FR2994500A1 - METHOD OF OPTIMIZING THE EXECUTION OF A GRAPHIC PIPELINE - Google Patents

Abstract

The subject of the present invention is a method for rendering images in real time from a three-dimensional virtual scene comprising a set of objects, comprising: a preparation step (S1) of at least a part objects in the scene as a first set of polygons; a first rasterization step (S2) of the first set of polygons at a first resolution less than the final resolution of the image so as to obtain a first depth buffer; information being collected during the first rasterization step corresponding to the identifiers of the visible polygons of the first set of polygons, that is to say to the polygons defining at least one pixel of the first depth buffer, so as to obtain a second buffer identifying and defining a second set of polygons among the first set of polygons; and - a second rasterization step (S4) of the second set of polygons at the final resolution of the image.

Description

The present invention relates to a real-time image rendering method from a three-dimensional virtual scene comprising a set of objects, as well as a system implementing such a method. More particularly, the present invention relates to a rendering method using a rasterization step. In most existing computer equipment, real-time rendering is at least partially performed by graphical computing acceleration means, such as GPU graphics processors, which allow parallel processing of certain operations, particularly rasterization. However, some equipment such as laptops or mobile phones include means of accelerating the graphical computation that is not very advanced, which leads to slow usage and leads to limiting the use of rendering in real time. However, it is desirable to use such a type of rendering in certain nomadic uses requiring mobility of the user for which portable equipment is needed. For this type of mobile use, the equipment must make it possible to efficiently explore dynamic scenes and to ensure a high quality of display of the scenes, with an immobile object rendering close to reality in order to allow manipulations. virtual objects. Several solutions have already been proposed to speed up rendering processing. In particular, [Seiler et al. 2008, Larrabee: a many-core x86 25 architecture for visual computing. ACM Trans. Graph. 27, 3 (Aug.), 18: 118: 15] presented an architecture adapted to rasterization programs. However, the use of this specific architecture involves significant production costs and limited portability. More recently, a reverse approach of adapting the graphical pipeline to the architecture of existing computers has been adopted, as in [Liu et al. 2010, Freepipe: Programmable parallel rendering architecture for efficient multi-fragment effects. ACM SIGGRAPH Symposium on Interactive 3D Graphics and Game, ACM, New York, NY, USA, 13D'10, 75-82] and [Laine and Karas 2011, T. 2011. High-performance software rasterization on GPUs. In Proceedings of the ACM SIGGRAPH Symposium on High Performance Graphics, ACM, NY, NY, USA, HPG '11, 79-88]. These two documents present a programming and parallel computing platform known as CUDATM based on the current 5 graphics processors. [Peng and Cao 2012, A GPU-based Approach for Massive Mode! Rendering with Frame-to-Frame Coherence. Computer Graphics Forum 31, 2, 393-402] introduce a hybrid CPU / GPU approach based on the minimal data exchange between the CPU core processor and the GPU graphics processor, using the consistency of successive frames on a graphics processor. High end GPU. Nevertheless, these high-end GPU graphics processors are not present in the majority of mobile devices, which limits a possible use of this technique in most mobile devices. EA Digital Illusions CE (CE) [Andersson 2009, Parallel graphics in frostbite - Current & Future. Course Notes SIGGRAPH 2009] has created a game engine named "Frostbite Engine" based on a hybrid pipeline for their cross-platform, especially for the PlayStation e game console. However, this graphic pipeline uses, for a masking calculation step, additional preprogrammed geometries in order to efficiently calculate the masked objects, which implies the implementation of a specific processing for each new scene. The object of the present invention is to solve all or some of the disadvantages mentioned above by proposing a hybrid graphic pipeline that can efficiently exploit the processors available on a current computer, namely the so-called GPU graphics processor and the so-called CPU central processor. so as to allow real-time rendering of complex scenes, for example in nomadic uses, even on low-end equipment. For this purpose, the present invention relates to a real-time image rendering method from a three-dimensional virtual scene comprising a set of objects, comprising: a step of preparing at least part of the objects of the scene as a first set of polygons; a first step of rasterizing the first set of polygons at a first resolution less than a second final resolution of the images so as to obtain a depth buffer; information being collected during the first rasterization step corresponding to the identifiers of the visible polygons of the first set of polygons, that is to say to the polygons at the origin of at least one fragment of the first depth buffer, so defining a second set of polygons among the first set of polygons; and a second rasterization step of the second set of polygons at the second final resolution of the images. This arrangement makes it possible to perform the calculations on a set of minimized polygons and sufficient to achieve quality rendering, which optimizes the machine resources, in particular by balancing the data processing loads between the CPU central processor and the GPU graphics processor according to the present invention. the specialization of each processor. Moreover, the performance of a rendering process is all the more important as it makes it possible to reduce the amount of computation data and thus of operations to be executed by the GPU graphics processor. According to one aspect of the invention, the method comprises a masking calculation step for determining the visible objects as a function of the visibility of a volume encompassing the object. According to one aspect of the invention, the visible or invisible state of an object is determined for a given image based on the rendering of a previous image. These provisions make it possible to limit the processing to be performed during the first rasterization step by performing a simple prior test on the boxes enclosing the objects, in particular by maintaining a list or a buffer of visible objects and invisible objects, and by performing, for the invisible objects, a test by comparing a volume encompassing the object, and in particular an oriented bounding box, with respect to the depth buffer established for the image I based on the visible objects. According to one aspect of the invention, the preparation step comprises a step of dividing the image into tiles, a subset of polygons of the first set being defined corresponding to the polygons likely to be visible in this tile.

