Incremental learning of novel activity categories from videos

Abstract

We present a methodology for learning novel human activities incrementally. In many real-world scenarios (e.g. YouTube), new videos of novel activities are provided additively, and the system must incrementally adjust its activity models rather than retraining the entire system after each addition. We introduce our incremental codebook learning algorithm for an efficient mining of important visual words for human activities, and propose a method that incrementally trains activity models using them. The experimental results show that our approach successfully learns human activities from increasing number of training videos, while maintaining its recognition performance comparable to previous non-incremental systems.