Case Studies

For deciding the best face detection algorithm for the thesis, see the case studies created (created in JupyterHub files). The two main factors for a real-time face detection system is efficacy and accuracy. These factors were also discussed and agreed upon with the bachelor thesis client. There are two case studies

1. Case Study - Comparing Accuracy for Real-Time Face Detection Models This case study measures the FPS over a dataset with ~2800 images. It presents the result in plotted in multiple graphs. See the case study for more information.

2. Case Study - Comparing Efficacy for a Real-Time Face Detection System This case study measures the memory usage of each algorithm. It plots the result as peak mega byte usage during a stress test of the algorithms. For this case, the python package memory_profiler is used. The result is plotted in multiple graphs. See the case study for more information.

3. Case Study - Calculating the focal length This case study calculates the focal length of camera based on our own dataset. The dataset is labeled with ground truths of the eye position and the known distance between the user and the camera. Iterating over this dataset, we are able to calculate the focal length of the camera with minimal error. See the case study for more information.

4. Case Study - Compare Tracking This case study compares each algorithms ability to track the users head. It uses a single video to create tracking videos. Each video can be manually verified - all videos are saved in the folder ./results/tracking_videos/. See the case study for more information.

Note: The case studies do not represent an absolute truth. There are most likely something that makes it not objectively true. Use them with caution.

Requirements

To successfully run this project, the following requirements must be met:

• Python Version 3.10.X
• Installed all packages listed in requirements.txt

Installation

All packages needed are in the requirements.txt. To set up the project with Visual Studio Code:

1. Go on a case study (file with the .ipynb)
2. Click on the Kernel button in the top right corner
3. Click Create Python Environment
4. Select Venv for the virtual environment
5. Select the Python with version 3.10.X
6. Check of requirements and press Ok to install dependencies
7. Wait a couple of minutes until the environment is set up

Then you can compile each code block in the case studies. Alternatively you can run the project in the browser by running the command:

jupyter notebook

Workflow

Code Owners are allowed to use the Github Workflow that triggers the compiling of the case studies. The workflow also downloads the dataset needed. The code for the workflow can be found here.

Estimated time of compiling both case studies (Note: the repository includes compiled notebooks):

~ 2h, 04 min

About the Git History

As of PR #55 in the OpenCV Server, this was originally part of the OpenCV Server code. The decision for moving it out of the repository was based on the following arguments:

• Reduced Repository Size: Separating case studies into their own repository keeps the main server lightweight, enhancing performance and ease of cloning and setup for new users.
• Improved Code Cohesion: By isolating the case studies, we can maintain a high level of cohesion within each repository, ensuring that each component focuses on a specific set of responsibilities.
• Simplified Version Control: With separate repositories, tracking and reverting changes becomes more straightforward, as the history of commits in each repo is more relevant to its specific content.
• Focused Issue Tracking and Documentation: Issues, pull requests, and documentation can be more precisely tailored to the repository's content, improving clarity and effectiveness in project management.
• For the User Convenience: Separating the case studies into their own repository simplifies the user experience. The user can utilize each repo without the added complexity of the other repository.

Because of this transition between repositories the commit history is quite short. See PR, Issues and commit related to the case studies before PR #55 (before 24th April 2024)