Method and system for optimizing accuracy-specificity trade-offs in large scale visual recognition
As visual recognition scales up to ever larger numbers of categories, maintaining high accuracy is increasingly difficult. Embodiment of the present invention include methods for optimizing accuracy-specificity trade-offs in large scale recognition where object categories form a semantic hierarchy consisting of many levels of abstraction.
1. A method for classifying images, comprising:
receiving an input image to classify using a computer system;
scoring a likelihood of each individual node in a plurality of nodes of a classifier using a computer system, where the classifier includes a semantic hierarchy in which the plurality of nodes correspond to a hierarchy of named entities and a set of individual object classifiers to classify a likelihood that the input image contains a named entity in one of a plurality of leaf nodes from the plurality of nodes, where the plurality of leaf nodes correspond to a set of mutually exclusive named entities in the hierarchy of named entities;
selecting an individual node from the plurality of nodes most descriptive of the image using a computer system, where the individual node is determined by:
iteratively estimating a reward weight within the classifier that achieves a predetermined accuracy, where the accuracy of the classifier is determined by classifying a validation data set using the estimated reward weight;
determining reward weighted likelihoods using the estimated reward weight that achieves the predetermined accuracy; and
selecting as the individual node most descriptive of the image the individual node within the plurality of nodes in the semantic hierarchy that has the highest reward weighted likelihood;
classifying the input image as a named entity corresponding to the individual node most descriptive of the image using a computer system; and
returning the named entity as a classification of the input image using a computer system.
2. The method of claim 1 , wherein the plurality of nodes within the semantic hierarchy further comprises a plurality of internal nodes and a unique root node, wherein each of the internal nodes correspond to a named entity that is a union of named entities corresponding to at least one leaf node.
3. The method of claim 2 , wherein the input image is an unknown named entity, and wherein the individual node most descriptive of the input image is one of the plurality of internal nodes.
4. The method of claim 1 , wherein the set of individual object classifiers is a set of support vector machines (SVMs).