Active Vision in the Era of Convolutional Neural Networks

摘要

In this work, we examine the literature of active object recognition in the past and present. We note that methods in the past used a notion of recognition ambiguity in order to find a next best view policy that could disambiguate the object with the fewest camera moves. Present methods on the other hand use deep reinforcement learning to learn camera control policies from the data. We show on a public dataset, that reinforcement learning methods are not superior to a policy of adequately sampling the object view-sphere. Instead of focusing on finding the next best view, we examine a recent method of quantifying recognition uncertainty in deep learning as a potential application to active object recognition. We find that predictions with this technique are well calibrated with respect to the performance of a network on a test-set, showing that it could be useful in an active vision scenario.