Face Recognition using Computer-Generated Database

Transcription

1 Face Recognition using Computer-Generated Database Won-Sook LEE Scool of Information Tecnology and Engineering University of Ottawa, 800 King Edward Ave Ottawa, Ontario, K1N 6N5, Canada Tel: (1-613) ext 6227 Fax: (1-613) Kyung-A SOHN Human-Computer Interaction lab Samsung Advanced Institute of Tecnology San 14-1, Nongseo-ri, Gieung-eup Yongin-si, Gyeonggi-do, Korea Tel: (82-31) Fax: (82-31) Abstract Many face database and recognition systems ave been constructed in specially designed studios wit various illuminations, poses, and expressions. However, none of tese databases yet satisfies a large variation of poses and illuminations tat permit te study of systematic 3D uman face information, wic results in an unsatisfactory success rate. It is caused primarily by te difficulty of data collection of facial images to satisfy te large variation of poses and illuminations to fully represent te 3D caracteristics of uman faces. We present ow computer-generated database can be used for face recognition experiment focused on multi-view face recognition/descriptor. We sow multi-view face data collection using rendering of 3D models. We also illustrate our approac ow to build a face descriptor containing 3D information of uman face using multi-view concepts. Tis multi-view face recognition descriptor is a 3D descriptor using te concept ow muc powerful a view influences over nearby views, so called as quasi-view size. 1. Introduction A variety of tecniques borrowed from computer vision ave been brougt to bear on te computer grapics problems inerent in syntesis and animation. Neverteless, tese areas of computer grapics, toug relatively young compared to computer vision, ave made dramatic progress in recent decades in teir central tecnology, tat of rendering [4], and ave also seen an entry into virtual uman and virtual environment simulation [21][1][19], wic results tat computer grapics are now providing tecnology to oter areas, suc as computer vision and medical imaging. Te fusion and synergy of computer grapics (virtual) and computer vision (real) as given rise to intense activity. By te late 90 s, face/body modeling and animation on one and, and face/uman detection and recognition in te oter, were separate endeavors, but ad started to merge teir tecnologies. For instance, te quality of uman face modeling as been enanced using computer vision tecnology and te statistical approac to face recognition tecnology, wereas traditional computer grapics relied on conventional geometrical and deformation approaces. Supplementary views to provide complete 3D information ave been critical, but te statistical approac based on a large database of conventional sapes as sown potential for determining individual face sapes wit limited information (views). Conversely, computer vision researc on face/uman detection/recognition as benefited from computer animation tecnology for te extraction and description of moving sapes using 3D uman anatomy. For te example of faces, see [21][2][18][3] and for bodies see [15]. Wen we closely examine te classic face recognition problem, one of te most significant extrinsic factors in variety of different people s faces is te pose, or te viewing direction at wic te face is seen and illuminated. Tis is te basic reason tat te typical face description algoritms based on a single view, especially frontal view, wic as a fundamental limitation in its performance in a real situation. Some approaces suc as novel view generation wit given limited view [13] were proposed but reveal limitations in obtaining a ig success rate. Tis is because te face is a 3D-object and it cannot be reconstructed from

2 2D information. Tere are oter approaces to describe te 3D caracteristics of uman face for recognition purpose suc as building a 3D model te subject. Some approaces using 3D morpable models [2] obtained using a laser scanner sow quite good result but due to teir eavy computational costs, it is still unavailable publicly. A 3D model reconstruction wit stereoscopic camera is also used to retrieve te face identity [10], but te successful result as not come out yet Our Approac Here we discuss ow a computer generated 2D database can be used for face recognition. Te 2D database is derived from te 3D database of sapes captured from te real world. Te raw date from capture equipment is usually neiter directly usable nor error-free; all te more reason for te great investment we plan in 3D data collection, especially te texture terms, based on realistic rendering. Our approac not only economizes data collection time and reduces te expense of a dedicated studio set up, but it also avoids over-fitting to any pysically-fixed environment by varying te virtual environment. Our approac for describing a face is by te use of a 2D-viewpoints mosaic [16] and we sow wic criteria are used to extend