Method and apparatus for acquiring a set of consistent image maps to represent the color of the surface of an object

[0002] This invention is related to copending U.S. patent application Ser. No. 09/631,043, filed on even date herewith, entitled “Method and Apparatus for Acquiring Normals Maps Consistent with Low Spatial Resolution Meshes” by Holly E. Rushmeier and Fausto Bernardini.

[0003] This invention relates generally to methods and apparatus for processing image data and, more particularly, relates to methods and apparatus for computing surface color from a plurality of sets of photometric data.

[0004] The creation of three-dimensional digital content by scanning real objects has become common practice in graphics applications for which visual quality is paramount, such as animation, e-commerce, and virtual museums. A significant amount of attention has been devoted to the problem of accurately capturing the geometry of scanned objects.

[0005] Three-dimensional scanners are used increasingly to capture digital models of objects for animation, virtual reality, and e-commerce applications for which the central concerns are efficient representation for interactivity and high visual quality. Most high-end 3D scanners sample the surface of the target object at a very high resolution. Hence, models created from the scanned data are often over-tesselated, and require significant simplification before they can be used for visualization or modeling.

[0006] In general, a variety of techniques can be used to capture digital models of physical objects, including CAT scans and structure from motion applied to video sequences. The following description has been restricted for convenience to techniques involving instruments that capture range images (in which each pixel value represents depth) and intensity images (in which each pixel is proportional to the incident light). A detailed summary of such methods can be found in G. Roth, “Building models from sensor data:an application shared by the computer vision and computer graphics community”, In Proc. of the NATO Workshop on the Confluence of Computer Vision and Computer Graphics, 2000.

[0007] The basic operations necessary to create a digital model from a series of captured images are as follows. After outliers are removed from the range images, they are in the form of individual height-field meshes. A first step aligns these meshes into a single global coordinate system. In high-end systems registration may be performed by accurate tracking. For instance, the scanner may be attached to a coordinate measurement machine that tracks its position and orientation with a high degree of accuracy. In less expensive systems an initial registration is found by scanning on a turntable, manual alignment, or approximate feature matching. The alignment is then refined automatically using techniques such as the Iterative Closest Point (ICP) algorithm of Besl and McKay.

[0008] After registration, scans do not form a single surface, but interpenetrate one another, due to acquisition errors primarily along the line-of-sight in each scan. To form a single surface, in a next step overlapping scans are averaged. In stitching/zippering methods this averaging is performed between pairs of overlapping meshes. In volumetric/occupancy grid methods line-of-sight errors are averaged by letting all scanned points contribute to a function of surface probability defined on a single volume grid. An advantage of volumetric methods is that all scans representing a surface point influence the final result, rather than simply a pair of scans.

[0009] In this step the scans are integrated into a single mesh. The integration may be performed by zippering/stitching, isosurface extraction from volumes, or interpolating mesh algorithms applied to error-corrected points.

[0010] If a texture map is to be used with the integrated mesh, in a next step the surface is parameterized with respect to a 2D coordinate system and texture coordinates are interpolated between mesh vertices. A simple parameterization is to treat each triangle separately and to pack all of the individual texture maps into a larger texture image. However, the use of mip-mapping in this case is limited since adjacent pixels in the texture may not correspond to adjacent points on the geometry. Another approach is to locate patches of geometry which are height fields that can be parameterized by projecting the patch onto a plane. Stitching methods use this approach by simply considering sections of the scanned height fields as patches. Other methods could be built on tiling methods developed for multiresolution analysis or interactive texture mapping.

[0011] Parallel to acquiring the geometry of the model, intensity images are captured to obtain information about the reflectance of the surface. Such images may be recorded with electronic or traditional cameras, or by using polychromatic laser technology. In a next step these images are aligned to the corresponding geometry. In some cases the image acquisition is decoupled from the geometry acquisition. The camera intrinsic and extrinsic parameters for the images are estimated by manual or automatic feature matching. The advantage is that acquisition modalities that cannot capture surface reflectance can be used for capturing geometry.

[0012] In most cases, however, the alignment is performed by calibration. Geometry and intensity are captured simultaneously from scanners with a measured transformation between sensing devices. The resolution of the intensity image may be the same as that of the range image or even higher.

[0013] One benefit of obtaining intensity and range images simultaneously is that the intensity information can be used in the registration process in the first step described above. Various approaches have been developed to use intensity images in registration. For example, it is known to use color as an additional coordinate in the ICP optimization. This avoids local minima in the solution in areas that have no geometric features, but have significant variations in the intensity. For models with pronounced geometric and intensity features, the method has proven to be very effective. A drawback is having to combine position and color data with different ranges and error characteristics. For subtle feature variations, these can cause one type of data to erroneously overwhelm the other.

