Indoor Video Layout Estimation Based on Plane Features and Motion Information
Weidong Zhang, Yulei Qiao · 2023
Estimating the 3D layout of an indoor scene is a critical task in computer vision. Its objective is to reconstruct the 3D structure of an enclosed space from an RGB images. At present, image-based layout estimation has been greatly developed. Unfortunately, the study of layout estimation based on video scenes that better match the real life have received little attention. In this paper, we focus on the task of indoor layout estimation in video scenes. In an indoor video, the room layout has no distortion and the only movement is from the camera, hence it is feasible to propagate the room layout information between the neighbouring video frames. Building on this understanding, we create a video-based dataset for 3D layout estimation, where the ground truth layout information is obtained by plane annotation and fitting as well as inter-frame layout transfer based on the camera pose. We also propose a method for video-based 3D layout estimation with a combination of intra-frame layout estimation and inter-frame camera pose estimation. The experimental results show that our method can efficiently estimate the indoor 3D layout in the videos. Compared to the existing image-based layout estimation methods, our method achieves significant performance gains.