Accurate Image Based Relighting through Optimization

Transcription

1 Thirteenth Eurographis Workshop on Rendering (2002) P. Debeve and S. Gibson (Editors) Aurate Image Based Relighting through Optimization Pieter Peers Philip Dutré Department of Computer Siene, K.U.Leuven, Belgium Abstrat In this paper we present a new relighting tehnique that, for a single viewpoint, aurately aptures the refletane field of real and virtual objets, without restritions on their geometrial omplexity or material properties. As a result the objets an be relit under arbitrary lighting onditions. To apture the refletane field, we take photographs of an objet lit by several lighting patterns on a surrounding disretized hemiube. The illumination of eah pixel due to emitted radiane from eah disrete path on the hemiube is approximated by a refletion oeffiient and a retangular support whih are found through an optimization proedure. The disretization of the hemiube ensures suffiient angular sampling to apture diffuse material properties. The use of a sub-path support aommodates small solid angles of inoming light important for speular materials. To relight the objet, the target illumination, a high-dynami range environment map, is averaged over the support and multiplied by the refletion oeffiient, per pixel and per disrete path on the hemiube. The results obtained show aurate relighting of diffuse, glossy and speular objets. Categories and Subjet Desriptors (aording to ACM CCS): I.3.7 [Computer Graphis]: Three dimensional graphis and realism I.4.1 [Image Proessing and Computer Vision]: Digitization and Image Capture 1. Introdution Image-based relighting represents a lass of tehniques that apply new lighting onditions to a sene, given a set of basis images. Possible appliations range from lighting design to augmented reality. Lighting design an be a tedious task, sine moving objets and light soures is not always feasible. Objets an be very fragile (e.g. arhaeologial artifats), the illumination is not ontrollable (e.g. the sun through a skywindow) or an objet is too massive to move around (e.g. a statue). Augmented reality appliations are onstantly evolving and pushing the limits of widely aepted tehniques. Plaing real objets in virtual environments influenes the appearane of the objets. It is not always possible to approximate the desired illumination at the moment an objet is aptured (e.g. speial effets in movies or omputer games whih ombine real and virtual environments). Relighting is an important part of omputer graphis appliable to a large range of appliation. In this paper we develop a relighting tehnique that ombines strengths from several tehniques, the Light Stage 2 5 {pieterp, phil}@s.kuleuven.a.be and environment matting 1 14, into a single framework. This allows us to apture omplex geometrial objets without plaing restritions on material properties. To apture the refletane field a hemiube around the objet is disretized in several light pathes. Eah path emits a series of illumination patterns and for eah pattern a highdynami range photograph 3 is reorded. Currently we use a alibrated CRT monitor per hemiube side to emit the illumination patterns (figure 1). For eah pixel and for eah light path the reeived pixel radiane is then approximated by a refletion oeffiient and a support area on the light path. This support area approximates the area on the light path that is important for the pixel under onsideration. A least square minimization proedure on the pixel radiane over different illumination patterns is used to determine the best approximation for the support. To relight the sene with an arbitrary high-dynami range light map, the average illumination emitted over the support area by the light map is omputed and multiplied by the refletane oeffiient. This is repeated for eah pixel and eah light path. The results show that the tehnique is able to aurately relight omplex geometry with diffuse, glossy or speular material properties.

