Global feature space neural network for active computer vision

Simultaneous object class and pose estimation for mobile robotic applications with minimalistic recognition

“…Other techniques [6], [8], [9] use a large number of labeled images taken from different poses and attempt to match specific images in order to determine the pose. The accuracy of these approaches increases with the amount of training and reference images.…”

“…In this example we assume x 1 is known but the true object class and orientation φ i is unknown. This is a reasonable approximation in many active object search problems 6 . In the next section, we consider a grid-based object search and orientation estimation problem where this assumption is explicitly realized.…”

Global feature space neural network for active object recognition

“…We collectively refer to non-pose distortions as "noise" since they cause undesirable variations in the features observed, and ultimately cause errors in both classification and pose estimation. We have detailed a probabilistic FST description [9] in which the observed feature values are random variables where zero-mean, white Gaussian noise (with standard deviation…”

“…We have used this conditional PDF to show that the FST classification procedure discussed in the introduction approximates the minimum probability of error classifier and that the FST pose estimation algorithm approximates the maximum a posteriori pose estimate [9].…”

“…Our prior work [9] described in detail our FST-based probabilistic object representation and our Bayesian methods for estimating pose from a single object view, classification from multiple views, and determining where to look next for more reliable class or pose estimates.…”