one-view to multi-views. Here, te multi-view face descriptor contains te 3D information, in a compact form, of all views of orizontal rotation [ ], and vertical rotation [ ]. Te database for tis multi-view face descriptor must cover a wide pose variation wit a dense pose distribution. Many kinds of face databases already exist but none of tem provides a satisfactory view range. Instead of building a complicated and ig-cost studio for data collection, we used a simple metod wit 3D mes models. In Section 2, we analyze problems wit existing face databases, wit Section 3 devoted to 3D capturing tecnologies. Section 4 details te face recognition database construction using a rendered face image database wit captured 3D facial models and includes a description of te multi-view face data collection procedure. Section 5 sows te experimental results using a 3D face descriptor, using systematic extension of a 2D face descriptor to multi-view one using quasiview concepts. Section 6 concludes te paper, wit a review of future work. 2. Face databases built in studio 2.2. Existing face databases Many 2D face images ave been obtained in a studio built entirely for tat specific purpose. Some selected examples of face databases intended to cover poses and illumination are KISA [23], YALE [6], and CMU [22]. Cameras and ligts are placed in different positions to simulate pose and ligt variances, tis quickly produces satisfactory images. Neverteless, certain constraints exist due to te studio setup: te total number of poses equals to te actual number of cameras, and ence limited number of pose variance suc as 7 to 13 pose images in tese examples are far too small to perform experiments for multi-view face recognition Ideal pose independent face recognition Ideal posed face recognition means to cover te wide cases in terms of people and environment, in oter words intrinsic and extrinsic parameters. We focus on extrinsic parameters in tis paper. Two main extrinsic parameters are pose and illumination variations. For te pose, te training set must be complete in every view covering any view of orizontal rotation [-90 o 90 o ] and vertical rotation [- 30 o 30 o ] ~ [-60 o 60 o ] as a view-mosaic in Figure 1 wile registration can be made wit only selected views. A omogenous background is more suitable, and permits te automatic segmentation using cromakeying, so tat te background can be replaced later. For te illumination, many ligts, for example 64 ligts, must be used to control various illumination environments. Figure 1: Ideal face recognition region in terms of poses. 3. Capturing systems We discuss about tecnologies to capture te 3D positional data of a uman face. Te tree main metodologies for capturing sape are: laser scanners suc as Cyberware Color Digitizer [8], an active ligt stripper [21], or a stereo camera [5][25]. Tey suffer from various problems and often reproduce te igly reflective surface of airy parts (suc as air and eyebrows) and te sady regions (suc as bottom

3 of nose and between two lips oles) as oles on te final model. Insufficient texture information, nonuniform material property or bad ligting conditions also result in a poor reconstruction quality. Distortion and occlusion also make te problem more difficult and computationally expensive Using 3D Mes Data No capture system is capable of producing errorfree raw data; ence post-processing is always required. For example, 3D Studio Max [9] provides functionality to automatically fill gaps, but it can also remove certain caracteristics from te mes data. One of more uman face sape oriented metodology is using an adaptive mes metodology. Te adaptive mes metod not only removes gaps, but it also provides a mes structure pre-designed for te uman face. Lee et al [17] propose a mes adaptation metod tat uses a two-step matcing procedure. Initially te geometrical deformation is created using a roug matc wit te elp of feature points, tis is ten refined using sopisticated projections on all surface points from te raw data. Te feature information obtained in te roug matcing step provides te partition of surface points on te raw data, so tat te projection can be done eiter to te raw data surface or to te deformed surface wit only feature points. Tis metod provides satisfactory results even on faces were large errors were sown in raw captured data; Figure 2 exemplifies tis procedure. Figure 2: Raw data containing large error is processed to be error-free smoot surface as well as feature information on surface to able to facial expression. 