FastPlane: A Fully Convolutional Network for Real-time 3D Plane Segmentation

Abstract

3D plane segmentation is a challenging computer vision task that involves the detection of planar regions in a scene and the prediction of their spatial position. Recent methods have addressed this problem as a multi-task learning problem in which convolutional neural networks with separate branches predict the plane segmentation mask and the dense depth maps, or plane parameters. However, existing methods often rely on a multi-stage object detection framework for plane segmentation, resulting in slow inference speeds that hinder real-time applications. To address this issue, we propose FastPlane, a real-time 3D plane segmentation network based on a fully convolutional structure. FastPlane comprises a shared backbone, a fully convolutional plane segmentation network, and a lightweight depth estimation network. Given a single image as input, FastPlane outputs both the plane segmentation mask and dense depth maps. Extensive experiments on public datasets demonstrate the effectiveness and efficiency of FastPlane, outperforming existing methods in terms of plane segmentation and depth estimation metrics. Notably, FastPlane achieves an impressive inference speed of 45.0 FPS, making it suitable for real-time applications in robotics, augmented reality, and virtual reality.