Adaptive Template for Parsing Object of Indoor Scene Image

Abstract

In this paper, we propose an adaptive template for semantic labeling of indoor scene objects and estimating their oriented bounding facets (OBFs). The proposed adaptive template encodes prior geometric information of objects based on statistics of the training images. Given an input image, we utilize the adaptive template on the detected bounding boxes to initialize the raw labeling and OBF estimation of objects. To refine the initial results, multiple cubes/faces that follows geometric principles of adaptive template are generated to make up OBFs hypotheses. Each of the OBFs hypotheses is scored by the consistency matched with its corresponding semantic labeling result. The OBFs hypothesis that has the highest matching score with the corresponding labeling result is selected as the final parsing result. We evaluate our method on the bed, sofa and tea table categories, on both real and rendered indoor scenes. The experimental results show that our method has improved performance compared with the state-of-the-art detectors, and can give reasonable 3D interpretations of objects.