3D S.O.M.—A commercial software solution to 3D scanning

Abstract

This paper describes the novel features of a commercial software-only solution to 3D scanning—the 3D Software Object Modeler. Our work is motivated by the desire to produce a low-cost, portable 3D scanning system based on hand-held digital photographs. We describe the novel techniques we have employed to achieve a robust software-based system in the areas of camera calibration, surface generation, and texture extraction.

Introduction

Conventionally generating photo realistic 3D models has been an expensive and time-consuming operation. To automate this process, a wide variety of 3D scanning devices is available, for example, using laser-based systems (e.g., Cyberware body scanners), structured light systems as well as passive systems (e.g., Geometrix LightScribe). Some of these high end devices can recover full-colour texture as well as geometry data. One particular problem with laser-based and structured light systems is that they do not work well with shiny reflective objects and it is often necessary to coat the surface with a non-reflective layer (e.g., dust with chalk) which can result in unsatisfactory texture data. Existing passive systems often require an expensive turntable component that limits the size of the object being scanned. These systems often require professional lighting (e.g., diffuse lighting) to produce satisfactory textures.

Our work is motivated by the desire to produce a low-cost, portable 3D scanning system based on hand-held digital photographs. The only hardware requirements for our system are a camera, a black and white calibration pattern printed on any standard printer, a uniform backdrop and a PC. We designed the system to be as easy to use as possible although a certain amount of skill is required in taking the photos and editing any artifacts in the output model. The target model quality is “reasonable looking” textured 3D models suitable for 3D on the web. In general the more expensive hardware systems are capable of producing very high quality 3D models. However, we have found that our low-cost software solution can often produce comparable results, especially when comparing models suited to bandwidth limited e-commerce applications. In this context the appearance (and file size) of the textured model is much more important than the geometric accuracy.

On a system level our approach is similar to that of Niem [1]. Like Niem we use a calibration object to ensure accurate camera parameter estimation for arbitrary objects. Similarly to ensure the system can handle untextured or reflective objects and uncontrolled lighting we also use a “shape from silhouettes” approach rather than rely on stereo feature matching [2] or colour consistency [3]. In contrast to Niem, our system benefits from a new robust calibration object, a novel batch technique for exact computation of the “visual hull” and a novel texture blending scheme.

An alternative approach to using a calibration object is to use “Structure From Motion” (SFM) techniques (e.g., Fitzgibbon et al. [4]). This has proved to be very effective in recovering camera motion from tracked feature trajectories. However in our application, we wish to be able to handle objects with little or no texture detail making feature tracking difficult. For feature tracking to be robust, video input is generally required limiting the resulting textured models to video quality. Furthermore SFM can fail for scenes consisting of highly reflective (e.g., metallic) objects.

Currently, we extract a simple texture map to represent the appearance of the 3D surface. Structured or controlled lighting techniques such as Bernardini [5] allow the recovery of albedo and normal maps for more accurate re-rendering of an object in new lighting conditions. However, we chose an uncontrolled lighting solution to reduce the system cost and simplify image acquisition.

An alternative to explicitly modeling the surface reflectance properties is to use an image based rendering approach. MacMillan describes such a technique where images of an object are acquired from multiple viewpoints and with multiple lighting conditions [6]. This approach can produce very high quality results but is not well suited for producing small sized models suitable for low bandwidth applications. The image acquisition step is also expensive and potentially slow. Proprietary image-based solutions such as QTVR suffer from large file size and also require large numbers of input images to ensure smooth viewpoint changes when rotating the model.