Dataset can be found in the following link http://vis-www.cs.umass.edu/lfw/
The pretrained model for alex net is automatically downloaded and loaded in training. 'alexnet': 'https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth'
dataset.py -- creates the image tensor by cropping the bounding box from the image and randomly apply data augmentation like crop with offset, flip ,brightening
LFWNet.py -- modified alex net for the LFW dataset
Training /Validation Error:
Accuracy:
An average 1% MSE is achieved on the testing data. Based on the loss, we are expecting 1%0.5 x 225 = 20 pixel absolute error. However, based on the percentage of detected points, the average error seems to be at 15 pixels, this is different than what we are expecting. One reason can be due to the variance of the testing data. We do expect better results to be achieved when more training data is given. For a future project, an increasing number of training data would be helpful. Also, applying random blurring for data augmentation can be implemented. Group the loss data on left eyes, right eyes, mouth, and nose can also be investigated. From our result, the prediction on the nose seems less accurate than other points.