[0014] It is also known to use intensity images to avoid the spatial search required by ICP. Intensity and intensity gradient images from approximately aligned scans are transformed into a common camera view. Locations of corresponding points on overlapping scans are inferred based on the difference between intensity values at a given pixel and the gradient at that pixel. This method works well only if the spatial variation of the gradient is small relative to errors in the alignment of the scans.

[0015] It is also known to employ a non-ICP method for using intensity images to refine an initial manual alignment. In this approach pairs of range images are aligned manually by marking three points on overlapping intensity images. The locations of the matching points are refined by searching their immediate neighborhoods with image cross-correlation. A least-squares optimization follows to determine a general 3D transformation that minimizes the distances between the point pairs.

[0016] Image registration techniques are also used for image mosaics in which only rotations or translations are considered.

[0017] After the intensity images are aligned to the geometry, illumination invariant maps are computed to estimate the surface reflectance. The number of scans versus the number of intensity images, as well as the resolution of the scans compared to the resolution of the images are considered at this stage. For a small number of scans and a large number of intensity images obtained under calibrated lighting conditions, a full Bidirectional Reflectance Distribution Function (BRDF) can be estimated.

[0018] If many scans are required to represent an object, and only a few high-resolution intensity images are captured per scan, photometric stereo techniques can be used to estimate Lambertian reflectance. Alternatively, if the range and intensity images have the same resolution, the geometry can be used to compute reflectance from a single image.

[0019] Problems arise when it is desired to employ a low complexity, low cost image capture system. In such a system variations in lighting (light source and colors) and positioning result in slight but discernible variations in photometric results from mesh to mesh. In that red, green and blue color maps can be computed from the photometric data, the variations in lighting and positioning, as well as inaccuracies and non-linearities in the color camera itself, can result in chromatic variations from map to map, which can further result in the production of visible artifacts.

[0020] Based on the foregoing, it can be readily appreciated that a need exists for improved methods to construct accurate digital models of multi-scanned objects, in particular digital models that exhibit high-quality color, and that eliminate visible artifacts that appear in the object model.

[0021] It is a first object and advantage of this invention to provide an improved method and system for constructing accurate digital models of multi-scanned objects, in particular digital models that exhibit high-quality and accurate surface color(s).

[0022] It is a further object and advantage of this invention to provide a method and system for computing surface colors from multiple image sets of photometric data, wherein the colors are made consistent across the multiple images for eliminating, when the images are combined, an appearance of objectionable visible artifacts.

[0023] The foregoing and other problems are overcome and the objects and advantages of the invention are realized by methods and apparatus in accordance with embodiments of this invention.

[0024] A method is disclosed for computing colors from multiple sets of images represented by photometric data, wherein the colors are made consistent between the images.

[0025] A method and system provide accurate and consistent surface color by processing a set of images of an object to eliminate illumination effects, projecting the images onto a lower resolution three dimensional model of the object, selecting a set of points on the surface of the model, listing all images that contain a particular point, and forming an equation for equating all the colors in each of the images in an area surrounding the particular point with a set of unknowns that are scaling factors for the color channels of each image. The set of all equations for all of the points are solved simultaneously to determine the optimum scaling factors for each image to make all of the color images consistent. The scaling factors are then applied to the images such that, when the images are combined, a seamless and artifact free description of the object color is produced.

[0026] In this invention colors are derived from photometric calculations, and are adjusted by a global color registration process.

[0027] The teachings of this invention preferably employ a low spatial resolution numerical representation of an object, such as one in the form of a triangular mesh, and combines information from the low spatial resolution mesh representation with, typically, higher resolution color images from the photometric system. The colors computed using the low spatial resolution mesh representation are made consistent from image to image such that, when the images are combined, the resultant image of the object exhibits substantially seamless, artifact-free surface color characteristics.

[0028] The above set forth and other features of the invention are made more apparent in the ensuing Detailed Description of the Invention when read in conjunction with the attached Drawings, wherein:

[0029] FIG. 1A is a block diagram of a computer system with graphics and 3D data acquisition capabilities that is suitable for practicing this invention;

[0030] FIG. 1B shows the graphics subsystem in greater detail;

[0031] FIG. 1C depicts the 3D data acquisition system in greater detail;

[0032] FIG. 1D is an example of the operation of the 3D scanner;

[0033] FIG. 1E shows the result of acquiring multiple stripes;

[0034] FIG. 1F shows, in the scan integration phase, the computation of a triangular mesh;

[0035] FIG. 1G shows a closeup view of the computed triangular mesh;

[0036] FIG. 2 illustrates various mathematical equations that are referred to in the description of the invention;

[0037] FIG. 3 is a diagram useful in understanding the geometry of light reflectance used in Eq. (1) of FIG. 2;

[0038] FIG. 4A depicts a two dimensional system used to evaluate the validity of assumptions made in computing surface normals, while FIG. 4B shows the results of applying the method for surface normals computation;

