Action Recognition in Streaming Videos via Incremental Active Learning
In this talk, we introduce a novel incremental and active learning classification approach that can be used with any local or global set of feature descriptors extracted from a segmented video stream. Our system is nonparametric: it covers the feature space with classifiers that locally approximate the Bayes optimal classifier. We focus on streaming scenarios, in which our approach features incremental model updates and on-the-fly addition of new classes. Moreover, predictions are computed in time logarithmic in the model's size (which is typically fairly small), and active learning is used to save labeling costs. A ``constant budget'' variant is also presented to limit the grow of model size over time, as an appealing feature in real-time applications. We apply this methodology to human activity recognition tasks. Experiments on standard benchmarks show that our approach is competitive with state-of-the-art non-incremental methods, and outperforms the existing active incremental baselines.