Adviser Networks: Learning What Question to Ask for Human-In-The-Loop Viewpoint Estimation

Abstract

Humans have an unparalleled visual intelligence and can overcome visual ambiguities that machines currently cannot. Recent works have shown that incorporating guidance from humans during inference for real-world, challenging tasks like viewpoint-estimation and fine-grained classification, can help overcome difficult cases in which the computer-alone would have otherwise failed. These hybrid intelligence approaches are hence gaining traction. However, deciding what question to ask the human in the loop at inference time remains an unknown for these problems. We address this question by formulating it as what we call the Adviser Problem: can we learn a mapping from the input to a specific question to ask the human in the loop so as to maximize the expected positive impact to the overall task? We formulate a solution to the adviser problem using a deep network and apply it to the viewpoint estimation problem where the question asks for the location of a specific keypoint in the input image. We show that by using the keypoint guidance from the Adviser Network and the human, the model is able to outperform the previous hybrid-intelligence state-of-the-art by 3.27%, and outperform the computer-only state-of-the-art by 10.44% absolute.