The field of 3D acquisition has emerged as an important source of both challenges and applications for point-based techniques. The high detail and ease of acquisition of 3D models of real-world objects have driven their acceptance in computer graphics productions, but the size and density of these datasets have required the development of efficient representations and processing algorithms.
The relations between 3D scanners and points are:
1) Size: 3D scans can be large and detailed, with models in the range of 106 to 109 samples becoming commonplace.
2) Noise: All scans have noise, and depending on the particular scanner design this noise can be on the order of the sample spacing.
3) Sparseness: Some 3D acquisition techniques, especially passive methods, can only return reliable geometry at a few locations on the surface.
A basic 3D model acquisition pipeline may include the following stages:
1) 3D scanning
2) Registration
3) Merging
4) View planning
5) Postprocessing
Two possible representations for points:
1) represent them as unorganized point clouds.
2) represent them using range p_w_picpaths.
Triangulation-based 3D scanners
They have the common element of possessing two (or more) viewpoints on the object being scanned, and determing correspondences between rays from those viewpoints.
The advantages:
1) Flexible working volume
2) Low complexity and cost
3) Precision scales with camera resolution
The disadvantages:
1) Two-line-of-sight problem
2) Sensitivity to shiny or translucent object
3) Difficulties for large scenes
Passive stereo
Constructing an effective 3D scanner based on stereo is essentially synonymous with solving the correspondence problem between p_w_picpaths taken with two cameras.
Characteristics: They differ from most of the other techniques considered in this section in that most of the characteristics of the datasets they produce, including density, accuracy, and presence of outliers, depend almost exclusively on the object or scene being acquired, rather than on the camera and setup of the scanner itself. In many applications the output of stereo is best suited to the unorganized point cloud representation. As scene contrast decreases, of course, so do data density and accuracy.
Active stereo
They permit additional light to be introduced into the scene, significant gains in robustness, accuracy, and data density may be obtained.
Characteristics: The data returned by active stereo systems can be of high quality, and it is not uncommon to have noise in depth less than one-tenth of the sample spacing. The scanning also tends to be robust, with few outlier points (with some exceptions near depth discontinuities). They return better data, at the cost of the requirement to introduce light into the scene and the inability to acquire the color of objects simultaneously with their shape.
Structured light
Conceptually, structured-light system may be thought of as active stereo systems in which the second camera is eliminated and the source of illumination is itself considered to be one of the viewports on the scene.
Characteristics: Datasets tend to be accurate, dense, and large. Outliers are few, and are mostly caused by depth discontinuities or specular reflections. They have “texture embossing” phenomena. In addition, they often suffer from misssing data in dark regions of the surface where the projected light patterns cannot be detected.
Light stripe
A simple way of thinking about light-stripe scanners is that they are structured-light scanners with a pattern that illuminates only one strip per frame.
Characteristics: The data produced by them are among the highest in quality, and the use of laser light results in little missing data (other than due to occlusion of the camera or laser) and few outliers. Datasets are dense, and are most naturally organized as range p_w_picpaths.
Multiview triangulation and structure from motion
Any of the above techniques can be combined with additional calibrated cameras, leading to multiview techniques.
Characteristics: Both structure from motion and multiview stereo systems return sparse, though accurate, depth estimates. In addition, SfM can return points from all around an object, rather than just from one direction. Unorganized point clouds are therefore the appropriate representation.
Other 3D scanning technologies
1) Pulsed and modulated time of flight
2) Shape from silhouettes and other space-carving methods
3) Shape from shading and photometric stereo
4) Shape from focus and defocus
Scan alignment
Why need it?
Since most types of scanners return data from only one part of the object at a time, it is necessary to obtain multiple scans and align them together.
Some alignment methods:
1) Initial alignment
Manual alignment + automated alignment
The general technique of automated initial alignment is as follows:
a) Find distinctive feature points on each scan
b) Compute local shape descriptors characterizing the surface geometry around the features.
c) Propose correspondences between features by matching similar descriptors.
d) Compute the rigid-body transform that best aligns corresponding features.
e) Verify that the proposed alignment is reasonable.
2) Iterative pairwise registration using ICP
3) Global registration
