Grant Perkins, Tian Yu Fan, Jean Claude Zarate, Mingjie Zeng
The accuracy of existing face detection and recognition models are greatly compromised when obstructions, like surgical masks, occlude facial features of the individual. This paper introduces a method which uses facial reconstruction as a new entry for face detection in a mask-wearing context. In the first phase of the model, an object detection network, called EfficientDet-D0, locates the position of the mask. In the second phase, a Gated Convolutional Network and SN-PatchGAN model, in a Generative Adversarial Network, work collaboratively to reconstruct the occluded region of the face. The EfficientDet-D0 model successfully detects masks with a prediction score of 0.966 mAP with an IoU threshold of 50%. The GAN was trained successfully and reconstructed the outline of a human face, but failed to reproduce detailed facial features due to time and hardware constraints.
- Docker
- built base docker image
- nvidia-docker