Recognition of cooking actions in video data
Mingyue Wei · Carolina Digital Repository (University of North Carolina at Chapel Hill) · 2019
This project focuses on using Machine Learning methods to recognize actions from cooking video sequences. The cooking videos and annotations are from MPII Cooking Activities Dataset, and a subset was constructed by selecting a group of the most frequently appearing actions, and generating matching video clips. Optical flows of those 11 actions’ videos were computed and were passed into a pre-trained Deep Learning Two-Stream Inflated 3D ConvNet to generate features of dimension 1024. Support Vector Machine classifiers and Random Forest classifiers were then applied to recognize and classify these actions using the extracted features. Binary SVM had the worst performance among the three classifiers we tested, whether considering imbalance in data or not. Multi-class SVM classifier and Random Forest classifier had similar results in precision of classification.