According to one aspect of the invention, in the first rasterization step: the calculation of the depth buffer includes the determination for at least one tile of a maximum depth value when the tile is filled. - we compare the maximum depth of the tile with the minimum depth of a polygon, the polygon is not taken into account if its minimum depth is greater than the maximum depth of the tile. These arrangements make it possible to implement a depth test by tile and to avoid the rasterization of triangles having a minimum depth greater than the maximum depth of a filled tile. According to one aspect of the invention, the polygons of the first set or subsets of the first set are classified by increasing value of minimum depth. These provisions make it possible in particular to easily determine all of the triangles which have a minimum depth greater than the maximum depth of a tile and to stop the rasterization of a given tile on the basis of this criterion. According to one aspect of the invention, the first rasterization step comprises a step of comparing an image fragment with the surrounding volume of the projection of a polygon of the first set in the image space. According to one aspect of the invention, a comparison step between the fragment and the edges of the polygon under consideration is performed in the case where the comparison between a fragment and the projection of the encompassing volume of a polygon of the first set indicates that the fragment is at least partially in the enclosing volume of the projection of the polygon in the space of the images. According to one aspect of the invention, the method comprises, between the first rasterization step and the second rasterization step, an intermediate step of filling the image intended to remove the areas of the image without visible geometry. According to one aspect of the invention, the filling step corresponds to an expansion of the polygons of the second set in the space of the images. The present invention also relates to a computer program product comprising code instructions arranged to implement the steps of a method as described above and a data carrier comprising the code instructions of such a program product. computer. The present invention also relates to a computer system in which the central processor and / or the graphics processor are arranged to implement instructions of the computer program product. According to one aspect of the invention, the central processor is arranged to implement instructions of a computer program product corresponding to the preparation step and the first rasterization step of the method, and the graphics processor is arranged to implement instructions of a computer program product corresponding to the second process rasterization step. According to one aspect of the invention, the central processor is arranged to implement instructions of a computer program product corresponding to the process hole filling step. According to one aspect of the invention, the graphics processor is arranged to implement instructions of a computer program product corresponding to the process hole filling step. According to one aspect of the invention, the central processor is arranged to implement instructions of a computer program product corresponding to the masking calculation step of the method. In any case, the invention will be better understood with the aid of the description which follows, with reference to the appended schematic drawing showing, by way of non-limiting example, the steps of a method according to the invention. Figure 1 shows a flow chart of a preferred implementation of the method according to the invention. FIG. 2 illustrates a step of preparation of the method according to the implementation of the method illustrated in FIG. 1. FIG. 3 illustrates a first step of rasterization of the method according to the implementation of the method illustrated in FIG. 4 is an illustration of the organization of the data structures representing the depth buffer used by the method illustrated in FIG. 1. FIG. 5 is a representation of an object in the scene.

FIG. 6 illustrates an operation performed during the first rasterization step. FIG. 7 illustrates an intermediate step of the method according to the implementation of the method illustrated in FIG.

