Recognition self-awareness for active object recognition on depth images

We propose an active object recognition framework that introduces the recognition
self-awareness, which is an intermediate level of reasoning to decide which views to
cover during the object exploration. This is built first by learning a multi-view deep 3D
object classifier; subsequently, a 3D dense saliency volume is generated by fusing together
single-view visualization maps, these latter obtained by computing the gradient
map of the class label on different image planes. The saliency volume indicates which
object parts the classifier considers more important for deciding a class. Finally, the
volume is injected in the observation model of a Partially Observable Markov Decision
Process (POMDP). In practice, the robot decides which views to cover, depending on the
expected ability of the classifier to discriminate an object class by observing a specific
part. For example, the robot will look for the engine to discriminate between a bicycle
and a motorbike, since the classifier has found that part as highly discriminative. Experiments
are carried out on depth images with both simulated and real data, showing that our
framework predicts the object class with higher accuracy and lower energy consumption
than a set of alternatives.