Multi-level framework for object detection

摘要: The disclosure provides an approach for detecting objects in images. An object detection application receives a set of training images with annotations. Given these images, the generates semantic labeling detections, where includes lower-level subcategories and higher-level visual composites. In one embodiment, identifies using exemplar support vector machine (SVM) based clustering approach. Identified are used to initialize mixture components models which trains latent SVM framework, thereby learning number subcategory classifiers that produce, any given image, candidate windows associated labels. addition, learns structured model captures interactions among discriminative composites, labels spatial relationships between reason about interactions.