A discriminative framework for object recognition

In this paper we present a discriminative approach to learn and recognize the object classes from unsegmented cluttered scenes in a scale invariant manner. A multiscale algorithm for the selection of salient regions of an image is used to select keypoints and their scale in the image. The PCA-SIFT method is used to describe these keypoints in a compact form. For each object class the probability of local features is modeled by a Conditional Random Fields (CRF). In the learning stage, the parameters of CRY are estimated from feature vectors given the labels in a maximum likelihood framework. In the recognition stage, we take the label for the image to the most likely class under the CRF models. This method achieves good classification results on motorbikes and airplanes database.