[0039] FIGS. 5A and 5B illustrate two typical spectra measured on a test object (sculpture), wherein intensity is plotted versus wavelength;

[0040] FIG. 6 is a simplified block diagram of a 3D data acquisition and reconstruction pipeline;

[0041] FIG. 7A is a logic flow diagram showing the adjustment of illumination invariant maps of surface color;

[0042] FIG. 7B shows an example of a geometry produced with a 3D scanner, with four (random) points selected;

[0043] FIG. 7C shows three surface color maps containing some of the random points of FIG. 7B, with image locations corresponding to the same random point being identified; and

[0044] FIG. 7D shows the three surface color maps of FIG. 7C after scaling factors are found an identified.

[0045] FIG. 1A is a block diagram of a computer system 100 with a graphics subsystem 110 and 3D data acquisition (photometric) subsystem 112 that is suitable for practicing this invention. A system bus 106 interconnects a CPU 120 (such as a Pentium™-type microprocessor) with the graphics subsystem 110 and a system memory 114. The acquisition subsystem 112 is preferably, but not necessarily, interfaced through an I/O subsystem 118.

[0046] It can be appreciated that the memory 114 can include or be loaded from a computer readable medium that embodies a computer program for constructing or obtaining a digital model of an object in accordance with the teachings herein, in particular a computer program for constructing consistent color maps.

[0047] More particularly, the computer program that is embodied on the computer readable medium has instructions for implementing a method for obtaining a digital model of an object. This method includes steps of (a) generating a low spatial resolution numerical representation of an object, such as a triangular mesh representation; (b) generating higher spatial resolution images of the object using the photometric subsystem 112, having a camera 112band a lighting source 112a;and (c) computing consistent color maps that when combined yield a visible artifact-free object representation.

[0048] FIG. 1B shows the graphics subsystem 110 in greater detail. A bus interface 110 a connects the system bus 106 to a graphics control processor 110band to a geometry subsystem 110c.A graphics control processor 110bcontrol bus 110galso connects to the geometry subsystem 110cand to a rasterizer 110d.A depth or z-buffer 110eand a frame buffer 110fare also typically coupled to the rasterizer 110d,and cooperate to render the object models for display.

[0049] FIG. 1C depicts in greater detail a presently preferred low complexity and low cost 3D data acquisition system 112. A tripod-mounted frame 112dsupports a plurality (5) of light sources 112la used for photometric stereo imaging. A suitable total width of the frame 112dand the supported light sources 112acan be 60 cm. Also supported by the frame 112dis a color camera 112band a scanner with, for example, six black and white cameras and a light striping flash. Reference in this regard can also be had to FIG. 3 of H. Rushmeier et al., “Acquiring Input for Rendering at Appropriate Levels of Detail: Digitizing a Pieta” in Rendering Techniques '98, Eds. G. Drettakis and N. Max, Springer, Wien, pp. 81-92, 1998.

[0050] In operation, and referring now as well to FIG. 1D, a light striping flash projects a set of stripes of light (structured light) 130 onto an object 132 (such as the vase depicted previously). The digital cameras 112C capture the shape of the stripes 130 from different angles. Digital image processing techniques are then used to compute the distance of each pixel on each stripe with respect to the sensor (cameras).

[0051] For all but the simplest objects 132, this process is repeated from multiple viewpoints, so that the sensor can “see” every portion of the object's surface. The multiple scans are then registered, or aligned, into a single coordinate frame 134. The measured points can be seen as an unorganized “point cloud”.

[0052] Next, in FIG. 1F, a scan integration phase forms a triangular mesh 138 that connects the measured points. The result of this operation is a 3D geometric model of the surface of the object 132. In subsequent operations, the mesh can be simplified, texture maps can be applied, or the triangular mesh can be converted to a curved surface representation (e.g., NURBS). FIG. 1G shows a closeup view of the computed triangular mesh of FIG. 1F.

[0053] The acquisition system 112 allows the capture of five images with each of the five lights turned on in sequence, registered to the geometry acquired by the shape camera. Each geometric mesh has a resolution of approximately 2 mm, and an estimated sub-millimeter accuracy. A set of normals and colors are computed from the five photometric images at a resolution of between 0.5 and 1 mm.

[0054] Variations in lighting and positioning result in slight but discernible variations in the photometric results from mesh to mesh. In the ensuing description a discussion is made of a technique to correct these variations.

[0055] Although not forming a portion of this invention per se, a description will first be provided of a technique to form consistent normals maps, followed by a description of the color processing aspects of this invention. Additional details concerning the normals maps calculations can be found in the above-referenced copending U.S. patent application Ser. No. 09/631,043, filed on even date herewith, entitled “Method and Apparatus for Acquiring Normals Maps Consistent with Low Spatial Resolution Meshes” by Holly E. Rushmeier and Fausto Bernardini.

