The model was built mainly using Convolutional Neural Networks with multiple layers. It contains convolutional layers specifically with ReLU activation and also batch normalisation and max pooling layers to extract important features from the input images. The features are then flattened and passed through fully connected layers with ReLu activation, which results in a final softmax layer that predicts our data.
The training process first consists of Data Preprocessing, where the input images are resized because they must have the same dimensions. Afterwards, the model splits the data into training and validation datasets using a split ratio of 0.2. Also, we add Data Augmentation to improve our model, which includes random rotations, shifts, zooms, horizontal flips, and brightness variations. Data augmentation aids us in increasing the diversity of our data and makes the model familiar with unseen images. Finally, model training is performed, where it is compiled with the Adam optimizer and categorical cross-entropy loss function. Accuracy is used as evaluation metrics. The model is trained on the training data using the fit() function, specifying the number of epochs, batch size and validation data. Number of epochs was chosen to be 60, because the higher the number is, the better our model trains (not applied to all cases (can overfit the model)).
After the completion of training, the model's performance was estimated using the platform “CodaLab”. The model achieved an accuracy of 37,9% on the validation set. However, the accuracy was improved over and over by adding various techniques and increasing the number of epochs.
The chosen CNN architecture was very effective, but further experimentation with different ones, such as wider deeper networks, may result in a better result. The hyperparameters that were used in the model can greatly affect the model’s performance. In different situations, different hyperparameters were used such that the model is improved. At the end, the fixed size of learning rate, batch size, and epochs were chosen and thus were shown as optimal for the model. Regularisation techniques such as Dropout also showed lots of improvement for our model that prevented overfitting and improved generalisation.
In conclusion, in this project a deep learning model was developed for action classification based on input images. A CNN architecture was used and trained the model on the augmented data.