User-adaptive models for recognizing food preparation activities

Recognizing complex activities is a challenging research problem, particularly in the presence of strong variability in the way activities are performed. Food preparation activities are prime examples, involving many different utensils and ingredients as well as high inter-person variability. Recognition models need to adapt to users in order to robustly account for differences between them. This paper presents three methods for user-adaptation: combining classifiers that were trained separately on generic and user-specific data, jointly training a single support vector machine from generic and user-specific data, and a weighted K-nearest-neighbor formulation with different probability mass assigned to generic and user-specific samples. The methods are evaluated on video and accelerometer data of people preparing mixed salads. A combination of generic and user-specific models considerably increased activity recognition accuracy and was shown to be particularly promising when data from only a limited number of training subjects was available.