Skip to Main Content
There are two big issues emerging in the field of computer vision: one is the explosively increasing large amount of visual data and the other is the demand of deep labeling of objects and scenes. In this paper, we propose a structured outputs prediction framework equipped with a discriminative model and a corresponding efficient online learning algorithm. Instead of doing simple multiclass classification as usual, we aim at outputting structured labels which means different label confusion mistakes may have different costs. Moreover, the online learning algorithm with efficient updating strategy and compact memory management mechanism makes the framework work well on large visual data. Experiments on two representative datasets show an exemplar application of our model.
Date of Conference: Sept. 27 2009-Oct. 4 2009