FIG. 8 is an illustration helping to understand the intermediate step of the method illustrated in FIG. 7. FIG. 9 illustrates a first system configuration implementing the method according to the invention. FIG. 10 illustrates a second system configuration implementing the method according to the invention. FIG. 11 illustrates a third system configuration implementing the method according to the invention. As illustrated in FIG. 1, according to one embodiment, the real-time I-image rendering method from a three-dimensional virtual Sc scene comprises: a preparation step If from least part of the objects 0 of the scene Sc as a first set of polygons EP1. a first rasterization step S2 at an intermediate resolution Ri less than the final resolution Rf 20; an intermediate hole filling step S3; and then a second rasterization step S4. In addition, a step of calculating a masking S5 to identify non-visible objects in an image based on the objects present in the previous image. In the remainder of the description, we consider that the polygons are triangles formed from three vertices, the triangle being the most used type of primitive for rendering two-dimensional images of images from objects 0 of a Sc scene in three dimensions. However, other types of polygons could be envisaged according to alternative embodiments of the invention. Each of the objects 0 of the scene is thus formed by a plurality of triangles, each triangle T being composed of three vertices V and associated with a normal N on the surface of the triangle T representative of the orientation of the triangle T in the scene Sc.

Each of the vertex V of the triangles T of an object 0 has coordinates that can be expressed in a frame specific to the object or in an orthonormal frame having for reference the scene Sc. Moreover, we consider here that the term image I describes generically the successive images of the scene Sc defined by the rendering method from the characteristics of a camera C. The process steps are described below. As illustrated in FIG. 2, the preparation step Si comprises a plurality of sub-steps. These sub-steps are detailed below. During a first substep SS1, the coordinates of each vertex V are transformed from the orthonormal space having for reference the scene Sc to the orthonormal space of coordinates x, y, z having for reference the position of the camera C , then projected on the (x, y) plane of the image I while also preserving a depth information z associated with the vertex V in the reference frame having for reference the camera C. Each vertex V projected onto the plane (x, y ) is expressed by its coordinates x, y, z. This first sub-step SS1 therefore makes it possible, after this transformation, to obtain a first list of triangles T of objects 0, the vertices V of these triangles T being expressed in the reference frame having the image I as the reference with coordinates x, y corresponding to a position in the image I and a depth information z. During a second substep SS2, the field of view or frustum of the camera is defined by a pyramidal trunk delimited by a near plane NearP and a farP FarP. In this second substep SS2, the potentially visible objects 0, that is to say which are wholly or partly in the field of vision or frustum of the camera C, are marked as such. During the rendering of an image I of the scene Sc, all the triangles T of the objects 0 of the scene Sc which are outside the field of view or frustum are stored separately for later use, in particular during a masking calculation step S5 described later in the text. In a third substep SS3, the triangles T which do not face the camera C are eliminated, the orientation of the triangles T with respect to the camera C being deduced from the normal N on the surface of each triangle T in the reference frame having for reference camera C. In a fourth substep SS4, the so-called degenerate triangles T, that is to say having surfaces that are zero or too flattened in the reference frame having as their reference image I are they also eliminated. These triangles T are eliminated by converting the real coordinates (x, y) of each vertex V into integer coordinates of the nearest fragment Fr of the image I, and by deleting after this conversion the triangles T of zero or too weak surface. A second list of triangles is obtained at the end of the fourth substep containing the triangles of the first list which have not been eliminated in substeps SS2 to SS4. In a fifth substep SS5, the image I is also divided into a plurality of tiles T1 of the same determined resolution. Each tile T1 is defined in the image I by a rectangle of coordinates xmin, ymin and xmax, ymax. Each tile T1 can thus be covered in whole or in part by triangles T of the second list. Likewise, a triangle T of the second list may cover one or more tiles T1. In order to easily determine whether all or part of a triangle T may be contained in a tile T1, a bounding surface TSB of the projection of The x, y coordinates of the bounding surface TSB are compared with the coordinates xmin, ymin and xmax, ymax of the tile T1 to determine the potentially visible triangle portions in the tile. This calculation can be carried out simultaneously for all the tiles T1 of the image I. At the end of the fifth sub-step SS5, a first set EP1 of triangles T potentially visible in the image I with subsets SEP1 of triangles T likely to be visible in a tile Tl determined from the image I. However, the triangles T of the subsets SEP1 are only potentially visible in the image I.

