Deep style estimator for 3D indoor object collection organization and scene synthesis

Estimating the style compatibility between a pair of cross-category 3D indoor objects has received wide interests from the field of computer graphics in these years. Many previous works solve this task by extracting and analyzing the style-aware structures or elements from the input 3D models. In this paper, we propose a novel approach to solve this task by training a deep neural network to quantitatively assign a compatibility score between arbitrary pair of cross-category 3D objects. By entirely learning from raw data, the trained network is able to capture various compatibility conditions influenced by global style features, such as ergonomics and object category relation. The proposed deep estimator is generally robust and can facilitate various high-level tasks. We first show its application for object collection organization. After that, we show how layout-guided, style-consistent object retrieval for indoor scene synthesis can be achieved by integrating pairwise style estimations into a novel submodular formulation. Our experiments demonstrate the usability of the proposed approach, demonstrating results superior than previous works and even comparable with suggestions made by human observers.