TransferLearning ResNet18: Facial Keypoints Detection with ResNet18

Project1: Pretrained ResNet18 in Pytorch for Facial Keypoints Detection

Project Experience

This is the version with pretrained ResNet18 for my first computer vision project which allows me to apply deep learning computer vision architectures (using Pytorch) to build a facial keypoint detection system for

• tracking
• pose recognition
• filters
• emotion recognition

Image Augmentation

Introduce generalization(randomness) to detect and learn structures

Transform Steps

Only the following transformation techniques were chosen to avoid unnecessary impacts on keypoints

• Rescale to 250 for width and height
• RandomCrop to 224
• Normalize & Convert ToTensor

Architecture & Performance

Pretrained ResNet18 is applied here as TransferLearning. Compared to the model without using TransferLearning: https://github.com/mminamina/Udacity_ComputerVision_PeerReviewed_Project---Facial_Keypoints_Detection Compared to the model with NaimishNet TransferLearning: https://github.com/mminamina/Transfer_Learning_NaimishNet--Keypoints_Detection

BatchSize, Epochs, Loss & Optimization Functions(using GPU)

• BatchSize : 32 for train dataset and 20 for test dataset
• Epochs : 8 (can train longer for better performance)
• Loss : SmoothL1Loss (significant loss reduction since earlier epochs)
• Optimizer : Adam

Worklist

For project instructions, please refer to https://github.com/udacity/P1_Facial_Keypoints

Files in Order

• models.py;
• Notebook 2. Define the Network Architecture.ipynb
• Notebook 3. Facial Keypoint Detection, Complete Pipeline.ipynb

Results

• model without TransferLearning vs NaimishNet vs Pretrained ResNet18
• Pretrained ResNet18 produces the best performance and the most accurate keypoints among all three models
• Both NaimishNet and Pretrained ResNet18 generate more accurate keypoints than the simple model without transfer learning

Additional Helpful Resources

• https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html
• https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf
• https://arxiv.org/pdf/1512.03385.pdf
• https://towardsdatascience.com/an-overview-of-resnet-and-its-variants-5281e2f56035

Packages

Python, Pytorch

LICENSE: This project is licensed under the terms of the MIT license.