In fact, the triangles T of the subassemblies SEP1 may be located at different depths z relative to the camera C and have overlaps. A part of a triangle T at a lower depth z can therefore mask a portion of triangle T at a greater depth z if their projection in a tile TI of the image I comprises a covering. In order to determine the visible portions of the triangles T for each fragment Fr of the image, a first rasterization step S2 is performed on the first set EP1 of triangles T at a first resolution Ri less than the final resolution Rf of the image I so as to obtain a depth buffer ZB1 and an identification buffer of the visible triangles. By way of example, the depth buffer ZB1 may have a resolution two times smaller than that of the final image I, that is to say the width and height of which are half the width and the width. height, respectively, of the final image I. The details of the sub-steps of this first rasterization step is shown in FIG. 3. In order to increase the memory coherence, and since the rasterization is performed by parallel processing of fragments With 4X4 resolution at a time for efficient use of SIMS functions, the ZB1 depth buffer is stored in memory in a non-linear manner. As shown in FIG. 4, the depth buffer ZB1 is organized by TI tiles, and comprises several levels of TI tiles. By way of example, the following levels can be defined: a first level of TI tiles comprising 64X64 fragments of the depth buffer ZB1; a second level of TI tiles comprising 16X16 fragments of the depth buffer ZB1; a third level; of TI tiles including 4X4 fragments of depth buffer ZB1. It should be noted that TI tiles of the same level do not overlap, which allows for parallel processing by separate threads. On the other hand, in the case where a triangle covers several tiles T1, only the triangle portion T covering the tile T1 is rasterized.

The rasterization is carried out as follows. For each tile Tl of the first level, consider each triangle T of the subassembly SEP projecting on said tile. For each triangle T, all the fragments Fr of the tile lying inside the bounding surface TSB of the triangle T are compared to the equations of the sides of the triangle T. For each fragment Fr lying inside the triangle T , a depth value is stored in the first depth buffer ZB1, if the current depth of the considered triangle is less than a depth value possibly already stored corresponding to the same Fragment. As illustrated in FIG. 3, if a depth value is updated in the depth buffer ZB1 for a given Fragment Fr, an identification information ID of the triangle T corresponding to this update 15 is recorded in a buffer. IDB identification for the same Fragment Fr. It should be noted that the identification buffer may have a structure similar to that of the depth buffer illustrated in Figure 4 and described above. Thus, during this first rasterization step S2, identification information ID is collected, in the identification buffer IDB, these identification information corresponding to the identifiers of the visible triangles T, that is to say triangles T defining at least one the value of a fragment or pixel of the first depth buffer ZB1. According to one embodiment, the rasterization is performed by processing T1 tiles of 4X4 Fr fragments at a time, each group of four Fr fragments being processed in parallel by a single "thread" or processing wire, using SIMD functions allowing four operations to be performed simultaneously during a CPU central processor cycle. In order to minimize the calculations to be made to determine the visible triangles T for each fragment Fr of the depth buffer ZB1, a tile depth test is set up. To allow this test to be performed, the triangles T of each subset SEP1 of each tile T1 are classified according to their minimum depth zmin.

During rasterization, when a TI tile is filled, that is when the depth buffer ZB1 comprises depth information for each fragment Fr of said tile T1, a value corresponding to the maximum depth zmax present in the buffer depth for Fr fragments of the tile is determined. Subsequently, this value zmax is compared with the zmin value of each triangle T that has not yet been processed. If the zmin value is less than the zmax value, a triangle rasterization is performed, which may lead to an update of the zmax value.

