Fast 3D Layout Estimation through Instance-Level Parameter Learning in Lightweight Network

Indoor layout estimation serves as a fundamental task in the realm of scene understanding, aiming to infer the overall spatial structure of the indoor scene from an input image. Many recent studies focus on learning high-resolution maps representing pixel-level planar depth parameters for layout estimation. However, the high model complexity and the error-prone parameter aggregation process make these methods challenging to deploy on mobile devices. In this study, we propose a novel framework that directly learns low-resolution instance-level planar depth parameters with a lightweight model, aiming to reduce computational costs and accelerate layout estimation. We further employ knowledge distillation of spatial attention to enhance the accuracy of layout estimation. We train a more complex teacher network, and impart the spatial attention learned by the teacher network to the lightweight student network, thereby improving its performance. Experiments demonstrate that the proposed framework achieves fast computational speed while maintaining a high level of accuracy, indicating its superiority in addressing both speed and accuracy challenges.