Facial Keypoint Detection using Pytorch, there are two models implemented:-

• Standard Model (CNN --> DNN)
• Convolutional Implementation (Only CNN)

The dataset is from Kaggle (https://www.kaggle.com/c/facial-keypoints-detection/data). The comprises of a csv file with all features along with the Image Pixel data, the images were transformed to 96x96 image with 15 keypoint feature pairs (i.e 30 output features).

It's architecture is defined as shown:-

96x96x1 -Conv5x5x32-> 94x94x32 -MaxPool2x2-> 46x46x32 -Conv3x3x64-> 44x44x64 -MaxPool2x2-> 22x22x64 -Conv3x3x128-> 20x20x128 -MaxPool2x2-> 10x10x128 -FC12800-> -FC1000-> -FC128-> -FC30->

Architecture is a defined:-

96x96x1 -Conv5x5x32-> 92x92x32 -MaxPool2x2-> 46x46x32 -3x3x64-> 44x44x64 -MaxPool2x2-> 22x22x64 -3x3x64-> 20x20x64 -7x7128-> 14x14x128 -5x5x32-> 10x10x32 -MaxPool2x2-> 5x5x32 -5x5x30-> 1x1x30

• Standard Model:-

  • Optimizer := SGD
  • Loss := MSELoss

• Convolutional Implementation Model

  • Optimizer := Adam
  • Loss := MSELoss