This thesis focuses on incorporating target prior knowledge into structured output learning models. The output of the structured output learning model is a structure of several interdependent output variables. Obviously the interdependencies among the output variables play the key roles in the whole model. The target prior knowledge describing the output dependencies can just help the structured model construct more reliable dependencies, such that can enhance the model performance. Considering that the prior knowledge is always dependent on specific problems, this thesis studies three specific structured output learning problems, to reveal the correlation between target prior and structured output learning models. The three problems are video processing, multi-label learning and coupled conditional random field model respectively. The main contributions of this thesis are highlighted in three aspects: 1. We find that two related structured output learning models can provide addition- al target prior knowledge to each other. Such that if they can be jointly learned, then the overall performance of both models can be enhanced. To formulate this idea, we present a coupled hidden Markov random field model (CHMRF), which couples two different HMRF models based on the dependency between them. Specifically, we study two related problems in video processing, includ- ing face clustering and face tracking. Several target prior knowledge in videos is explored, including spatiotemporal knowledge, example/constraint smoothness assumption, the dependencies between cluster labels and tracklet associations, as well as the constraints among tracklet associations. Aforementioned target prior knowledge is systemically incorporated into the CHMRF model. Fur- thermore, based on CHMRF, we formulate the joint problem of simultaneous clustering and tracklet linking as a Bayesian inference problem, which can be effectively solved by a coordinate descent algorithm. 2. In multi-label learning, we study a general setting, i.e., multi-label learning with missing labels (MLML). MLML assumes that the class labels of training examples are partially provided, while other labels are missing. The positive, negative and missing labels are explicitly distinguished in our MLML setting. Such that the label bias of treating missing as negative labels that often occurs in existing works can be avoided. Furthermore, based on label consistency and example/class-level label smoothness, we formu...
修改评论