Code for the paper "Modeling Multiple Temporal Scales of Full-body Movements for Emotion Classification"
Full Paper and Supplementary Material
Video Presentation
Proposed method employs a two-branch architecture. It consists of two CNNs, each of them is composed of three convolutional layers followed by fully connected layers. The shape of the convolutional filters are extended along the time axis to form 3x5 rectangles. The reason for having rectangular filters is that we expect the network to learn and extract features in the time domain rather than among successive markers. It is also important to highlight that the input image is always rectangular. The first convolution is applied to the input image with 16 filters. The "same" padding, which makes the size of outputs the same as that of inputs, is used. A max pooling operation with a stride of 2 is performed, which reduces both x and y dimensions by half. The obtained result is given to the next layer after applying a ReLU function.
In the second layer 32 filters are used while in the third layer 64 filters are used. Such increase in the number of filters allows us to identify more complex features in the deeper layers. After the third and final max pooling and ReLU layers, the image size is reduced to 3x12, but with 64 layers. The output is then flattened out. In this two-branch architecture, separate convolutional layers are used before the weights are flattened out and these weights are added together in a fully-connected layer. Finally, another fully connected layer of dimension cx1 (c equals to the number of emotion classes) is used as the output layer such that each output value corresponds to one emotion class. A softmax function in the output layer determines the final emotion class for the given input image.
The input of our model at a time is a part of a data segment in the form of two RGB images in logistic position image format. The two-branches of the proposed architecture take images both having the size M=NxK (while K is defined by the number of markers). Starting from an image I having the size MxK) corresponding to a certain data chunk duration and I_2 that is a part of I corresponding to the last portion of I (e.g., the last quarter), thus its size is M=NxK, where 1<N<M, first, image resizing with bi-cubic interpolation is applied to I, resulting in I_1, which is M=NxK. Then, I_1 and I_2 are given to the network as the inputs, simultaneously.
Data consisting of the 3D-positions of e.g., 30 markers at 100 fps was converted into RGB images, which is a common input format for CNNs. This includes dividing the MoCap segments as identified by the experts, which have a variable duration, into chunks of fixed duration. Then, a chunk of data is converted into an RGB image. Various values were tested for the duration of a single chunk while overlapping in time was also applied.
The procedure for constructing RGB images includes bodycentred relative normalization. The value of marker e.g., CLAV at the first frame of each chunk is taken as a point of reference. CLAV is situated on the lower part of the chest on the xiphoid process. So in the first frame, the position of CLAV is zero. The positions of the other markers are taken with respect to this new origin. By using body-centred relative positioning, the range of the marker values is reduced, thus, it is no longer required to map all the positions of the workspace.
An 8-bit RGB image format is used to represent the data such that the X, Y and Z coordinates of the markers are associated with the R, G and B layers, respectively. Markers are represented on the y-axis, while the consecutive frames of the sequence are represented on the x-axis. For example, a row of the R layer in the resulting image represents the temporal evolution of the X coordinate of the marker associated with that row. Then, logistic position is used to fit the information in this 8-bit image format. We use a logistic function that maps the positions into the -127 to +127 interval (see paper for the equation used).
obtainRGBImages_twoBranch_Demo.m: Generates RGB images in logistic position format and also applies data augmentation. The input Mocap data is stored under the folder MocapData. The resuting images are saved into the Images folder.
applyImageGeneration.mm: To generate images in logistic position format.
augmentData.m: To apply proposed data augmentations.
resizeFirstBranch.m: To resize the inputs of the first branch to the size of the images input to the second branch.
traintest_DEMO.py: Train and test the proposed architecture in a single file when the inputs are the images saved into the Images folder.
- Python 3.5
- Tensorflow 1.12
- Matlab 2015a or later
1- Use the Matlab code (obtainRGBImages_twoBranch_Demo.m) to obtain RGB images in logistic position format.
2- Use the Python code (traintest_DEMO.py) to train and test the method.
C. Beyan, S. Karumuri, G. Volpe, A. Camurri and R. Niewiadomski, "Modeling Multiple Temporal Scales of Full-body Movements for Emotion Classification", in IEEE Transactions on Affective Computing, DOI: 10.1109/TAFFC.2021.3095425.
@ARTICLE{affectiveBodyMovements,
author={Cigdem Beyan and Sukumar Karumuri and Gualtiero Volpe and Antonio Camurri and Radoslaw Niewiadomski},
journal={IEEE Transactions on Affective Computing},
title={Modeling Multiple Temporal Scales of Full-body Movements for Emotion Classification},
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/TAFFC.2021.3095425}}