4. Computer-generated face recognition database In tis section, we demonstrate ow computer generated rendering images wit 3D captured facial mes models can be used to build face recognition systems. Firstly a multi-view face data collection procedure is described, followed by a detailed explanation of te 3D face descriptor construction; tis takes systematic extension of a 2D face descriptor to a multi-view using quasi-view concepts. One way to collect large range of poses and also illumination is using rendering of 3D facial mes models. We apply rotation of te desired degree to te mes and ten render it using 2D projection. Te advantage of tis metod is tat it ensures exact rotation degree of te face, various illumination conditions can be simulated virtually and ence any background can be used. Te database contains 108 laser scans wit texture images of size 480 by 400. Te model is aligned to te front view using te information of a nose and two eyes found as facial features in te texture image; tese ave been directly calculated by teir 3D locations on a mes model, and ten rendered wit te desired view direction Detection of Features Texture images of our mes models are carefully aligned in te process of scanning, tat is, te face center is located in te middle of te image and its orientation is uprigt as in te first image of Figure 3. A euristic approac using gray values is used to find te features. First, te minimal rectangular searc region for nose location is set and ten summates te gray pixel values of eac row in te region. Te sum around te nose-end row will be te smallest, or te darkest, due to nostrils and sadows. Te darkest pixel block along tat row corresponds to te real nose-end position. Wit similar metods, te left and rigt eye positions are captured in separate searc regions. For te eyes, searc regions are defined automatically from te acquired nose position. And column-wise summation is initially performed to reduce false detection caused by te eyebrow. In post-processing step for correction, if te eigt of two detected eyes differs too muc, te iger detection is from eyebrow regions, ence, moving te searc space to lower region. If te detected eyes are too close, tey are from nostrils, not real eyes, terefore moving te searc region iger. By reexamining te updated searc region, we were able to locate feature positions successfully for eac image in te database Alignment of te 3D Mes Model Due to te fact we know te 3D coordinates of eyes and nose, say, p left _ eye, prigt _ eye, pnose, from teir locations in te image, te alignment process is quite straigtforward: if we define v = p left p, v2 = 1 _ eye rigt _ eye te vector from p nose and ortogonal to v 1,

4 wic is similar to ( prigt _ eye + pleft _ eye ) / 2 pnose, and v 3 = v1 v2, te rotation given by v1 v2 v3 will transform te model to nearfrontal view. Te resulting model, owever, is sligtly v1 v2 v3 tipped forward for te up-vector approximation and is not parallel to te real up-direction in te face model; terefore we rotate te model to upper view using a fixed angle. Wit te face being aligned and we are able to render it using an arbitrary view direction. We align te face location on te image by fixing two eye locations for eac view as sown in te last column of Figure 3. + Figure 3: Detected feature points in te texture image are mapped to te 3D mes model, wic is aligned to te frontal view and ten rendered wit various view directions 4.3. Definition of te Face Localisation Te images ave size 56x56 and te positions of two eyes in te front view are on (0.3, 0.32) and (0.7, 0.32) wen widt and eigt are considered as 1.0. For example, (0.3, 0.32) means coordinates (17, 18), wic is 17 t pixel in orizontal coordinate from left and 18 nd pixel in vertical coordinate from top for tis image size case. Te left eye position of te positive orizontal rotation is retained at (0.3, 0.32) wile te rigt eye position of te negative rotation does (0.7, 0.32). Vertical rotation as te same eye positions as te ones on zero vertical rotation images. Figure 4 sows one example of face localization (face alignment) definition. 30 o 20 o 10 o (0.3, 0.32) 0 o -10 o -20 o (0.7, 0.32) -30 o -40 o -50 o -90 o -60 o -30 o 0 o 30 o 60 o 90 o Figure 4: eye-positions on view-mosaic of te center face of rendered images of 108 3D-facial mes models (a) (b) Figure 5: (a) Rendered images of a virtual face in unlimited illumination conditions and poses. (b) Face images taken in pre-built studio wit 576 viewing conditions

5 A big advantage using 3D mes data is te freedom to easily set te environment; some examples of various rendered backgrounds are sown in Figure 6 witout te need of croma-keying. Figure 6: different background options: Model is oriented wit orizontal ration 70 o and te ligt source is set wit orizontal rotation 40 o wit directional ligt. A comparison in illumination variation between rendered images from 3D models in a virtual environment and potograps taken using 64 ligts [6] is also sown in Figure 5. We attempted to simulate similar results wit te Yale Face Database [6] taken in a pre-built studio, but wit larger orizontal rotation to sow te freedom of any poses and any illumination possible as sown in Figure 5. We ave used a noncommercial rendering