If the zmin value of the triangle is greater than the zmax value of the tile, then the considered triangle and all subsequent triangles can be ignored for rasterization of the considered tile. Figures 6a and 6b illustrate an exemplary implementation of rasterization with an identification buffer and a tile depth test, limited to a two-dimensional case to simplify understanding. Thus, the tiles T1 as the triangles T are represented by segments. The rasterization of a tile T1 is done by considering the triangles T projecting on this tile starting with the triangle T having the minimum depth zmin going to the triangle having the highest minimum depth, the depth of a triangle. T being determined from the minimum depth zmin of one of these vertex V among the three vertices V composing the triangle T. So we first project the triangle Ti with the lowest zmin. This triangle modifies the depth buffer value for fragments Fr of the tile T1, and thus the triangle identifier IDT1 is entered in the IDB ID buffer for the corresponding Fr fragments. We then project the triangle T2 with a zmin2> zmin1. This triangle modifies the value of the depth buffer for Fragments Fr of the tile T1, and thus the triangle identifier IDT2 is entered in the IDB ID buffer 30 for the corresponding Fr fragments. Moreover, it appears that the tile T1 is filled following the rasterization of the two triangles Ti and T2, a depth information being present in the depth buffer for each fragment. A first value of zmax is therefore determined for the tile. A third triangle T3 is then considered, whose value zmin3 is greater than that of the triangles Ti and T2, but less than zmax. This triangle modifies the value of the depth buffer for Fragments Fr of the tile T1, and thus the identifier IDT3 of this triangle is entered in the IDB ID buffer for the corresponding Fragments Fr. In the case presented, the rasterization of T3 modifies the value of 5 zmax, which is represented by a value zmax '. We then consider a triangle T4 whose value zmin4 is greater than zmax '. The rasterization of this triangle, like all the others with values of zm in greater than zmax ', can be eliminated because these are necessarily hidden by another triangle T. Thus, the IDB buffer will contain only identifiersIDT1, IDT2 and IDT3 of triangles T1, T2 and T3. As illustrated in FIG. 3, after having rasterized all the triangles T of a tile TI, a substep of the IDB ID buffer 15 is carried out for each tile TI to define a new buffer by tile of the visible triangles. for this tile. Then, all the visible triangle buffers of the set of TI tiles are merged and the duplicates corresponding to triangles visible on several TI tiles are deleted so as to define a second buffer or set of EP2 visible polygons or triangles for the image I, which corresponds to the result of the first rasterization step. It should be noted that the triangles T of the second set of triangles EP2 can also be classified according to their belonging to an object O. All the objects which do not contain visible triangles are marked as invisible and will be processed during the step of S5 masking. Following the first rasterization step, a step of filling the holes S3 can be performed. Indeed, since the processing performed during the first rasterization step 1 is performed at a lower resolution Ri, it is possible for some small triangles T to be considered degenerate during the preliminary step 51 and thus removed during of the first rasterization step S2, which can generate holes in the image I of final resolution Rf higher during the second rasterization step S4. The step 53 of filling the holes to remove the zones of the image I without visible geometry or holes is illustrated in FIG.

This filling step consists in dilating all the visible triangles composing the image I. The dilation in itself is carried out using, for example, the method described in the article [Hasselgren, J., Akenine-Moller, T., and Ohlsson, L. 2005. Conservative Rasterization. In GPU Gems 2, M. Pharr, Ed. Addison-Wesley, 677-6901. As illustrated in FIG. 8, the algorithm used calculates quadrants to which each edge normal belongs and, depending on their position, determines one, two, or three vertices to be inserted for each original vertex. The expansion factor depends on the resolution ratio between the depth buffer ZB1 used in the first rasterization step and the depth buffer used in the second rasterization step. For example, for a depth tap ZB1 that has half the size of the width and half the height of the final image, an expansion of 0.4 of the pixel size is sufficient to cover all holes and does not create significant visual artifacts in motion. As illustrated in FIG. 7, the new coordinates of the projections of the vertices V of the visible triangles T of the image I are stored in a vertex buffer VB3 for each of the objects 0 of the image I.

Following this hole filling step, the second rasterization step S4 at a second resolution corresponding to the final resolution Rf of the image I based on the second triangle set EP2 is performed. Thus, this rasterization step S4 is performed with a minimum of triangles T to rasterize, which increases the rendering speed of the image I.

