“Video Activity Recognition with Limited Data for Smart Home Applications,” a Presentation from Comcast

This document discusses using machine learning and computer vision to recognize relevant events from home surveillance video with limited training data. It describes developing models that can identify the top 3-5 alert-worthy videos each day by incorporating object, spatial, and temporal context. The models aim to filter out unimportant motions like passing cars and focus on meaningful events like package deliveries. The techniques discussed include models that detect relevant motion events, incorporate scene layout into video representations to generalize to new environments with limited data, and leverage temporal patterns to identify anomalies. The goal is to deliver notifications of only the most significant events to enhance the customer experience of home security systems.