Recognition self-awareness for active object recognition on depth images

Abstract

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