2 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization The paper is strutured as follows: In the next setion we disuss previous work. In setion 3, a general overview of the tehnique is given, and is presented as a three step algorithm in setions 4, 5 and 6. The auray is verified in setion 7 and the relation with the Light Stage and environment matting is further explored in setion 8. Some results obtained with our relighting tehnique are found in setion 9. Finally diretions for future work and a onlusion is given in setion Previous Work In this setion we give an overview of work related to image based relighting. A first lass of tehniques requires or reonstruts a geometrial model of the sene. Losos et al. 8 9 reonstrut a simplified geometrial model, approximate the illumination and alulate unoluded illumination textures with an adapted radiosity algorithm. Relighting is performed by modifying light intensities. Yu et al fous mainly on reovering the refletane properties of materials by using an iterative optimization proedure to find parameters to represent the refletane properties of surfaes. The sene with the reovered material properties is re-rendered using a global illumination rendering system. Reonstrution of geometry is advantageous when visualization from multiple viewpoints is desired, but is a disadvantage when a sene onsists out of geometrial omplex objets. Another lass of tehniques irumvents the need for omplex geometrial models by diretly manipulating images. Nimeroff et al. 10 introdued the onept of linearly weighting and ombining basis images of a sene. The lighting ondition in these basis images are hosen in suh a way that any desired novel lighting ondition an be onstruted by linearly ombining and weighting these basis images. Wong et al. 11 based their relighting method on light field rendering 4 6. By sampling the sene under different viewpoints and illumination an apparent BRDF for eah surfae pixel is found. Lin et al. 7 define a refleted irradiane field, the dual of a light field, and derive upper sampling bounds for eah BRDF to reprodue it truthfully. The Light Stage 2 5 uses basis images of a sene lit by lights regularly spaed on a sphere surrounding the sene. A light map is applied by linearly ombining these basis images into a single resulting image. Environment matting, originally introdued by Zongker et al. 14 and extended by Chuang et al. 1, apture the refletane properties of speular and refrative materials by lighting them with different illumination patterns. The diretion and solid angle important for the illumination of a pixel an be derived from this information in the form of a refletion oeffiient and a support on the bak or side-drops. Our tehnique ombines the strength and features of the Light Stage and of environment matting. Similar to the Light Stage inoming light diretions are suffiiently disretized to approximate the refletane properties of diffuse materials, where as the use of illumination patterns, similar to environment matting, makes it possible to find small solid angles important to approximate speular refletane properties. 3. Overview of the method The goal of our relighting method is to apply a new light map (in the form of an environment map) to an existing sene, photographed from a fixed viewpoint. To do this, we need to know, for eah diretion in the light map, how muh it ontributes to the radiane seen through a pixel. To find these ontributions, the sene an be lit by light soures loated around the objet on a bounding volume and the ontribution to the illumination through a pixel an be reorded. We an now formulate the goal of our relighting method: for eah pixel, the goal is to ompute the weight (ontribution to the total radiane seen through the pixel) of eah point on the bounding volume for an existing sene. With this it is possible to relight the image using any environment map. The bounding volume is disretized in a number of light pathes that an be lit independently of other light pathes. Consider for a moment that only diret illumination emitted from a path ontributes to the radiane seen through a pixel. If the refletion to the eye is aording to a diffuse BRDF, then all diretion in the path will be equally signifiant. If on the other hand the refletion to the eye is aording to a speular BRDF, then only a few diretions are signifiant (loated around the perfet refleted diretion). The set points on the light path that are signifiant for the ontribution to the pixel radiane are alled the support of that light path. In this paper we model the refletane for eah pixel and eah light path by a refletion oeffiient and a support. The refetion oeffiient represents the global influene of the light path to the observed pixel radiane, where as the support represents the subset of important diretions within a light path. This an be written as: I p R p k k S p k L k The intensity I in pixel p equals the sum over all light pathes k of the refletion oeffiient R p k multiplied by the average of the inoming radiane L k over the support S p k. To relight a pixel, the illumination values in the light map need to be averaged over the support. We an generalize this averaging as a filter operation. We all this filtering the evaluation of the support. We used a box filter in the text above, but in priniple any filter an be used. We will now develop a method to apture, ompute and use these sets of refletion oeffiients and supports to relight eah pixel. First a number of basis images are reorded

3 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization (setion 4). From this gathered input data the refletion oeffiients and supports are extrated using a least square minimization algorithm (setion 5) and stored for relighting afterwards. This relighting (setion 6) an be done an arbitrary number of times with different light maps and requires minimal proessing time. 4. Basis Image reording 4.1. General overview We measure the refletane fields as follows: a hemiube is assumed around the objet for whih the refletane field will be aptured. Eah side of the hemiube is subdivided in a regular grid of onneting light pathes. From eah of these light pathes a series of illumination patterns are emitted (while emitting no radiane from other light pathes) and a high-dynami range photograph 3 is reorded. If n illumination patterns are emitted from eah light path and there are m light pathes then a total of n m high-dynami range photographs will be reorded Pratial setup In our experiments we used a CRT monitor for eah side of the hemiube (figure 1). Figure 1: A CRT monitor is used for eah side of the hemiube. A 4 4 grid of light pathes emits a