Patrick’s Status Reports

Completed: Set up AWS instance Processed 2058 of each emotion and trained Accuracy around 50 to 60% Used log loss for SVM Todo: Balance emotions in train/test split Currently shuffled split may not be balanced Migrate code to AWS and expand EBS storage Train second classifier for contempt, disgust, and Read more…

Completed: Trained with 60000 images Balanced out images of each emotion Confusion matrix Accuracy of around 60-70 percent 7000 images – got around 80-90 percent but could be overfit Tested on my own face – contempt, anger, disgust, fear are confused Todo: Make second SVM for classifying hard to differentiate Read more…

Completed: Trained with 10,000 images from AffectNet Accuracy still around 70% Lowest around 30% Highest at 77% Confusion matrix shows highest accuracy for happy, then anger, then sadness, surprise, fear, then contempt and disgust are not detected Downloaded 7 out of 11 parts of AffectNet so far To-do: Start batch Read more…

Completed: Personal SOW Downloaded AffectNet manually-annotated images Save trained SVM locally Testing different test/train splits Just testing and training further To-do: Train with more images Cross-validation on AffectNet Confusion matrix

Completed: Refocus SOW Setting up Zoom Migrating FER algorithm to home desktop computer Setting up PyCharm, Tensorflow, Keras, SkLearn, OpenCV Downloading datasets Reworking code to work in new environment To-do: Continue using more of AffectNet database Find more ways to improve accuracy Tweaking SVM parameters like bias, loss, etc Perform Read more…

Completed tasks: Test preliminary trained SVM Have SVM prediction in real-time Include AffectNet dataset in SVM training Design report To-do: Train with more AffectNet images Figure out how to output with probabilities for each emotion instead of single output Improve accuracy Set up AWS and train on server

Completed tasks: Automated I/O of CK+ dataset Localized pre-processed VGG19 features of CK+ dataset for faster training Preliminary training of SVM with 327 labeled pictures from CK+ dataset Working on evaluating trained SVM Implemented real-time prediction with camera input To-do: Evaluate performance of preliminary trained SVM Speed up real-time SVM Read more…

Completed tasks: Verified performance of VGG19 Working on PCA of features from 4th pooling layer of VGG19 100000 features to 200 principal components Still working on writing SVM To-do: Finish PCA and SVM algorithm Download CK+ database and begin preliminary training Design presentation and report