[0056] In general, the normals maps are made consistent with the underlying mesh by locally adjusting the light source intensities used in the calculations, using data from the underlying mesh. By using a consistent global underlying mesh computed by registering and remeshing the underlying base meshes, the normals maps are consistent with one another, as well as with the global mesh.

[0057] In accordance with the teachings of this invention, corrected red, green and blue color maps are also computed from the photometric data. Chromatic variations from map to map are corrected with a global color registration that is analogous to a global geometric registration used to obtain the global base mesh.

[0058] The photometric stereo method for computing three dimensional shapes was originally developed by Woodham (B. K. P. Horn and M. J. Brooks, Shape from Shading, MIT Press, 1989). The essence of the method is to take a series of images from the same camera view, each with lighting from a different, known, direction. Assuming an approximately Lambertian bidirectional reflectance distribution function (BRDF) and identical light sources, the relative reflectance and normal at each pixel location can be computed from three images. Specifically, the set of equations shown in Equation (1) of FIG. 2 are solved, where ρ_p/π is the BRDF at pixel p, L_ois the light source radiance, L_r,i,pis the radiance reflected from the point visible through pixel p in image i, Δω is the solid angle subtended by the light source, n_pis the surface normal at p and l_iis the direction vector to the (infinitely distant) i-th light source. The constant α accounts for the camera scaling of the reflected radiance to a value from 0 to 255, after the images are adjusted if necessary for gamma values other than 1. The equations can be solved directly for (ρ_pL_oΔω/απ)n_p. Since the magnitude of n_pis one, from this result one can obtain the normal vector n_pand a value ρ_rel,p=ρ_pL_oΔω/απ, that is, the reflectance at pixel p relative to the other pixels in the image. FIG. 3. is useful in understanding the geometry of light reflectance used in Eq. (1).

[0059] Two difficulties with this approach are the presence of a highly specular component in the BRDF of the surface being measured, as well as shadows. Following the approach used by other researchers (e.g. see R. Epstein et al., Learning object representations from lighting variations, ECCV 96 International Workshop, pages 179-199, April 1996) one obtains extra images from additional light source positions. For each pixel location, the intensity values from the source images are sorted. Very high relative values are not used to exclude specular reflections, and low relative values are not used to exclude shadows. It is preferred to use relative value only, rather than color, to distinguish possible specular reflection, since there may be spectrally flat regions in the object, in this example a statue, for which such methods would fail. It is preferred to not use in the computations that follow those pixel locations where fewer than three images have pixel values that fall in the valid range. This is not, however, a limitation upon the practice of these teachings.

[0060] In the presently preferred embodiment a system of five lights 112ais used, providing two redundant reflectance measurements per pixel location. This number was used because it allowed a set of lights with a significant physical separation, while keeping to an overall manageable size for mounting on a tripod. FIG. 1C, described previously, illustrates the presently preferred embodiment of the image acquisition system 112. In other embodiments more or less lights could be used, and the spacing between lights may be different.

[0061] In a controlled environment, the term L_ocan be made the same for each source for all points on the scanned target by using controlled, identical, isotropic light sources.

[0062] The term Δω can be made constant for all points and all sources by placing the sources at a distance from the target that is large relative to the sizes of the light sources and the target.

[0063] In an inexpensive image capture system used in an environment such as a museum, any number of variations in light source radiances and position can occur:

[0064] The light sources 112amay not be identical. The radiance from individual bulbs of the same type will vary slightly, particularly for inexpensive off-the-shelf bulbs.

[0065] The light sources 112amay not be isotropic. In the instant case halogen bulbs were used for their brightness and ease of use. The emitted radiance is approximately uniform in a 30 degree cone around the direction that the source is aimed. There are small variations across the scanned area, and small disturbances caused slight changes in the source direction.

[0066] The light sources 112amay vary over time. The emittance radiance varies with the age of the bulb. It also varies with each use as a function of time since the bulb was turned on.

[0067] The light source 112aradiance varies with the electrical power level. A regulated power supply is required to maintain a consistent source level.

[0068] The distance from the light sources 112ato the area being scanned varies across the object. Requiring the distance to the light sources to be large relative to the scanned area would result in very small areas being scanned in each shot to obtain the spatial resolution required.

[0069] As well as overall strength, the light source 112aspectrum varies from bulb to bulb, with time and with level of power input.

[0070] All of these factors can be treated by high precision characterization and control of the lighting system, and by restricting the spatial extent of data acquired for each set of photometric images. However, such precision adds greatly to the cost of the system, and to the time required to acquire the data for the full object, such as a piece of sculpture. Instead, the inventors have dealt with these problems by exploiting the existence of the underlying, lower resolution mesh, and by performing a global registration of results.

