TransferLearning Project: Facial Keypoints Detection with NaimishNet

Project 1: Facial Keypoints Detection with NaimishNet

Project Experience

This is the version with NaimishNet 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

NaimishNet is applied here as TransferLearning. (Compared to the model without using TransferLearning: https://github.com/mminamina/Udacity_ComputerVision_PeerReviewed_Project---Facial_Keypoints_Detection)

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

• BatchSize : 32 for train dataset and 20 for test dataset
• Epochs : 6 (can train longer for better performance)
• Loss : SmoothL1Loss (significant loss reduction since earlier epochs)
• Optimizer : Adam

NaimishNet with four Conv2D layers, ReLU Activation, MaxPooling+BatchNorm in every layer, three fully-connected layers, and Dropout to prevent overfitting.

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

• NaimishNet seems to generate more accurate keypoints compared to my original work without TransferLearning
• Please refer to https://github.com/mminamina/Udacity_ComputerVision_PeerReviewed_Project---Facial_Keypoints_Detection

Additional Helpful Resources

• https://arxiv.org/pdf/1603.07285.pdf
• http://cs231n.stanford.edu/slides/2016/winter1516_lecture11.pdf
• http://cs231n.stanford.edu/reports/2017/pdfs/300.pdf
• https://heartbeat.fritz.ai/5-regression-loss-functions-all-machine-learners-should-know-4fb140e9d4b0
• 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

Packages

Python, Pytorch

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