Analysis of the initial systems' failures revealed a serious limitation: inaccuracies in the sensor and object models were responsible for most of the recognition errors and reduced efficiency. My current research is aimed at addressing this problem by replacing the ray tracer with a real sensor--building the recognition model from real images of the object. Much of the previous work on learning/acquiring object models uses the 2D view-based approach where the model is represented by a large number of representative views. Though these techniques have shown promise within restricted domains, I believe that a 3D representation is necessary for localization and, thus, reliable verification. Accurate appearance prediction is the key to reliable localization and verification; thus, the recognition model must enable us to predict the appearance of object primitives (e.g., range data points or edge points) given the object position.
For the range image domain, a model of the object's surface is sufficient for appearance prediction. I have built upon previous techniques to merge multiple range image views of an object into a single geometric surface model. This geometric surface model is then used to build the model of the edge-point appearance for the intensity image domain. The surface model provides a mapping from the 2D edge points to 3D coordinates. I have developed techniques to merge multiple views of intensity edge images into a unified geometric edge-point representation. Current work is focused on inferring the geometric and photometric appearance constraints from the sample images. The appearance model will also be used for selecting statistically significant groups of features for feature matching.