series of illumination patterns. For eah light path a refletion oeffiient and support are omputed. Our urrent setup limits monitor plaement to 4 sides, exluding the side where the amera is loated. The CRT monitor needs to be radiane alibrated, beause the mapping between pixel values and emitted radiane is a non-linear funtion for most CRT monitors. To alibrate the CRT monitor, a high-dynami range photograph is reorded of the CRT monitor displaying a alibration pattern. All our illumination patterns are transformed using the non-linear mapping between radiane and pixel values. In our experiments we use a 4 4 grid on eah hemiube side. This hoie is motivated by the fat that it produes Figure 2: The illumination patterns used to reord the basis images. The solid white pattern is displayed in the enter, while inverse patterns are displayed on opposite sides. good results for a reasonable amount of reorded photographs. Due to pratial limitations the side faing the amera was omitted Illumination Patterns For eah light path a refletion oeffiient and support needs to be alulated for a pixel. The illumination patterns need to provide enough information to extrat this data. The hoie of whih illumination patterns depends on the hoie of the speifi filter used and on the representation of the supports. In this paper use an axis-aligned support and a box filter. It is possible to alulate the refletion oeffiient using just one illumination pattern. The refletion oeffiient is defined in setion 3 as the global influene of a light path to the pixel radiane. By emitting a solid white pattern, the refletion oeffiient an be diretly omputed from the observed pixel radiane. There are many kinds of illumination patterns that an be used to ompute a support. In environment matting were three different kinds of illumination patterns used: horizontal and vertial stripes (ombined with a box filter), Gaussian bands sweeping over the sreen at 0, 45, 90 and 135 degrees (ombined with a elliptial Gaussian filter) and a single olored gradient (restrited to olorless materials). We used a variation on gradients, and use 4 quadrati (gray-sale) gradients oriented at 45 n 90 degrees with n 0 3 (figure 2). Sine dark regions in a high-dynami range photograph are prone to noise, we also used 4 inverse patterns, beause the inverse patterns are bright where the normal patterns are dark and vie versa. This way we are sure that the effets of noise are minimized.

4 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization A detailed analysis of the hoie of supports is beyond the sope of this paper. 5. Calulation of the refletion oeffiients and supports One the basis images are reorded, the refletane oeffiient and its support is found for eah pixel and for eah light path through an optimization proedure. A pixel onsists out of a number of hannels (e.g. RGB) and two approahes are possible. The first possibility is to treat eah hannel independently and alulate a support and refletion oeffiient for eah hannel. The seond possibility is to use the same support for eah hannel and treat the refletion oeffiient as a vetor of refletion oeffiients for eah hannel. We used the latter approah. As said, the refletion oeffiient an be diretly derived from the basis image lit with the solid white pattern. The evaluation of any support over a solid white illumination pattern equals a onstant fator. The observed pixel value lit by the solid white illumination pattern equals the refletion oeffiient multiplied by the onstant fator. To find the support within a light path, a least square optimization is used. We use a steepest deent optimization algorithm (algorithm 1). Figure 3: The refletion oeffiients and supports of 4 different pixels. The size of eah square represents the support and the intensity the refletion oeffiient (a non-linear mapping was applied to the intensities). The upside down T represents the unfolded hemiube (top; left, bak, right). is squared and summed for the trail support over all illumination patterns. Algorithm 1: SquaredError S 2 R p P j S I p P j j 1) Selet an initial support 2) Calulate the least-squared error 3) while not minimized a) Test all expansion possibilities of the support b) Calulate least-squared errors and selet minimum error ) if minimum error < error then update support else minimized = true There are a number of possible initial supports. In our implementation we use the following: if the refletion oeffiient of the previous pixel is almost equal as the refletion oeffiient of the urrent pixel then the initial support is set equal to the support of the previous pixel. If the refletion oeffiients differ too muh then the initial support is set equal to the entire light path. The idea is that if the refletion oeffiients are almost equal then it is likely that the material properties of these pixels are almost the same and thus have a simular behavior with respet to supports. To ompute the least square error in eah optimization step, a trial support is evaluated for eah illumination pattern and saled by the known refletion oeffiient. The omputed value will differ from the observed pixel value in the basis image lit by the same illumination pattern. This error The summed squared error is alulated over the illumination patterns P j for a trial support S saled by a refletion oeffiient R p. I p P j denotes the observed pixel p lit by the illumination pattern P k. The use of a box filter and an axis-aligned support makes it possible to use summed area tables on the illumination patterns whih signifiantly speed up alulation of the squared error. In figure 3 the refletion oeffiients and supports are plotted for a number of different pixels. The size of the squares represents the size of the different supports and the intensity represents the magnitude of the refletion oeffiients. Pixel A represents a speular surfae. The bright white support is around the prefet refleted diretion. The other supports visible for this pixel are less then 0.5 perent in intensity and are due to noise in the reorded basis images. Pixel B represents a transparent surfae. One an learly see the two distint sets of supports representing the influene from the refletion (left) and refration (right). Pixel C represents a diffuse surfae. All supports are almost maximal in size. The reason that there are almost no supports visible for the bakside is due to olusion by the mask. Finally, pixel D represents a diffuse surfae at the base of the glass sphere. The reader an learly see the effets of diret (upper hemiube side) and indiret illumination (right hemiube side) due to refration through the glass sphere in the different supports.

