Models for multi-view object class detection

Abstract

Learning how to detect objects from many classes in a wide variety of viewpoints is a key goal of computer vision. Existing approaches, however, require excessive amounts of training data. Implementors need to collect numerous training images not only to cover changes in the same object's shape due to the viewpoint variation, but also to accommodate the variability in appearance among instances of the same class. We introduce the Potemkin model, which exploits the relationship between 3D objects and their 2D projections for efficient and effective learning. The Potemkin model can be constructed from a few views of an object of the target class. We use the Potemkin model to transform images of objects from one view to several other views, effectively multiplying their value for class detection. This approach can be coupled with any 2D image-based detection system. We show that automatically transformed images dramatically decrease the data requirements for multi-view object class detection. The Potemkin model also allows detection systems to reconstruct the 3D shapes of detected objects automatically from a single 2D image. This reconstruction generates realistic views of 3D models, and also provides accurate 3D information for entire objects. We demonstrate its usefulness in three applications: robot manipulation, object detection using 2.5D data, and generating 3D 'pop-up' models from photos.