The previous chapters described an approach to object recognition based upon top-down predictions and iterative refinement. That discussion centered around the best system attributes for solving recognition problems in two example domains. Now a switch is made from constructing the algorithm to a more thorough evaluation of performance based on a wider range of scenes. Currently, the dataset contains 80 test problems (see Chapter 3). This chapter evaluates the entire Render-Match-Refine (RMR) system on those 60 images not used by the previous two chapters. Within these 60 images, there are 190 instances of 17 different objects.
KeywordsGround Truth Object Recognition Object Type Actual Image Translation Error
Unable to display preview. Download preview PDF.