This paper presents a segmentation-free approach to optical character recognition (OCR) based on the concept of occluded object recognition, in which objects are recognized and then segmented out from the image. In applying the concept of occluded object recognition to the problem of OCR, we treat characters as touching or occluded objects that are subject to special constraints on their poses, i.e., they are juxtaposed with little or no freedom in rotation. Based on these characteristics, we combine two very powerful techniques used in occluded object recognition -- indexing and voting (pose clustering) -- and tailor them to the problem of OCR. This results in a segmentation-free OCR approach that is both highly efficient and robust. We note that recently some techniques have been proposed for handwritten OCR that conceptually are also segmentation-free, although these techniques are quite different from ours.