Real-time engagement-intensity prediction.
The project uses the award-winning approach proposed by the winners [1] of the engagement prediction task, a sub-challenge of the Emotion Recognition in the Wild Challenge (EmotiW 2019).
Video input is taken using the webcam mounted on the device. This input video stream is divided into ten seconds segments for evaluation. The divided segments are further sent in the order further down the model.
To obtain facial features, the video is fed into OpenFace. OpenFace gives 300 feature-points about eye-gaze-angles, eye-direction-vectors, 2-D eye-landmarks, 3-D eye-landmarks, and head pose. This is done for each video segment which consists of 100 frames. Further, the video is divided into 15 segments, for each segment, the standard deviation and mean is calculated for all the features to calculate the changes in values across the frames, this gives us about 60 feature-points. Finally, after the preprocessing is done, we get a 15x60 feature-set for each 10-second video segment.
The feature-set is then fed into a deep neural network. The architecture of the neural network consists of two LSTM layers, two fully-connected layers, and a global average pooling layer. The model then predicts the cognitive engagement intensity as a value between 0 and 1. Two different models were trained with slightly different hyperaprameters. Average of the values predicted by each of the two is taken.
The engagement-level is classified into four levels: Disengaged, Barely-engaged, Engaged and Highly-engaged. The value range for each level is as follows:
0 <= engagement-intensity < 0.4 : Disengaged
0.4 <= engagement-intensity < 0.6: Barely-Engaged
0.6 <= engagement-intensity < 0.83: Engaged
0.83 <= engagement-intensity <=1.00 : Highly-engaged
- Linux
- Nvidia GPU
-
Install OpenFace
-
Make the following modifications in OpenFace:
a. Replace the FeatureExtraction.cpp in OpenFace/exe/FeatureExtraction with FeatureExtraction.cpp in OpenFace mods folder.
b. Do the same for SequenceCapture.cpp in OpenFace/lib/local/Utilities/src and for SequenceCapture.h in OpenFace/lib/local/Utilities/include
c. Go to OpenFace/build and exectute
make. -
Delete the folder OpenFace/build/processed
-
Open terminal in OpenFace/build directory and run:
./bin/FeatureExtraction -wild -device 0 -pose -gaze -2Dfp -3DfpThis starts the video input and starts storing preprocessed data in the OpenFace/build/processed directory.
- From another terminal in the repository folder, run:
python predict.pyThe model was trained on a closed dataset “Engagement in the wild dataset” [2]. The data was recorded with a webcam on a laptop or computer, a mobile phone camera while the student participants were watching five minutes long MOOC video. The environment in which students watched the course videos varied from the computer lab, canteen, playground to hostel rooms. The dataset includes 91 subjects (27 females and 64 males) with 147, 48 and 67 videos, each approximately 5 minutes long, for training, validation, and testing, respectively. Four levels of engagement {0,1,2,3} were labeled by annotators. In this task, the problem was formulated as a regression problem with the output in the range [0,1] corresponding to 4 engagement levels (0 : 0, 0.33: 1, 0.66: 2, 1 : 3). The system performance was evaluated with mean square error (MSE) between the ground truth, and the predicted value of the test set.
[1] Van Thong Huynh, Hyung-Jeong Yang, Guee-Sang Lee, and Soo- Hyung Kim. 2019. Engagement Intensity Prediction with Facial Be- havior Features. In 2019 International Conference on Multimodal In- teraction (ICMI ’19), October 14–18, 2019, Suzhou, China. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3340555.3355714
[2] Amanjot Kaur, Aamir Mustafa, Love Mehta, and Abhinav Dhall. 2018. Prediction and localization of student engagement in the wild. In 2018 Digital Image Computing: Techniques and Applications (DICTA). IEEE, 1–8.
- th2l/SML_EW_EmotiW2019 Update: This repository has been taken down by its owner.
