Probabilistic Perspective-N-Lines for Indoor Camera Pose Estimation

Abstract

Indoor localization from a single RGB image via Perspective-n-Lines (PnL) remains a fundamental yet challenging problem in computer vision. To address this, the PnL-IOC algorithm was introduced, demonstrating remarkable performance by jointly estimating 3D correspondences of image outer corners (IOCs)--the intersection points between image borders and room layout boundaries--while optimizing camera pose within an iterative Gauss-Newton framework. However, existing learning-based indoor layout estimation methods often struggle to achieve high accuracy, resulting in unreliable line correspondences, which significantly limits the effectiveness of PnL-IOC. To overcome this limitation, we propose InPro-PnL, a probabilistic PnL layer for indoor camera pose estimation. Our approach introduces a novel framework that models the pose distribution on the SE(3) manifold, effectively extending the categorical Softmax function into the continuous domain. This is achieved by integrating a probabilistic PnL layer, where denoised 2D-3D correspondences serve as intermediate variables and are optimized by minimizing the KL divergence between the predicted and target pose distributions. Experimental results demonstrate that our method outperforms the geometric PnL-IOC approach and holds significant potential for further enhancing room layout estimation accuracy.