Active object recognition (AOR) refers to problems in which an agent interacts with the world and controls its sensor parameters to maximize the speed and accuracy with which it recognizes objects. A wide range of approaches have been developed to re-position sensors or change the environment so that the new inputs to the system become less ambiguous [1, 2] with respect to goals such as 3D reconstruction, localization or recognition of objects. Many of the active object recognition methods are built around a specific hardware system, which makes the replication of their results very difficult. Other systems use off-the-shelf computer vision datasets, which include several views of objects captured by systematically changing object's orientation in the image. However, these datasets do not offer any active object recognition benchmark per se.In this paper, we present and make publicly available the GERMS dataset (see Figure 1), that was specifically developed for active object recognition. The data collection procedure was motivated by the needs of the RUBI project, whose goal is to develop robots that interact with toddlers in early childhood education environments [4]. To collect data, we asked a set of human subjects to hand the GERM objects to RUBI in poses they considered natural. RUBI then pretends to examine the object by bringing it to its center of view and rotating the object. The background of the GERMS dataset was provided by a large screen TV displaying video scenes from the classroom in which RUBI operates, including toddlers and adults moving around. We also propose an architecture (DQL) for AOR based on deep Qlearning (see Figure 2). To our knowledge, this is the first work employing deep Q-learning for active object recognition. An image is first transformed into a set of features using a DCNN borrowed from [3] which was trained on ImageNet. We add a softmax layer on top of this model to recognize GERMS objects; the output of this softmax layer is the belief over different GERMS objects given an image. This belief is combined with the accumulated belief from the previous images using Naive Bayes. This accumulated belief represents the state of the AOR system in each time step.The accumulated belief is then transformed by the policy learning network into action values. This network is composed of two RectifiedLinear-Unit (ReLU) layers followed by a Linear-Unit (LU) layer. Each unit in the LU represents the action value for a given accumulated belief and one of the possible actions. In order to train this module, we implement the Q-learning iterative update:Figure 2: The proposed architecture for DQL.into the following stochastic gradient descent weight update rule for the network:Here, W is the weights of the policy learning network, Q(s, a) is the action-value learned by the network for action a in state s, γ is the rewarddiscount factor and R t is the reward at the tth time step. The number of output units in the policy learning network is equal to the number of possible actions. Each output uni...

No abstract

No abstract

No abstract

Active Object Recognition (AOR) has been approached as an unsupervised learning problem, in which optimal trajectories for object inspection are not known and are to be discovered by reducing label uncertainty measures or training with reinforcement learning. Such approaches have no guarantees of the quality of their solution. In this paper, we treat AOR as a Partially Observable Markov Decision Process (POMDP) and find near-optimal policies on training data using Belief Tree Search (BTS) on the corresponding belief Markov Decision Process (MDP). AOR then reduces to the problem of knowledge transfer from near-optimal policies on training set to the test set. We train a Long Short Term Memory (LSTM) network to predict the best next action on the training set rollouts. We sho that the proposed AOR method generalizes well to novel views of familiar objects and also to novel objects. We compare this supervised scheme against guided policy search, and find that the LSTM network reaches higher recognition accuracy compared to the guided policy method. We further look into optimizing the observation function to increase the total collected reward of optimal policy. In AOR, the observation function is known only approximately. We propose a gradient-based method update to this approximate observation function to increase the total reward of any policy. We show that by optimizing the observation function and retraining the supervised LSTM network, the AOR performance on the test set improves significantly.