ABSTRACT
We present a discriminative latent variable model for classification problems in structured domains where inputs can be represented by a graph of local observations. A hidden-state Conditional Random Field framework learns a set of latent variables conditioned on local features. Observations need not be
independent and may overlap in space and time. We evaluate our model on object detection and gesture recognition tasks.
HCRF原理
Differently, an HCRF models the distribution P (c, h/x) directly, where c is a category and h is an intermediate hidden variable modeled as a markov random field globally conditioned on observation x. The parameters µ of the model are trained discriminatively to optimize P (c/x).
Our hidden-state discriminative approach for object recognition is related to the work of Kumar and Herbert [12], [11], who train a discriminative model using fully-labeled data where each image region is assigned a part label from a discrete set of object parts. A CRF is trained and detection and segmentation are performed by finding the most likely labeling of the image under the learned model. The main difference between our approach and Kumar’s is that we do not assume that the part assignment variables are fully observed and are instead regarded as latent variables. Incorporating hidden variables allows use of training data not explicitly labeled with part (hidden-state) structure.
In previous work on CRFs label sequences are typically taken to be fully observed on training examples. In our approach category labels are observed, but an additional layer of subordinate labels are learned. These intermediate hidden variables model the latent structure of the input domain; our model defines the joint probability of a class label and hidden state labels conditioned on the observations, with dependencies between the hidden variables expressed by an undirected graph. The result is a model where inference and parameter estimation can be carried out using standard graphical model algorithms such as loopy belief propagation.