[0071] Because of the foregoing variations from an ideal photometric setting, a method was developed to solve the system in Eq. (1) expressed with variable light source radiances and solid angles. It was assumed that the radius of the light sources 112ais the same for all the sources, and that d is small (i.e. less that 10 percent) relative to the distance r_i,pfrom surface point p to each light source i so that the solid angle Δω_ifor each source can be approximated by πd²/r^2i,p. One may denote the radiance of light source i at pixel location p with L_o,i,p. Eq. (1) can then be rewritten as Equation (2) in FIG. 2. All of the quantities in Eq. (2), except for α and d, vary from pixel to pixel across the image.

[0072] Using the system in accordance with these teachings, one obtains a geometric mesh at a spatial resolution of approximately 2 mm registered with each set of five photometric images obtained. The individual meshes are registered to one another using a two pass method that first uses a set of laser dots projected onto the object being scanned for an initial alignment, and then a form of the above referenced Iterative Closest Point (ICP) technique can be employed for a final more precise registration (see, e.g. R. Bergevin et al., Towards a general multi-view registration technique, IEEE Trans. on Pattern Analysis and Machine Intelligence, 18(5):540-547, May 1996). The meshes are combined into one single mesh using, preferably, a Ball-Pivoting Algorithm described by F. Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva and Gabriel Taubin, The Ball-Pivoting Algorithm for Surface Reconstruction, IEEE Trans. on Visualization and Computer Graphics, 5(4): 349-359, October-December 1999, which is incorporated by reference herein in its entirety, as is copending U.S. patent application Ser. No. 09/549,432, filed Apr. 14, 2000, entitled Method and Apparatus for Reconstructing a Surface Using a Ball-Pivoting Algorithm.

[0073] In general, the Ball-Pivoting Algorithm (BPA) computes a triangle mesh interpolating a given point cloud. Typically the points are surface samples acquired with multiple range scans of an object. The principle of the BPA is as follows: three points form a triangle if a ball of a user-specified radius touches them without containing any other point. Starting with a seed triangle, the ball pivots around an edge, that is, it revolves around the edge while keeping in contact with the edge's endpoints until it touches another point, forming another triangle. The process continues until all reachable edges have been tried, and then starts from another seed triangle, until all points have been considered.

[0074] For each set of photometric images there is obtained a consistent underlying mesh, and a camera transformation that gives the camera parameters for the photometric image capture in terms of a global coordinate system. Given these data, one can readily compute the following for each pixel of the photometric images:

[0075] The distances r_i,pfrom the visible surface point p to each of the light sources i. While there may be some error in this distance approximation, the error is well under one percent over the typical 0.75 meter distance from the surface of the object to the camera 112c.

[0076] The direction l_i,pfrom the visible surface point to each light source 112a.

[0077] An approximate surface normal n′_pat the visible point computed from the lower resolution surface.

[0078] One can immediately use the underlying values of l_i,pand r_i,pas good estimates for the values in Eq. (2). However, a need remains to find estimates for the relative values of L_o,i,pthat account for the temporal, directional and bulb to bulb variations of the light sources 112a.To do this, it is preferred to equate the relative reflected radiances from the underlying surface illuminated by ideal sources, to the relative image radiances in the neighborhood of the pixel.

[0079] Consider ideal light sources with uniform radiances L_uin the same positions as the physical light sources 112a.If these ideal sources illuminated the underlying base surface coated with a Lambertian reflectance of l, the reflected radiances would be as shown in Equation (3) of FIG. 2. Eq. (3) may be considered to enable the generation of a synthetic image defining light reflected from the low resolution mesh for each grey scale image corresponding to a different lighting condition. The entire synthetic image need not be calculated at once, as the synthetic image can be calculated pixel-by-pixel, on the fly, as it is needed.

[0080] For the i-th photometric image, the reflected radiance in the neighborhood around pixel p is shown in Equation (4) of FIG. 2, where a consideration is made of a neighborhood of pixels Q that approximately represents the area on the surface of a disk with a radius equal to the base geometry resolution. In the preferred, but not limiting, system the base resolution is 2 mm, corresponding to a radius of 5 pixels. Notice that from the images, L_r,i,pis not obtained directly, but instead the quantity αL_r,i,pas recorded by the camera 112c.Reference in this regard may be had to FIG. 7C, discussed below, which shows blurred images produced by an averaging process.

[0081] The values αL_r,i,pare the result of recording the reflection from the true physical sources in accordance with Equation (5) in FIG. 2.

