We observe how the Viola-Jones face detection algorithm performs on "non-human human" faces, i.e. cartoon characters. Our data consists of screenshots from three cartoons with very different drawing styles: Frozen, The Simpsons and South Park. We evaluate our detections using Intersection over Union (IoU) and also by counting misdetections and correct detections compared to number of all faces in the dataset. In the end we also try to enhance the detector by combining the Viola-Jones face and eye classifier.

• Anaconda Python Distribution, Python 3.7.4, installed from: https://www.anaconda.com/distribution/.
• R version 3.6.1, installed from: https://www.r-project.org/.

• Setup the environment using environment.yml file you find in the directory: conda env create -f environment.yml.
• Activate the environment: conda activate cartoon-face-detection.

.
├── data                    
│   ├── raw         		# screenshots in cartoon subfolders
│   └── test                # preprocessed images in cartoon subfolders, annotation json files
├── output                  # detection output in cartoon subfolders, results json file, plots
├── src                    	# Python and R scripts
├── environment.yml			# conda environment file
├── example.png
└── README.md

Run script.py to obtain detections. The script will take care of preprocessing, detection and will also evaluate the detections. Its output is a json file results.json containing Intersection over Union (IoU) values for each of the input images.

Before running the script, user can change the following parameters:

• cartoons list List of cartoons we want to use in our experiment.
• max_width integer Maximum width for the input images in px or None if you don't want to resize the images.
• rgb2gray boolean True if you want to convert the input images to grayscale.
• use_combined_detection boolean True if you want to use our combined VJ detector.
• save_faces boolean True if you want to draw the detections on the images and save them.

Result analysis is done by three R scripts:

• IoU_analysis.R: reads results.json, bootstraps the IoU values for each cartoon, plots 95% bootstrap percentile confidence intervals, saves the plot.
• count_detections_analysis.R: reads {cartoon}_table.csv files, which we made manually by counting detections from our output images. Plots number of misdetections and number of correct detections compared to number of all faces in the dataset per cartoon, saves the plot.
• vj_combo_analysis.R: reads frozen_table.csv and frozen_table_2.csv (file, made manually by counting detections if we used the combined VJ detector). Plots a barplot for detector comparison, saves plot.

To run the scripts set the value of variable path (first line in all scripts) to path to your project root directory. Project directory structure is assumed to stay the same.