Point Cloud Noise and Outlier Removal for Image-Based 3D Reconstruction

Point sets generated by image-based 3D reconstruction techniques are often much noisier than those obtained using active techniques like laser scanning. Therefore, they pose greater challenges to the subsequent surface reconstruction (meshing) stage. We present a simple and effective method for removing noise and outliers from such point sets. Our algorithm uses the input images and corresponding depth maps to remove pixels which are geometrically or photometrically inconsistent with the colored surface implied by the input. This allows standard surface reconstruction methods (such as Poisson surface reconstruction) to perform less smoothing and thus achieve higher quality surfaces with more features. Our algorithm is efficient, easy to implement, and robust to varying amounts of noise. We demonstrate the benefits of our algorithm in combination with a variety of state-of-the-art depth and surface reconstruction methods.

1. Introduction

Acquiring the 3D geometry of real-world objects is a long standing topic in computer vision and graphics research, with many practical applications, ranging from scanning small objects up to modeling complete cities. Consequently, there is an abundance of 3D reconstruction techniques, which can be roughly classified into active techniques [3] relying on illuminating the scene (e.g. by lasers or structured light), and passive techniques that analyze a multitude of images of the scene and are thus referred to as multi-view stereo or pho-togrammetry methods [33]. The latter, image-based methods have a number of benefits compared to active techniques. One main advantage is that the capture process is simple and cheap, only requiring standard imaging hardware like consumer digital cameras. Additionally, image-based methods provide color information of the scene and offer high resolution scanning thanks to the advances in image sensors. A popular approach to image-based 3D reconstruction is to first compute camera poses and then estimate per-view depth maps by finding corresponding pixels between views and triangulating depth [14]. All pixels are then projected into 3D space to obtain a point cloud, from which a surface mesh is extracted using point cloud meshing techniques [2].