A Review of Geometry Recovery from a Single Image Focusing on Curved Object Reconstruction

Abstract

Single view reconstruction approaches infer the structure of 3D objects or scenes from 2D images. This is an inherently ill-posed problem. An abundance of reconstruction approaches has been proposed in the literature, which can be characterized by the additional assumptions they impose to make the reconstruction feasible. These assumptions are either formulated by restrictions on the reconstructable object domain, by geometric or learned shape priors or by requiring user input. In this chapter, we examine a representative set of state-of-the-art reconstruction approaches, which are applicable to real-world images. We classify the approaches according to their reconstruction objective and compare them based on a variety of important criteria. Finally, we show experimental comparisons for five curved object reconstruction approaches.

Learning 3D Shape Completion Under Weak Supervision

“…Single-View 3D Reconstruction: Single-view 3D reconstruction has received considerable attention over the last years; we refer to (Oswald et al, 2013) for an overview and focus on recent deep learning approaches, instead. Following Tulsiani et al (2018), these can be categorized by the level of supervision.…”

Pixel2point: 3D Object Reconstruction From a Single Image Using CNN and Initial Sphere

“…The challenge appears when a single input image is just available for the reconstruction process. Many approaches were proposed with restrictions and special assumptions on the input image to predict 3D geometry [19]. Single-view 3D reconstruction is a hard problem and it mainly depends on the available information and the imposed assumptions on the target object.…”

“…Single-view 3D reconstruction is a hard problem and it mainly depends on the available information and the imposed assumptions on the target object. This information or cues provide prior knowledge that helps in generating 3D shapes with plausible precision [19].…”

3D Room Geometry Reconstruction Using Audio-Visual Sensors

“…Computer vision techniques using visual sensors such as cameras and RGBD sensors have played an important role in geometry reconstruction and various approaches have been proposed in the 3D vision community. Recovering geometric information from a single photograph relies on learnt cues such as silhouettes, shading and texture [26]. Recent progress in deep learning has accelerated this field [38,33], but this approach still works in very limited environments and relies on large corpuses of training data for similar scenes.…”