Regression Forest Based RGB-D Visual Relocalization Using Coarse-to-Fine Strategy

Abstract

Visual relocalization plays an important role in computer vision and robotics. However, feature ambiguities have made it remain challenging. In this work, we propose a novel regression forest based visual relocalization method that is performed in a coarse-to-fine manner. A topological regression tree is proposed to predict `coarse' subscenes where the camera locates. The pixel-coordinates correspondence regression tree is next employed to accomplish the camera pixel-coordinates predictions. By only considering the predictions within the predicted subscenes, we perform `fine' camera relocalization. We further propose to refine the pixel-coordinate predictions with graph-cut, which helps generating better pose hypotheses. We evaluate the proposed method on two public RGB-D datasets, including the 7Scenes (Microsoft) and the 12Scenes (Stanford University) datasets. We observe that the proposed method achieves accurate relocalization results and shows superior to or on-par accuracy with the state-of-the-art methods.