[0082] An assumption is made that although ρ_p, L_o,i,p, r_i,p, and l_i,pvary pixel by pixel over the entire image, they vary relatively little in the small neighborhood around pixel p. An assumption is also made that the average over Q of the dot products l_i,q·n_qis approximately equal to the dot product of l_pand the underlying surface normal n′_p. Note that these assumptions are the same basic assumptions made in all photometric stereo calculations, with the exception that the assumptions are made in a (small) region of pixels rather than for just the surface represented by a single pixel. With these assumptions one has the relationship shown in Equation (6) of FIG. 2. Equation (7) then shows the ratio of the radiance reflected from the underlying (low resolution) surface given in Eq. (3) to the average recorded radiance in the pixel neighborhood given in Eq. (6). Eq. (7) can be used to express the unknown source radiances in terms of quantities that can be computed, or which are the same for all light source directions, as shown in Equation (8) of FIG. 2. Using Eq. (8) in the right hand side of Eq. (2), and simplifying terms, results in Equation (9) of FIG. 2.

[0083] The values of α have been retained in Eq. (9) since the quantities are the values actually available from the recorded images. Eq. (9) can be solved for a vector in the direction of the surface normal, and normalization gives the result n_p. The effect of the surface reflectance ρ_pis in both the individual pixel values and the averaged pixel values. As a result the reflectance term is canceled out of the equation, and the length of the vector in the direction of the surface normal no longer has a useful physical interpretation.

[0084] The accuracy of the resulting normals depends on the validity of the assumptions made in the approximation in Eq. (6). The small size of the neighborhood used for averaging relative to the distance to the camera 112c(e.g., 2 mm versus 750 mm in this example), justifies the assumptions that L_o,i,p, r_i,p, and l_i,pare uniform. The reflectance is also reasonably assumed to be uniform in a small neighborhood. Exceptions to this are abrupt changes due to, for example, a change in surface coating of the object being scanned. However, it has been found that changes in reflectance cause a consistent change in all of the photometric images and can be detected. That is, a sharp increase in reflectance results in an increase in reflected radiance for all light source directions. The main assumption is the equivalence of the average of the dot product of the light direction and the normals on the high resolution surface and the dot product of the light direction with the normal of the underlying lower resolution surface.

[0085] FIG. 4A depicts the two dimensional system used to evaluate the validity of assumptions made in computing the normals, while FIG. 4B shows the results of applying the method for normals computation.

[0086] To evaluate the assumption of equating the average and single dot products consider the simple two dimensional system shown in FIG. 4A. Consider a horizontal surface, of width 2, that is approximated by the low resolution surface connecting the points (0,0) and (2,0). The surface is illuminated by two lights: l_oin direction (0,1), and l₁in direction (−(sqrt 2)/2, (sqrt 2)/2). One can then apply the foregoing calculation method to a higher resolution surface that has two facets formed by a displacement h in the center of the surface. By examining the accuracy of the result one obtains the normal of the left facet as a function of h.

[0087] Specifically, the following quantities are computed:

[0088] The underlying surface normal n′=(0,1) for all h.

[0089] The true facets normals $n_0=\frac{\left(-h,1\right)}{\sqrt{1+h^2}}\mathrm{and}{n}_1=\frac{\left(h,1\right)}{\sqrt{1+h^2}}$

[0090] The reflected light for each facet for each light, L_r,i,j=l_i·n_j; and

[0091] The approximate surface normal from Eq. (9), using

[0092] {overscore (L)}_r,i=L_r,i,0+L_r,i,1

[0093] and

[0094] and setting the constant factor to one.

[0095] {overscore (L)}_r,i=l_i·n′

[0096] The results of these calculations are shown in FIG. 4B. The plot compares the angle between the estimated and true normal from using the underlying surface normal, and from using the approximate normal from the photometric method. The non-linearity of the dot products with normal vectors accounts for the errors in the photometric method, and for the asymmetry between negative and positive values for h. Although the error in the photometric method increases with the absolute value of h, the photometric estimate of the normal is in all the cases a significant improvement over using the underlying normal.

[0097] The above calculations do not account for interreflections that would occur when h is negative. In the exemplary application where the object was a statue, the measurement was made on a surface with low reflectance and, as a result, there is relatively little indirect illumination from interreflections. For surfaces with high reflectance and steep slopes in the high resolution surface there would be additional error in the computed normals.

[0098] Having thus described the presently preferred techniques for computing the underlying geometry of the object, as well as a presently preferred technique for computing surface normals, a description is now made of the presently preferred techniques for performing the color calculations in accordance with the teachings of this invention.

[0099] Given an estimate of the surface normal at each pixel, one may compute versions of the photometric input images with the shading effects due to the directional lighting removed. By taking a weighted average of these corrected images, one obtains an estimate of reflectances in the red, green and blue color channels. Unlike scans with a polychromatic laser, one does not obtain the absolute value of the reflectance in each channel, or a precise spectral definition of each channel. However, using some additional spot measurements of spectral reflectance one can adjust the corrected images to approximate absolute reflectance in defined spectral ranges. The adjustment of specific corrected images by the spot measurements is propagated over all images by a global color registration process.

