expression recognition?,”
in
Proceedings of the IEEE International Conference on Computer Vision Workshops
, 2015, pp. 19–27.  [10] James A Russell and Albert Mehrabian, “Evidence for a threefactor theory of emotions,” Journal of research in Personality, vol. 11, no. 3, pp. 273–294, 1977.
 [11] Samira Ebrahimi Kahou, Christopher Pal, Xavier Bouthillier, Pierre Froumenty, Çaglar Gülçehre, Roland Memisevic, Pascal Vincent, Aaron Courville, Yoshua Bengio, Raul Chandias Ferrari, et al., “Combining modality specific deep neural networks for emotion recognition in video,” in ICMI, 2013, pp. 543–550.
 [12] Fabien Ringeval, Florian Eyben, Eleni Kroupi, Anil Yuce, JeanPhilippe Thiran, Touradj Ebrahimi, Denis Lalanne, and Björn Schuller, “Prediction of asynchronous dimensional emotion ratings from audiovisual and physiological data,” Pattern Recognition Letters, 2014.
 [13] Fabien Ringeval, Björn Schuller, Michel Valstar, Shashank Jaiswal, Erik Marchi, Denis Lalanne, Roddy Cowie, and Maja Pantic, “Av+ ec 2015: The first affect recognition challenge bridging across audio, video, and physiological data,” in Proceedings of the 5th International Workshop on Audio/Visual Emotion Challenge. ACM, 2015, pp. 3–8.
 [14] Shizhe Chen and Qin Jin, “Multimodal dimensional emotion recognition using recurrent neural networks,” in Proceedings of the 5th International Workshop on Audio/Visual Emotion Challenge. ACM, 2015, pp. 49–56.

[15]
Lang He, Dongmei Jiang, Le Yang, Ercheng Pei, Peng Wu, and Hichem Sahli,
“Multimodal affective dimension prediction using deep bidirectional long shortterm memory recurrent neural networks,”
in Proceedings of the 5th International Workshop on Audio/Visual Emotion Challenge. ACM, 2015, pp. 73–80.  [16] Linlin Chao, Jianhua Tao, Minghao Yang, Ya Li, and Zhengqi Wen, “Long short term memory recurrent neural network based multimodal dimensional emotion recognition,” in Proceedings of the 5th International Workshop on Audio/Visual Emotion Challenge. ACM, 2015, pp. 65–72.
 [17] Samira Ebrahimi Kahou, Vincent Michalski, Kishore Konda, Roland Memisevic, and Christopher Pal, “Recurrent neural networks for emotion recognition in video,” in Proceedings of the 2015 ACM on International Conference on Multimodal Interaction. ACM, 2015, pp. 467–474.
 [18] Fabien Ringeval, Andreas Sonderegger, Jens Sauer, and Denis Lalanne, “Introducing the recola multimodal corpus of remote collaborative and affective interactions,” in Automatic Face and Gesture Recognition (FG), 2013 10th IEEE International Conference and Workshops on. IEEE, 2013, pp. 1–8.

[19]
Davis E King,
“Dlibml: A machine learning toolkit,”
The Journal of Machine Learning Research, vol. 10, pp. 1755–1758, 2009. 
[20]
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton,
“Imagenet classification with deep convolutional neural networks,”
in Advances in neural information processing systems, 2012, pp. 1097–1105.  [21] Xiao Zhang, Lei Zhang, XinJing Wang, and HeungYeung Shum, “Finding celebrities in billions of web images,” Multimedia, IEEE Transactions on, vol. 14, no. 4, pp. 995–1007, 2012.
 [22] HongWei Ng and Stefan Winkler, “A datadriven approach to cleaning large face datasets,” in Image Processing (ICIP), 2014 IEEE International Conference on. IEEE, 2014, pp. 343–347.
Comments
There are no comments yet.