ssaHumanFace

ssaHumanFace contains trained computer vision models based on the YOLO11s architecture for detecting and analyzing human faces. The models are trained on a mixed dataset of real and stylized (sketched/animated) faces, primarily sourced from the Sesame Street Archive.
These models are designed to handle challenging cases in detecting and differentiating between real and stylized human faces, making them suitable for research on cross-domain face analysis.

Directory Structure

├── models/                                           # Pretrained YOLO11s weights
│   ├── All-Face Detection Phase/
|      ├──real_first.pt                               # Model for all face detection, trained on real faces first
|      └──stylized_first.pt                           # Model for all face detection, trained on stylized faces first
│   ├── Classification Phase/
|      ├──real_first.pt                               # Model for real/stylized face classification, trained on real faces first
|      └──stylized_first.pt                           # Model for real/stylized face classification, trained on stylized faces first
│   ├── Separated Phase/
|      ├──real.pt                                     # Model for real face detection
|      └──stylized.pt                                 # Model for stylized face detection
├── scripts/ 
|   ├──classify.py                                    # Script for detected images classification
|   ├──collect.py                                     # Script for collect images 
│   └──pred.py                                        # Script for a sample prediction
└── README.md

Example Usage

These pretrained models can be directly used through the YOLO pipeline. See the official YOLO website for more instructions on setting up the YOLO environment.

• Download the pretrained model you would like to try.
• Change the line model = YOLO("model_name.pt") # pretrained YOLO11n model into the pretrained model from this repository.
• Test the model on the images you selected.
• Sample predictions:

Application

• Human vs. stylized face detection
• Educational image gathering
• Research on cross-domain representation learning

Acknowledgement

• Ultralytics YOLO for the YOLO11s implementation
• iCartoon Dataset for pretraining on stylized faces