[0100] Denoting the wavelength of light as λ and using the subscript C to denote the color channel that is either red, green or blue, the recorded color value αL_r,i,p,Cof each pixel p in each photometric image i is given by Equation (10) shown in FIG. 2, where S_C(λ) is the spectral sensitivity of the color camera 112bin channel C.

[0101] Using the normals as described above, a next step computes an approximate relative reflectance ρ_relfor each channel using Equation (11) shown in FIG. 2. This reflectance is relative to other pixels in the same channel in the same image. As in the case of the calculation of normals, it is preferred to not use high or low pixel values to avoid using data in areas of specular reflection and shadow. However, unlike the normals calculations, one does not meed a pixel estimate from three images to estimate the relative reflectance. Since one can afford to use less of the image, it is possible to avoid including the effects of directionally varying light source intensity by only using pixels that fall in a small cone around the direction in which the light source 112ais aimed.

[0102] The five relative reflectance images computed with Eq. 11 are adjusted relative to one another to compensate for variations in light source strength. The pixels for which all five of the images have non-zero values are found, and the averages for these overlapping pixels are computed.

[0103] The overall level of the images is adjusted by making all of the average values equal.

[0104] Once the images levels are adjusted, the five images are combined with a weighted average to produce a single corrected RGB image for the position of camera 112c.The weight assigned to each pixel in each image increases with distance to the nearest black edge in the image. This weighting is preferred in order to avoid small but sudden changes at edges where each photometric image stops contributing to the combined average.

[0105] Clearly the relative reflectances computed from Eq. 11 are not scaled to the actual percentage of visible light that would be reflected. The images also still contain the effects of the light source spectra and the color camera 112bspectral sensitivity. Furthermore, the red, green and blue channels are not defined in terms of wavelengths. To deal with such factors, separate spot measurements can be made of spectral reflectance using a Colortron color ruler (Light Source, Inc. Colortron (http://www.ls.com/colortron.html)). Two typical spectra measured on the test object (sculpture) are shown in FIGS. 5A and 5B. The spectral reflectances are simple functions in this case, and one can readily define spectral bounds for red, green, and blue for which the spectral reflectances vary little from the average for the range. That is, one can define λ_C,minand λ_C,maxand compute ρ_ave,Cin accordance with Equation (12) shown in FIG. 2.

[0106] By taking the ratio of ρ_ave,Cand ρ_rel,Cat the location in the image where the spot measurement was taken, one is enabled to estimate the scaling for light source 112aand color camera 112bspectral effects in the image. Using the scaling for each channel in the image, one can obtain at least approximate values for the absolute values of reflectance in clearly defined spectral ranges.

[0107] Once at least a few images have been adjusted to approximate reflectances in the three color channels, one can adjust all of the images for which ρ_rel,Chas been computed by requiring the colors in matching points in overlapping images to be the same. As in the case of geometric registration, it is generally inadequate to simply perform these adjustments pairwise, since small errors accumulate. The most significant errors in the relative color images have been found to arise from biases in the captured images that occur even when the automatic color compensation of the camera 112bis turned off. Some captured images are overall too blue, some too red, etc. Additional bias in the image may occur as the result of variations in the light bulbs 112aover time. To minimize the effect of these errors, it is preferred to perform a simultaneous, global color registration.

[0108] Forming equations for the color registration can be facilitated by the availability of the points used in the initial geometric alignment. Alternatively, and as will be described below, a set of randomly (truly random or pseudo-random) distributed points can be used as well. One result of the initial alignment is a list of points in overlapping images that represent the same location on the base geometry. One may then compute a color for each of these points by taking the average of a small neighborhood of pixels in the corrected image. A neighborhood is preferred for use rather than a point sample since very small alignment errors (e.g. of the size of one pixel projected onto the surface) can produce a very large color error when there are abrupt changes in colors between pixels. Let β_C,kbe the level adjustment for the k^thcorrected photometric image in the C channel. Each matching pair of points gives an equation of the form shown in Equation (13) of FIG. 2 for images m and n. There are far many more matching points than corrected images, so a least squares solution is preferred in order to compute the values of β. In Eq. (13) no level has been set for the system. While one could set the value of β to one for the image (or images) that have been adjusted with the Colortron measurements, it is desired to have the errors in the results to be distributed evenly over the model, rather than being skewed by distance from a particular image. To accomplish this, the Equation (14) shown in FIG. 2 is added to the set of equations, where N is the total number of images. Using N on the right hand side of Eq. (14) produces values of β around 1 (e.g., from about 0.6 to about 1.4). After all of the images have been adjusted to one another, global correction values for red, green and blue can be made to match the Colortron measurements.

