Discovering Efficiency in Coarse-To-Fine Texture Classification

We introduce a model for joint texture classification and segmentation that learns not only how to classify accurately, but when to classify efficiently. This model, combined with a complementary efficient feature representation that we describe, allows us to move beyond naive slidingwindow classification strategies into sub-linear coarse-tofine classification of an entire image. Recognition is formulated as a scale-space traversal through the image in which we can “stop short ” at coarse scales, dramatically increasing both the speed and the accuracy of classification. Unlike other models, ours is constructed such that the classification produced when stopping-short is exact (that is, equivalent to the classification produced when not stopping-short), because coarse-to-fine efficiency is directly incorporated into the model. Classification is demonstrated on partially- and fully-annotated datasets of satellite and medical imagery. for accurate classification that is, on average, of sub-linear complexity relative to the size of the image. A quad-tree[5] is used as a medium for multiresolution inference, and we features inspired by the integral image technique in [1] for efficient feature generation. The output of this system is shown in Figure 1, and a cartoon depiction of classification is shown in Figure 2. Multiresolution models have long been used for compact image representation[6], motion estimation[7], as well as classification and segmentation[8]. In [9], multiscale random fields are used for belief propagation through a quadtree hierarchy over an image. In [10], mixtures of treestructured belief networks are used for structural scene de-1.