View-Action Representation Learning for Active First-Person Vision

In visual navigation, a moving agent equipped with a camera is traditionally controlled by an input action and the estimation of the features from a sensory state (i.e. the camera view) is treated as a pre-processing step to perform high-level vision tasks. In this paper, we present a representation learning approach that, instead, considers both state and action as inputs. We condition the encoded feature from the state transition network on the action that changes the view of the camera, thus describing the scene more effectively. Specifically, we introduce an action representation module that generates decoded higher dimensional representations from an input action to increase the representational power. We then fuse the output from the action representation module with the intermediate response of the state transition network that predicts the future state. To enhance the discrimination capability among predictions from different input actions, we further introduce triplet ranking loss and $N$ -tuplet loss functions, which in turn can be integrated with the regression loss. We demonstrate the proposed representation learning approach in reinforcement and imitation learning-based mapless navigation tasks, where the camera agent learns to navigate only through the view of the camera and the performed action, without external information.