5 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization 6. Relighting One the refletion oeffiients and supports are alulated for eah pixel and eah light path it is easy to apply a new light map to the sene. To relight the sene using an arbitrary high-dynami range environment map, the support is evaluated over the new illumination and saled by the refletion oeffiient, for eah light path and for eah pixel. The results from the different supports are summed for eah pixel, resulting in the total observed radiane for that pixel. In ase a box filter was used over the different supports a signifiant speedup an be obtained by using summed area tables for the light-map, beause the supports differ for eah pixel and thus for eah pixel, the illumination pattern has to be evaluated over the support. 7. Verifiation Different degrees of auray are possible with the proposed tehnique. The disretization rate of the hemiube and the number of illumination patterns diretly influene the auray. In this setion we verify the auray. We use an artifiial sene ontaining 4 spheres, eah with different material properties, plaed on a diffuse surfae (figure 4) and lit by a light map plaed at the righthand side of the sene. A referene image of this sene was rendered using a global illumination renderer (figure 4, left piture). The basis images lit with the illumination patterns were also rendered using the same global illumination renderer. The refletion oeffiients and supports were alulated with our method using these basis images. Finally the same light map that was used in the referene image was used in the relighting step of our tehnique (figure 4, right image). No differene is visually notieable, soft shadows are reprodued aurately (figure 4 zoom-ins on the lefthand side) and glossy/speular refletion are reprodued without visual differene (figure 4 zoom-ins on the righthand side). The average relative error over all pixel onfirms the visual verifiation and was less than 1 perent. The verifiation shows that our method an aurately relight diffuse, glossy and speular materials. Soft shadows are also reprodued faithfully. 8. Disussion In this setion we disuss the impliations of our developed method. One an see that the proposed tehnique is a superset of methods like the Light Stage and environment matting. The disretization of the inoming light diretions diretly maps to the regular sampling on a sphere surrounding the sene by the Light Stage. The Light Stage is good in reproduing diffuse material properties, but fails for speular materials beause the oarse angular sampling is not suffiient to apture the high frequeny spikes of speular refletion funtions. Illumination patterns to determine a support are also used in environment matting. This method is very good in reproduing speular and transparent material properties, beause it is able to find a single small solid angle important for the refletion funtion of speular materials. Environment matting however fails for diffuse materials beause the support would over the entire hemisphere and a single refletion oeffiient (ombined with a simple filter) is unable to approximate the refletion funtion suffiiently. A straightforward ombination of the Light Stage and environment matting (proposed by Debeve et al. 2 ), where a two step data aquisition stage first preforms a step simular to the Light Stage and then as a seond step preforms an environment matting step, would only be able to orretly handle diffuse and speular materials. Glossy materials would still be diffiult to apture beause their angular frequeny ould be too small for the Light Stage but still be too large for environment matting to represent it by a single refletion oeffiient and a support (and a simple filter). Using more omplex filters (e.g. oriented Gaussian filters as proposed by Chuang et al. 1 ) would minimize this problem, but not ompletely solve it. Certain materials are almost impossible to represent by a single refletion oeffiient and a filter. The proposed method differs from the straightforward ombination due fat that for eah disrete part of inoming light diretions a support and a refletion oeffiient are alulated. Glossy materials an still be approximated with simple filters and with suffiient auray beause they still