Modified models for our paper Automated Detection of Equine Facial Action Units
Environment:
- pytorch >= 1.0
Modify config.py for configuration.
Call python main.py for training and test for the AU binary classification in cropped eye/lower face regions.
We modified the AlexNet and DRML for the binary classification in cropped eye/lower face regions.
The repos of the original models we referred to can be found at:
AlexNet: https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py
DRML: https://github.com/AlexHex7/DRML_pytorch
We employed Yolov3-tiny for ROI Detector: https://github.com/ultralytics/yolov3
Remember to change the filters and classes in the yolov3-tiny.cfg for the detection of one class.
Since Yolov3-tiny does not work well for small objects, we first detected the face regions and then detected eye/lower face regions from cropped face regions. The cropped eye/lower face regions will have the same label as the raw frames within these regions, and will be resized as 64x64 as the input to these modified models.