Facial landmark detection is the task of detecting key landmarks on the face, given the static facial images. As a fundamental component in face analysis, facial landmark detection is applied in many areas, such as facial recognition, facial positional tracking, and facial editing.
In line with the recent success of neural networks, many methods are proposed to address this problem. However, developing a practical facial landmark detector remains challenging, as detection accuracy should always be a concern. As a supervised and regression task, our project investigates the CNN-based approach and the combination of CNN and the Transformer. Moreover, to obtain more promising performance in facial landmark detection, we construct a residual neural network (RNN) model, which eventually displays a significant decrease in loss during the validation process. After building these three models, we detect facial landmarks with the new data.
In 'dataset' file, we upload the full dataset of this project we used. Since the dataset is about 17GB, we upload the website directly. The data is from Kaggle.
In 'code' file, we upload all the code with some reports and interpretations, which contains all three models with data preprocessing and reality experimental results.