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:
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.
- MTCNN face detection
- OpenCV face detection
- dlib face detection
- Erroneously predicted landmark points
S. Jha, N. Agarwal, and S. Agarwal, “Towards Improved Cartoon Face Detection and Recognition Systems,” 2018, arXiv:1804.01753v1.