software POV-Ray [11] to create sadow. As a reference, te Yale Database [6] used 64 ligt cones in a spere sape studio and a total 9 poses of 10 people were captured; five poses at 12 and tree poses at 24 from camera s axis. 5. Experimental Results Using a statistical approac is one of te most popular metods for face recognition metods, and many approaces are based on Eigenfaces [20], wic uses Principle Component Analysis as a feature extraction metod. Conventional statistical face recognition systems ave a training step and test step. Te training step is used to build a feature extraction system wit given images tat maximize te difference between various identities wile minimizing te (intrinsic and extrinsic) variation of one identity. Te test step is used to perform classification in given images using te feature extraction obtained in te training step. Better yet, it classifies unknown face image to te proper identity, wit a iger accuracy of face recognition. Most systems, based on training-test, often suffer from lack of variety and over-fitting in given in circumstances. Wit te elp of freedom in virtual environment, we are able to create a large range of various posed facial images. Figure 7 sows rendered images from 3D facial model in various view points. Figure 7: Rendered images from a 3D facial model in various view points. Wit many 3D models, we can generate numerous images in any virtual environment. However tere is anoter problem we ave to consider to build a face descriptor wic is able to recognize te identity. Tere ave been various approaces focused on front view face recognition and we need to take systematic extension to build all-view-face recognition. Wit large-pose-variation database, we create a multi-view face descriptor for any-view face recognition. If we register every 10 o apart, we ave to register 19 x 7 = 133 views to cover orizontal rotation [ ] and vertical rotation [ ], wic results 133 single view descriptor size. Ten te descriptor becomes too large. In tis paper, we focus on feature extraction and descriptor optimization among many issues in multi-view face recognition. As we learn from frontal-view face descriptors, a registered view is used to retrieve nearby views (quasifrontal) wit ig retrieval rate and we extend te concept of quasi-frontal to quasi-view, from frontal view to general view. Definition: Quasi-view wit error rate K - Quasiview V q of a given (registered) view V wit error K means faces on V q are retrieved wit V wit error retrieval rate less tan or equal to K Feature Extraction We use a modified version of te best algoritm in te MPEG-7 advanced face descriptor XM wic is selected in competition among various algoritms to retrieve faces. More details are found in MPEG document [24].

6 normalized face image f(x,y) Subregion k parts f i (x,y) of te face image concentrated around tis range. Note tat tey ave cecked only orizontal rotation of uman eads. Fourier Transform Error rate 0.2 F(u,v) x 1 PCLDA projection P 1 F(u,v) x 2 PCLDA projection P 2 F j (u,v) x 1 j PCLDA projection P 1 j F j (u,v) x 2 j PCLDA projection P 2 j j = 1,.., k Quasi-view size Error rate 0.05 y 1 LDA Projection P 3 f Z Quantization y 2 HolisticFourierFeature Vector Normalization y 1 j LDA Projection P 3 j Z i Quantization y 2 j j t Subregion FourierFeature j = 1,.., k Figure 8: Feature extraction used for Multi-view 3D Face Descriptor Our Subregion-based Linear Discrimination analysis (LDA) on Fourier space as sown in Figure 8 is designed for multi-view purpose. Te biggest differences between te MPEG-7 XM [12] and our model are (i) feature extraction in luminance space is removed, (ii) te subregion decomposition, wic was in luminance space, is now in Fourier space, and (iii) te number and positions of subregions are varied depending on a given view. Te first two modifications give more efficient feature extraction wit smaller descriptor size. Te tird modification is caused by te non-frontal-view extension. Te new view-dependent subregion is sown in Figure Horizontal Rotation degree Vertical Rotation degree (a) (b) Figure 10: Quasi-view sizes depending on a registered view. Quasi-view size also depends on te view. We experiment quasi-view inspection wit error rate Figure 10 (a), wic as very similar pattern to te ones in [7], sows ow te quasi-view size varies wit orizontal and vertical rotations of a ead. To make a fair comparison between different views, we extracted 24 olistic features (witout using subregion features) for eac view. Te views (20, 0 ) ~ (30, 0 ) ave bot te largest quasi-view size and te largest Euclidean distance between te people in Eigenspace among views (0, 0 ), (10, 0 ),... and (90, 0 ). Figure 10 (b) sows te views (0, 0 ) ~ (0, 10 ) ave te biggest quasi-view size among views (0, -60 ), (0, -50 ), and (0, 60 ). One point to note is tat te ead-down views ave bigger quasi-view size tan ones of