Cartoon-Face-Detection-and-Recognition

This repo contains the codes necessary to reproduce the results of the paper Towards Improved Face Detection and Recognition Systems.

The project maintains the following directories:

• preprocessing contains the files for parsing the XML files to filter the images based on the attributes required by the sub-tasks of the project (i.e., class-wise, gender-wise, facial posture wise manners).

• landmark_extractor contains the code for managing the 15 facial landmarks extraction, arranging these to the format compatible of being merged with the Kaggle instances and training the 5 layer LeNet architecture for landmark extraction. The output csv is further used in the character and gender recognition of the cartoon faces.

• datasets hosts two files:

  • train_own.csv is the output of running append_pixels_new.py on onlyLandmarks.csv, that contains comma separated landmark coordinates manually extracted using LandmarkManuallyGetter.jar.

• face_detection contains the code for running the MTCNN, Haar and HOG based models.

• face_recognition contains the code for the character recognition of the cartoons based on the Inception v3+SVM and the proposed HCNN model.

• gender_recognition contains the code for the character recognition of the cartoons based on the Inception v3+SVM and the proposed HCNN model.

• outputs contains the accuracy and top-5 error rate graphs for the character recognition problem, the model architectures used as well as the results of the face detection models.

Outputs

1. MTCNN face detection

2. OpenCV face detection

3. dlib face detection

4. Erroneously predicted landmark points

References

S. Jha, N. Agarwal, and S. Agarwal, “Towards Improved Cartoon Face Detection and Recognition Systems,” 2018, arXiv:1804.01753v1.