Deep ConvBi-LSTM: A Robust 3D Room Layout Estimation Model for Indoor Environment

Abstract

Room layout estimation is importance in recent times due to its extended application area. This process is highly challenging due to several factors affecting the room image such as clutter, occlusions, illuminations, etc. It is important to accurately identify the 3D layout of the room from a single 2D room image. The available techniques focused on determining the 3D layout but with limited number of features. It is important for a model to be fed with large number of features to result in successful predictions. To this extent, the proposed model introduced a robust 3D layout estimation framework for indoor environment. Initially, the input image is pre-processed and then subjected to layout estimation where our proposed model predicted both the edge maps and semantic labels for the image. For prediction, the proposed framework utilized the Deep ConvBi-LSTM model and a score function is defined and maximized by remora optimization algorithm (ROA) to obtain the optimal 2D layout from the candidate set. Finally, the 3D output is reconstructed from the 2D layout based on the layout coordinates and camera orientations. The experimental results of the proposed model proved the efficiency of the model in providing the desired performance.