This project focuses on analyzing an existing facial landmarks dataset and exploring the impact of different scaling methods on subsequent analysis. Our aim was to compare the original scaling method with three alternative approaches and observe their effectiveness when applied to various studies. By determining the most suitable scaling method, we hope to contribute to improved accuracy and efficiency in applications like facial recognition, emotion detection, and computer vision tasks.
The dataset under analysis contains facial landmarks from hundreds of images, each comprising 68 distinct landmarks represented by their x and y coordinates. We began by examining the original scaling method, which utilized the No.33 landmark (nose) as the center and scaled the landmarks based on the range of the axis value.
To explore alternative scaling methods, we applied three new approaches:
- Multidimensional Scaling (MDS): Applied using
sklearn.manifold.MDS(“sklearn.manifold.MDS”). - Standardization: Implemented using
sklearn.preprocessing.StandardScaler(“Sklearn.Preprocessing.StandardScaler”) . - Normalization by Bounding Box: Landmarks were scaled based on the bounding box of the face, using separate x and y ranges.
The rationale behind selecting these methods was to investigate the effectiveness of various scaling techniques in aligning and representing facial landmarks, ultimately enabling more accurate analysis in different studies.
Similarly, with the previous project, we performed center alignment using the No.33 landmark (nose) as the center.
After aligning the landmarks, we applied the above three scaling methods to the centered landmarks too. Each of these scaling methods was applied to the dataset to explore their impact on the alignment and representation of facial landmarks, with the ultimate goal of identifying the most suitable method for various studies in the field.
After applying center alignment and scaling to the facial landmarks, the previous project introduced a rotation method to further enhance their representation. With applying the above 3 scale methods to the rotation data, we obtained another CSV file containing 3 scaling types, previous project data, and original data. This further enhances our analysis of the impact of various scaling and rotation methods on the alignment and representation of facial landmarks in different studies.
We generated visualizations to better understand the impact of the scaling and rotation methods on the facial landmarks. Four figures were plotted, showcasing the alignment and representation of the landmarks for different methods:
- Figure 1: Infant landmarks after aligning center
- Figure 2: Adult landmarks after aligning center
- Figure 3: Infant landmarks after Rotation
- Figure 4: Adult landmarks after Rotation
From the visualizations, we observed that the MDS scaling method altered the face orientation, making it unsuitable for our objective. Similarly, the Standardization method was found to be unsuitable due to its use of the mean as the zero point. The other two methods (excluding Original) showed similar results, making it challenging to determine which one is superior. These visualizations aid in understanding the effectiveness of each scaling and rotation method on the alignment and representation of facial landmarks, providing valuable insights for future work and applications.
In comparison to the previous project, our project retained the initial two preprocessing steps, aligning the center and rotation, to ensure minimal influence on the outcome. We experimented with alternative scaling methods, including MDS, standardization, and a different normalization approach. We discovered that MDS resulted in varying face directions, which increased complexity and rendered outlier detection ineffective. Standardization maintained the face direction but shifted the center slightly, yielding similar results. Normalization by bounding box provided more reasonable landmarks, and despite the minor differences, it proved to be the best method for keeping the landmarks in an organized range. Thus, scaling indeed impacts the outcome significantly, and the normalization method is the most effective choice for maintaining an organized landmarks after aligning and rotation.
In this section, we analyzed various scaling and rotation methods for facial landmarks, comparing the original method with three alternatives: MDS, Standardization, and Normalization by Bounding Box. The MDS method was found to be unsuitable, as was Standardization due to its use of the mean as the zero point. The other methods yielded similar results, making it challenging to determine the most effective approach.
- “sklearn.manifold.MDS.” Scikit-learn, scikit-learn.org/stable/modules/generated/sklearn.manifold.MDS.html.
- “Sklearn.Preprocessing.StandardScaler.” Scikit-learn, scikit-learn.org/stable/modules/generated/sklearn.preprocessing.StandardScaler.html.