te ead-up and it demonstrates it is easier to recognize people wen tey look downward more tan tey look upward Descriptor optimization Five subregion definition on View (0 o, 0 o ) Two subregion definition on View (80 o, 0 o ) Figure 9: Subregion definition depending on views is sown as superimposed on te center face of our rendering face database 5.2. Quasi-view Graam and Allinson [7] ave calculated te distance between faces over pose to predict te pose dependency of a recognition system. Te faces are furter apart tey will be easier to recognize using distance measures in te Eigenspace and consequently, te best pose samples to use for an analysis sould be Optimization criteria depends on were te empasis is placed, suc as te smallest number of registration poses, te smallest descriptor size, or te biggest coverage of poses by selecting views efficiently describing te 3D of te facial features. Our first experiment was made using 50/50 ratios wit rendered images from 3D mes models; tis means tat alf te images were used for training and te oter alf for test. We concentrate on te selection by quasi-view size, te number of subregions, and te number of features in eac subregion. So for some views, 5 olistic features and 5 features for 5 subregions are extracted and for oter views 5 olistic features and 2 features for 5 subregions are extracted. If a view is close to profile, we use 5 olistic features

7 erticalrotationerticalrotationand 5 features for 2. For one view, one image is selected. Horizontal rotation VFigure 11: Rendered image experiment: Te region (93.93%) covered by 7 quasi-views in te view-mosaic of positive orizontal rotation wit error rate Registration wit 13 views are enoug to retrieve over 90% of views of orizontal rotation [ ] and vertical rotation [ ] wit error rate Troug experiments wit various combinations of quasi-views, a set of views is selected to create multiview 3D descriptor. An example of several possible descriptors sown in Figure 11 as 240 dimensions wit rendered images. Tis descriptor is able to retrieve te rendered images in te test database wit error rate 0.05 covering % views of total region of view-mosaic of orizontal rotation [ ] and vertical rotation [ ]. For a reference, it covers 95.36% for error rate 0.1, 97.57% for error rate 0.15 and 97.98% for error rate 0.2. Horizontal rotationvfigure 12: Video image experiment: Te region (84.21%) covered by 14 quasi-views in te viewmosaic of orizontal rotation [ ] and vertical rotation [ ] wit error rate 0.2. For comparison, video images taken in real situation are also experimented in te same processes. Here te DB size was 114 wit 50/50 ratio, wic means 57 identities used for training and 57 identities for test. We position a video camera at te front of a uman subject and video streams are taken by rotating te uman subject s cair, so we were able to obtain images of orizontal at interval 1 o. We ave taken 7 video streams of eac uman subject were eac time te uman s neck is fixed in different vertical rotation. Due to te difficulty to collect large variation of poses in real environment, only 10 apart vertical rotation images are taken wile te orizontal rotation is as dense as te rendered images. Experiment wit images extracted from video images as a lower success rate. Te training data does not ave pose information as accurate as te rendered images from te 3D-facial mes models, especially vertical axis of views. In addition te uman subject as been sat for a wile and rotated to take video streams, so facial, air and body movement and deformation exist. So te quasi-view size of video images is smaller tan te one of rendered images. An example of many possible descriptors is sown in Figure 12. Oter examples are 300 dimensions wit 87.22% coverage and 270 dimensions wit 85.71% coverage wit te same error rate. As a reference for dimension, te current Advanced Face Descriptor [24][12] as 48 dimensions wit error rate, called as ANMRR, and 128 dimensions wit ANMRR for 50/50 ratio for potograp images. Here ANMRR is a MPEG definition for error rate. AFD is verified only wit limited pose variation suc as about (±30, 0 ) and (0, ±30 ), wic means te coverage of view is very small compared to our 180 orizontal rotation and 60 vertical rotation. 6. Conclusion We ave clearly sown te usage of computergenerated rendered images wit captured 3D mes facial data for face recognition purpose and multi-view face descriptor metodology wic is a systematic extension of conventional one-view face recognition descriptor. Te database construction for very dense poses using a rendering of a 3D facial mes obtained from a laser scanner is explained and also a comparison wit video images are also given. Our multi-view approac wit quasi-view dedicated to uman face structure sows acceptable size of descriptor and satisfactory recognition rate for recognizing identity of uman face in any view.

