Visual perception requires not only making inferences from observations, but also making decisions about what to observe. Though much of the computer vision literature implicitly assumes a well-captured visual observation as input, in reality a single view of a complex visual environment---or even multiple arbitrarily chosen views---may provide too little information for perception tasks. We aim to address the problem of "learning to look around" in the first place. Specifically, in a setting where a visual agent has the ability to voluntarily acquire new views to observe its environment, how can we train it to exhibit efficient exploratory behaviors to acquire informative observations? We treat this as a reinforcement learning problem, where a system is rewarded for actions that reduce its uncertainty about the unobserved portions of its environment. Based on this principle, we develop recurrent neural network-based systems to perform active completion of panoramic natural scenes and 3-D object shapes. Crucially, the learned policies are not closely tied to the particular semantic content seen during training; as a result, 1) the learned "look around" behavior is relevant even for new tasks in unseen environments, and 2) training data acquisition involves no manual labeling. Through tests in diverse settings, we demonstrate that our system learns useful and generic exploratory policies that transfer to new unseen tasks, an important step for autonomous embodied visual agents.

通过奖励代理的减少未观测环境部分的不确定性的行为，我们提出了一种基于循环神经网络的强化学习方法来实现对自然场景和三维物体的主动完成，并演示了我们的方法学习到的通用策略对于新的未见环境和任务具有较好的泛化性。

学习环顾四周：智能探索未知任务的未见环境