Multiple Visual Object Recognition For Poster Detection

We are aiming at an Augmented Reality application where multiple instances of a movie poster can be detected and localized. In that regard, we studied the problem of detecting and localizing multiple instances of planar objects, or objects which have repetitive patterns on it. Local image features, such as SIFT or SURF features are popular methods for image search and localization tasks. However classical methods which depend on initial feature matching, fail when there are multiple instances of the same object or when the objects have repetitive patterns on them. We extended and modified those methods in various ways to handle such situations. First of all, instead of an initial putative feature matching phase, we consider all possible meaningful matches in a Hough voting schema. Considering the nature of the target application which requires efficiency due to limited mobile platform CPU power and where the posters are distinctly apart, we do Hough transform only in translation space. Various subtle issues are also clarified, including an explicit formulation of the transformations. Techniques are verified with actual experiments.