8 Virtual environments are far easier to manipulate tan pysically built-in environments to produce variety in extrinsic parameters like poses and illumination, wic as been bottle-neck in database construction for face recognition. Our experiment for multi-view face recognition sows te potential of CG models in various applications. As te computergenerated 3D mes tecnologies progress to produce more and more realistic uman face modeling and animation, te intrinsic parameters suc as facial expression, air, and aging can be simulated in virtual environment for face recognition purpose as well as extrinsic parameters we ave discussed in tis paper. 7. References [1] Badler, N., Allbeck, J. and Bindiganavale, R., "Describing uman and group movements for interactive use by reactive virtual caracters," Army Science Conference, 2002 [2] Blanz, V. and Vetter, T. A Morpable Model for te Syntesis of 3D Faces, SIGGRAPH'99 Conference Proceedings, pp [3] Blanz, V. and Vetter, T., Face Recognition Based on Fitting a 3D Morpable Model, IEEE Transactions on Pattern Analysis and Macine Intelligence arcive, Volume 25, Issue 9, September 2003 [4] Bueler, C., Bosse, M., Gortler, S., Coen, M. and McMillan, L., "Unstructured Lumigrap Rendering," Proc. ACM SIGGRAPH '01, pp , [5] Cen, Q. and Medioni, G., Building Human Face Models from Two Images, IEEE 2 nd Worksop Multimedia Signal Processing, pp , Dec [6] Georgiades, A.S., Belumeur, P.N. and Kriegman, D.J., From few to many: Generative models for recognition under variable pose and illumination, In Proceedings of te 4t IEEE International Conference on Automatic Face and Gesture Recognition, [7] Graam, D. B. and Allinson, N. M., Caracterising Virtual Eigensignatures for General Purpose Face Recognition in Face Recognition: From Teory to Applications (H. Wecsler, PJ Pillips, V. Bruce, FF Soulie and TS Huang, eds.), Berlin: Springer-Verlag, pp , 1998 [8] ttp:// [9] ttp:// [10] ttp:// [11] ttp:// [12] Kamei, T., Yamada, A., Kim, H., Hwang, W., Kim, T.- K. and Kee, S. C., CE report on Advanced Face Recognition Descriptor, ISO/IEC JTC1/SC29/WG11 M9178, Awaji, JP, December 2002 [13] Kim, W.-Y., Lee, J.-H. Park, H.-S. and Lee, H.- J., PCA/LDA Face Recognition Descriptor wit Pose, MPEG-7 video, Oct [14] Lee Y., Terzopoulos D. and Waters K., Realistic Modeling for Facial Animation, In Computer Grapics (Proc. SIGGRAPH), ACM Press, pp , [15] Lee, M. W., Coen, I., Jung, S. K., Particle Filter wit Analytical Inference for Human Body Tracking, IEEE Worksop on Motion and Video Computing, [16] Lee, W.-S. and Kee, S. C., 3D-Face Descriptor: proposal for CE, ISO/IEC JTC1/SC29/WG11 M9422, Pattaya, Tailand, Marc 2003 [17] Lee, W.-S. and Magnenat-Talmann, N., Fast Head Modeling for Animation, Journal Image and Vision Computing, Volume 18, Number 4, pp , Elsevier Sceince, 1 Marc, [18] Lee, W.-S. and Son, K., Multi-view 3D Face Descriptor: proposal for CE, ISO/IEC JTC1/SC29/WG11 M9801, Trondeim, Norway, July, 2003 [19] O'Brien, J. F. and Hodgins, J. K., Grapical Modeling and Animation of Brittle Fracture. Te proceedings of ACM SIGGRAPH 99, Los Angeles, California, August 8-13, pp , [20] Pentland, A., Mogaddam, and B., Starner, T., View- Based and Modular Eigenspaces for Face Recognition, Proc. of IEEE Conf. on Computer Vision and Pattern Recognition (CVPR'94), [21] Proesmans, M. and Van Gool, L., "Reading between te lines - a metod for extracting dynamic 3D wit texture". In Proc. of VRST'97, pp , [22] Sim, T., Baker, S. and Bsat, M., Te CMU Pose, Illumination and Expression Database, IEEE Transactions on Pattern Analysis and Macine Intelligence, [23] Virtual Media, Face database constriction for researc, KISA, [24] Yamada, A. and Cieplinski, L., MPEG-7 Visual part of experimentation Model Version 17.1, ISO/IEC JTC1/SC29/WG11 M9502, Pattaya, Tailand, Marc 2003 [25] Zang, L., Curless, B. and Seitz, S., Spacetime Stereo: Sape recovery for dynamic scenes, CVPR 2003