an span several supports and refletion oeffiients, but aren t bound by the disretization rate of the inoming light diretions. Using more omplex filters (e.g. elliptial Gaussian filters) would allow even more aurate approximations at the ost of extra illumination patterns. The fat that everything is on a per-pixel-basis and that a different support is possible per disrete set of inoming light diretions makes it possible to use a muh oarser subdivision of the inoming light diretions. The proposed method ombines the strengths of the Light Stage and environment matting in an unified framework. The results in setion 9 show that the method is indeed able to apture a wide range of materials even with a limited number of photographs. 9. Results In this setion two results (figure 5 and figure 6) are disussed. Both images show a wooden toy ar plaed on a diffuse surfae. Figure 5 is relit using a light map reorded inside the Uffizi Galleria, where most of the light omes from diretly above. The zoom-in on the bumper shows the elongated glossy highlight aused by the left to right orientation of the illumination in the light map. The zoom-in on the bak of the toy ar learly show the relatively hard shadows aused by the illumination from above. Figure 6 is relit using an artifiial light map, where eah side of the hemiube emits a different olor. One an learly distinguish the four sides of the hemiube in the refletion

6 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization Figure 4: A omparison of a rendered (left) and relit image (right). Both use the same light map plaed at the righthand side of the sene. in the glossy bumper and the zoom-in on the bak of the ar shows almost no shadow beause the radiane emitted from eah side of the light map is approximately the same. To apture the basis images for these senes we used a 19 Iiyama CRT monitor. We moved this monitor around for eah side of the hemiube. By using only one monitor we avoid subtle differenes between different monitors. The photographs were reorded using a Canon EOS D30 amera. We automated the proess of apturing and, as mentioned before, only the monitor had to moved around. The reording time was about 10 hours (of whih at least 6 hours were due to overhead of the slow speed of the amera). Reording ould be sped up by using a digital video amera and using different F-stop filters to aquire high-dynami range images. The proessing took 20 hours on an Athlon 1.2Ghz system with 512meg ram. This omes down to approximately 1 pixel per seond per 200Mhz of pu power. Sine eah pixel an be proessed independently, it is possible to subdivide the proessing time among different omputers without loss in performane. The relighting an be done in less than a minute and ould be done in real-time if all refletion oeffiients and supports would fit in the main memory (16 4 NrPixels supports). When high-dynami range environment maps are of muh larger dynami range than the range of the monitor, visual artifats appear. Neighboring pixels an have relatively different refletion oeffiients due to noise in the reorded highdynami range photographs, ombine this with large intensities from the environment map for a ertain support, and large differenes beome visible for neighboring pixels. This problem ould be redued by applying noise redution filters as preproessing on eah reorded basis image. 10. Conlusions In this paper we developed a new and pratial tehnique for apturing the refletane field of objets for a fixed viewpoint, but without plaing restritions on material properties or geometrial omplexity. The method sales very well in terms of auray versus number of basis images. Higher auray an be obtained by using more omplex filters and/or a finer disretization. The pratial implementation an be done with off the shelf materials. Diffuse, glossy and speular materials are reprodued faithfully, as are more omplex illumination effets (suh as austis). A whole new range of appliations is possible: the digital apture and display of objets with a mix of diffuse and speular appearane (e.g. jewelry or sensitive arhaeologial objets), augmented reality appliations, (interative) relighting of fully virtual senes,... Future researh will fous on faster and easier apturing of the refletane field, the use of other projetion devies suh as VR-aves that an aommodate larger objets, reduing the number neessary photographs by using better patterns (e.g. oriented box filters), and better optimization proedures. More results in olor an be found at the dediated web page: graphis/ CGRG.PUBLICATIONS/ARTO/ Aknowledgments We would like to thank Frank Suykens, Vinent Masselus and Karl vom Berge for proof-reading the paper and many

