Finding Headache Moments from YouTube Videos Using Weak Supervision

Abstract

Smart Monitoring of actions of elderly people have become more important these days since they mostly live by themselves. Majority of them have some kind of illness and require monitoring in their daily life. Computer Vision and Deep Learning can play a vital role in monitoring critical actions of the elderly people and the detected information can be very useful for their primary doctors and their kith and kin to care of them. To build a strong deep learning model that can detect ’headache’ moments from videos, the biggest bottleneck is huge amount of labeled training data. The reality is that hand labeled datasets are expensive and may take many months or in some cases years to create. The practical deployment of deep learning is hindered by the cost and intractability of hand labeling such datasets. This bottleneck has led to many machine learning systems use some form of weak supervision. The present study aims to use multiple weak supervision sources such as a pretrained deep learning model and several handcrafted heuristic rules; integrate and model them using Snorkel [1] which helps to programmatically build training datasets without manual labeling, from YouTube videos. A False Positive Rate (FPR) of 0.08% and False Negative Rate of 0.01% were achieved using the DNN model.This research study has shown that the accuracy of the combined weak supervision model is superior than the single pretrained model for programmatically building training dataset consisting of headache moments.