[0109] FIG. 6 is a simplified block diagram of a 3D data acquisition and reconstruction pipeline in accordance with these teachings, which may also be viewed as a logic flow diagram of the method disclosed herein. The image acquisition system 112 captures partial shape information in block or step 140. These scans are aligned together at step 142 to form an unorganized set of point measurements. In step 144 a triangular mesh that connects these points into a triangular mesh is computed. A presently preferred for forming the triangulated surface is the above-referenced Ball-Pivoting Algorithm described by F. Bernardini, Joshua Mittleman, Holly Rushmeier, Claudio Silva and Gabriel Taubin, The Ball-Pivoting Algorithm for Surface Reconstruction, IEEE Trans. on Visualization and Computer Graphics, 5(4): 349-359, October-December 1999, incorporated by reference herein in its entirety, as well as is the copending U.S. patent application Ser. No. 09/549,432, filed Apr. 14, 2000, entitled Method and Apparatus for Reconstructing a Surface Using a Ball-Pivoting Algorithm. However, the teachings of this invention are not limited to only this particular technique for forming the triangulated surface. Further processing can occur at step 146 to transform the computed surface into one of a number of representations. In step 148 the photometric images, i.e., the images obtained from the color camera 112bare obtained with a higher spatial resolution than the mesh obtained in step 144. The mesh, or a derivative model of same produced in optional step 146, are combined to compute normals maps for representing the high spatial resolution shape of the object. Multiple normals maps are then combined to represent the complete surface of the object in step 150, and in step 152 the normals maps are used to compute illumination invariant maps of the surface color using the methods discussed above. The presently preferred techniques for computing the normals maps and for computing the illumination invariant maps have been described above, and is further described in the above-referenced copending U.S. patent application Ser. No. 09/631,043, filed on even date herewith, entitled “Method and Apparatus for Acquiring Normals Maps Consistent with Low Spatial Resolution Meshes” by Holly E. Rushmeier and Fausto Bernardini.

[0110] “Illumination invariant maps” refers herein to surface maps that represent an intrinsic property of the surface, generally the reflectance or albedo. Illumination invariant maps are images of an object in which the effects of any particular illumination conditions have been removed.

[0111] In step 154 the colors in each of the surface color maps covering the surface of the object are adjusted so that the colors in overlapping images are the same. Multiple surface color maps are then combined in step 156 to represent the computed surface of one object. The combination of color maps can be performed in accordance with the procedure described by Fausto Bernardini, Ioana M. Martin and Holly Rushmeier, High-Quality Texture Synthesis from Multiple Scans, IBM Report RC21656, incorporated by reference herein in its entirety, as is copending U.S. patent application Ser. No. 09/603,928, filed June 27, 2000, “High Quality Texture Reconstruction from Multiple Scans”, Fausto Bernardini, Ioana M. Martin and Holly Rushmeier.

[0112] The teachings of this invention reside most particularly in step 154, such that when step 156 is performed there are no visible artifacts along the seams between scans.

[0113] FIG. 7A is a logic flow diagram showing the adjustment of illumination invariant maps of surface color. In step 200 a plurality of points are selected, preferably random points that are selected automatically or by an operator so as to have some operator-specified approximate interpoint spacing. The points are preferably distributed over the entire surface of the object. In step 210 each of the surface color maps obtained with camera 112bare projected onto the object geometry and each image point that corresponds to one of the randomly selected points is identified. In step 220 the average color of pixels around each image location is computed, where the size of the area that is averaged preferably corresponds to the amount of possible misregistration between images. In step 230 an equation is formed for each random point to equate the colors of all surface color maps containing the point, with the unknowns in the equations being scalar factors for adjusting each color channel in each image (see Eq. (13) in FIG. 2). In step 240 all of the equations for the image color factors are solved simultaneously, preferably by computing a least squares solution (see Eq. (14) in FIG. 2). In step 250 the scaling factors thus obtained are applied to the surface color maps to produce a globally consistent set of maps (see, for example, the discussion of above of the use of the Colortron measurements).

[0114] FIG. 7B shows an example of a geometry produced with a 3D scanner, with four (random) points selected, while FIG. 7C shows three surface color maps containing some of the random points of FIG. 7B, with image locations corresponding to the same random point being identified. FIG. 7D shows the three surface color maps of FIG. 7C after scaling factors are found an identified.

[0115] While described above in the context of specific embodiments, it should be appreciated that these teachings apply to any 3D geometric model obtained by any suitable means, and to any set of color images acquired and aligned to the model by any suitable means. These teachings also apply to any color images, including images that record grey scale albedo only, red/green/blue images, and to images containing any number of channels encoding reflectance in any number of spectral bands.

[0116] Thus, while the invention has been particularly shown and described with respect to preferred embodiments thereof, it will be understood by those skilled in the art that changes in form and details may be made therein without departing from the scope and spirit of the invention.