Learning Depth of Edges for 3D Layout Estimation

The 3D layout estimation of indoor scenes aims to recover the overall structure of the scene from indoor RGB images, with a core focus on reconstructing the geometric structure of dominant planes. However, existing methods mostly concentrate on directly learning planar depth parameters or planar parameters, which depend on the position of the plane within the image and are sensitive to noise. To address these issues, we propose a method that learns the depths of corners and edges to reconstruct planar depth, which is unaffected by the position of planes and robust against noise. The edges and corners are generated through element-wise multiplication using the learned planar semantic labels. Hence, our framework is end-to-end trainable. Furthermore, to enhance the robustness and efficiency of the training process, we introduce a distance loss function that calculates the distance from points on layout edges to adjacent planes. Experimental results demonstrate that, compared to the current state-of-the-art methods, our method exhibits significant advantages.