7 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization onstrutive disussions. Furthermore we would like to thank Paul Debeve for his high quality environment maps ( All artifiial senes were rendered using RenderPark ( Referenes 1. Yung-Yu Chuang, Douglas E. Zongker, Joel Hindorff, Brian Curless, David H. Salesin, and Rihard Szeliski. Environment matting extensions: Towards higher auray and real-time apture. In Kurt Akeley, editor, Siggraph 2000, Computer Graphis Proeedings, Annual Conferene Series, Los Angeles, ACM Siggraph, Addison Wesley Longman. 2. Paul Debeve, Tim Hawkins, Chris Thou, Haarm- Pieter Duiker, Westley Sarokin, and Mark Sagar. Aquiring the refletane field of a human fae. In Kurt Akeley, editor, Siggraph 2000, Computer Graphis Proeedings, Annual Conferene Series, pages , Los Angeles, ACM Siggraph, Addison Wesley Longman. 3. Paul E. Debeve and Jitendra Malik. Reovering high dynami range radiane maps from photographs. In Turner Whitted, editor, Siggraph 97 Conferene Proeedings, Annual Conferene Series, pages ACM Siggraph, Addison Wesley, August Steven J. Gortler, Radek Grzeszzuk, Rihard Szeliski, and Mihael F. Cohen. The lumigraph. In Holly Rushmeier, editor, Siggraph 96 Conferene Proeedings, Annual Conferene Series, pages ACM Siggraph, Addison Wesley, August Tim Hawkins, Jonathan Cohen, and Paul Debeve. A photometri approah to digitizing ultural artifats. In In 2nd International Symposium on Virtual Reality, Arhaeology, and Cultural Heritage, Glyfada, Greee, November Mar Levoy and Pat Hanrahan. Light field rendering. In Holly Rushmeier, editor, Siggraph 96 Conferene Proeedings, Annual Conferene Series, pages ACM Siggraph, Addison Wesley, August Zhouhen Lin, Tien-Tsin Wong, and Heung Yeung Shum. Relighting with the refleted irradiane field: Representation, sampling and reonstrution. In Proeedings of IEEE Computer Vision and Pattern Reognition, Deember Céline Losos and George Drettakis. Low-ost photometri alibration for interative relighting. In Proeedings of the First Frenh-British International Workshop on Virtual Reality, Brest, Frane, July Céline Losos, George Drettakis, and Lu Robert. Interative virtual relighting of real senes. In IEEE Transations on Visualization and Computer Graphis, volume 6(4), pages IEEE Computer Soiety, Jeffry S. Nimeroff, Eero Simonelli, and Julie Dorsey. Effiient Re-rendering of Naturally Illuminated Environments. In Fifth Eurographis Workshop on Rendering, pages , Darmstadt, Germany, June Tien-Tsin Wong, Pheng-Ann Heng, Siu-Hang Or, and Wai-Yin Ng. Image-based rendering with ontrollable illumination. In Julie Dorsey and Philipp Slusallek, editors, Eurographis Rendering Workshop 1997, pages 13 22, New York City, NY, June Eurographis, Springer Wien. 12. Yizhou Yu, Paul Debeve, Jitendra Malik, and Tim Hawkins. Inverse global illumination: Reovering refletane models of real senes from photographs from. In Alyn Rokwood, editor, Siggraph 99, Annual Conferene Series, Annual Conferene Series, pages , Los Angeles, ACM Siggraph, Addison Wesley Longman. 13. Yizhou Yu and Jitendra Malik. Reovering photometri properties of arhitetural senes from photographs. In Mihael Cohen, editor, Siggraph 98 Conferene Proeedings, Annual Conferene Series, pages ACM Siggraph, Addison Wesley, July Douglas E. Zongker, Dawn M. Werner, Brian Curless, and David H. Salesin. Environment matting and ompositing. In Alyn Rokwood, editor, Siggraph 1999, Computer Graphis Proeedings, Annual Conferene Series, pages , Los Angeles, ACM Siggraph, Addison Wesley Longman.

8 Pieter Peers and Philip Dutré / Aurate Image Based Re-lighting through Optimization Figure 5: A wooden toy ar plaed on a diffuse surfae, relit with a light map reorded in the Uffizi Galleria. Note the glossy highlight and the soft shadows. Figure 6: A wooden toy ar plaed on a diffuse surfae, relit with an artifial light map.