The recognition scheme (shown in Fig.1) is based on two distinct stages. The first stage relies on indexing. However, to keep space requirements low we index only a sparse number of sampled views per model. To improve the quality of hypotheses generated during recognition, we have replaced hashing, which performs a range search, with a more powerful indexing scheme based on k-d tree. In the second stage, we learn the manifold formed by a dense number of sampled views per model using the EM algorithm. Learning takes place in a "universal", lower-dimensional space computed through random projection (RP). The only information that needs to be stored at this stage is just a few parameters for each manifold.

The main purpose of the first stage is to generate hypothetical matches between the models and the scene very fast. And the main purpose of the second stage is to filter out quickly and inexpensively as many invalid hypotheses as possible. This stage provides a way to rank each hypothesis prior to verification. This saves time since only hypotheses ranked high enough are considered for further verification. Verification is performed by matching the predicted model appearances with the scene.

Figure 1 The recognition framework