Subsequently, the S5 masking calculation step for determining the potentially visible objects 0 to be taken into account in the rendering of the next image is performed. This step S5 is based on the temporal coherence between an image I and the previous image I and assumes that an object 30 visible during a rendering n will certainly be visible during an n + 1 rendering. The same goes for the 0 not visible objects. Thus, the visible or invisible state of an object 0 is determined for a given image I when rendering a previous image I. However, for the invisible objects, a test is performed by comparing an OBB volume encompassing the object 0 and in particular an oriented bounding box, as illustrated in FIG. 5, with respect to the depth buffer established for the image I on the base of visible objects. Thus, the whole object is not tested. This approach provides a conservative result with respect to visibility, while requiring the rasterization of a small number of triangles. The process substeps are similar to those presented in the first rasterization step. However, the rasterization method using the bounding boxes of objects does not alter the depth buffer, but limits itself to marking the object as potentially visible if at least one of the depth buffer fragments has a depth greater than the depth of the depth buffer. bounding box. This allows in particular to stop the rasterization of an object as soon as a fragment fulfills the condition mentioned above. Since the depth buffer ZB1 is already filled when the comparison with the OBB bounding box of the objects 0 begins, it is possible to implement the tile depth test T1 described with reference to the first rasterization step. If an object is marked as potentially visible during this object masking calculation step, it will be rendered at the next image. It is possible that this shift of an image produces a "pop-up" artifact. However, this artifact corresponds to the duration of the rendering of an image I and concerns only the first visible pixels of an object. By maintaining a refresh rate proportional to the speed of the camera, this artifact remains almost imperceptible. It is also possible to correct it by performing a preparation step and a specific first rasterization step for the new visible objects 0. We will now describe three examples of hardware configurations on which the use of the previously described method has been tested. According to a first configuration illustrated in FIG. 9, a device used for aircraft maintenance is considered. For this application, the same procedures have been used for many years on the same hardware to avoid unpredictable problems. For this reason, relatively old and low performance computers are used. For example, an Intel® GMA type architecture is used, and more specifically an Intel @ 965 Express Chipset Family Mobile with an Intel @ Core2 duo T8300 2.4 GHz. In this configuration, the central processor executes two "threads" simultaneously during a CPU core cycle. All the steps of the method are processed by the central processor CPU except the second rasterization step S4 which is processed by the graphics processor GPU. According to a second configuration illustrated in FIG. 10, devices intended for training new recruits at a nuclear power plant site are considered. For this application, recent lightweight computers for increased mobility in hard-to-reach areas are used. For example, the target architecture considered is a 1.6Ghz Intel Core7 CPU processor with a graphics GPU on Intel HD Graphics 4000. The central processor CPU offers a hyperthreading technology that allows to use 8 threads in parallel. In this case, all the steps are processed by the central processor CPU except the second rasterization step S4 which is processed by the graphics processor GPU. According to a third configuration illustrated in FIG. 11, devices intended for the maintenance of automobiles are considered. For this application, less portable but better performing computers are used, as the tasks to be performed require less mobility and easier access. For example, the target architecture under consideration is an Intel @ Core7 1.6Ghz CPU processor with a GPU graphics processor ATI Mobility Radeon @ HD 4830 Series. For this architecture, the central processor can realize eight "threads" simultaneously during a CPU core cycle. However, the speed of the process becomes limited by the capabilities of the CPU given the capabilities of the GPU. Thus, to enable efficient load balancing, some parts of the process are delegated to the GPU. In particular, the step of filling the holes S3 and the second rasterization step S4 which are processed by the graphics processor GPU. It will be understood that the various implementations of the rendering method detailed above are only examples of implementations of the invention as defined by the appended claims. Variations of these different implementations can be envisaged and the various embodiments described can be easily combined by those skilled in the art.

Claims (13)

REVENDICATIONS1. A method for rendering images (I) in real time from a three-dimensional virtual scene (Sc) comprising a set of objects (0), comprising: a preparation step (Si) of at least part of the objects of the scene as a first set of polygons (EP1); a first rasterization step (S2) of the first set of polygons (EP1) at a first resolution (Ri) of less than a second final resolution (Rf) of the images (I) so as to obtain a depth buffer (ZB1) information (IDB) being collected during the first rasterization step (S2) corresponding to the identifiers (PID) of the visible polygons of the first set of polygons (EP1), that is to say to the polygons at the origin of at least one fragment (Fr) of the first depth buffer (ZB1), so as to define a second set of polygons (EP2) among the first set of polygons (EP1); and a second rasterization step (S4) of the second set of polygons (EP2) at the second final resolution (Rf) of the images (I). An image rendering method (I) according to claim 1, comprising a masking calculation step (S5) for determining the visible objects according to the visibility of a volume (OBB) encompassing the object ( 0). The method of claim 2, wherein the visible or invisible state of an object (0) is determined for a given image (I) based on the rendering of a previous image (I). 4. Method according to one of the preceding claims, wherein the preparation step (Si) comprises a step of dividing the image (I) into tiles (Tl), a subset (SEP1) of polygons of the first set (EP1) being defined corresponding to the polygons likely to be visible in each tile (Tl). 5. Method according to one of the preceding claims, wherein during the first rasterization step (S1): the calculation of the depth buffer (ZB1) includes the determination for at least one tile (Tl) of a value of maximum depth (Zmax) 35 when the tile (Tl) is filled.- the maximum depth (Zmax) of the tile (Tl) is compared with the minimum depth of a polygon (Zmin), the polygon not being taken into account. count if its minimum depth (Zmin) is greater than the maximum depth (Zmax) of the tile (Tl). 6. Method according to one of the preceding claims, wherein the polygons of the first set (EP1) or subsets of the first set are classified by increasing value of minimum depth (Zmin). 7. Method according to one of the preceding claims, wherein the first rasterization step (S2) comprises a comparison step 10 between a fragment (Fr) of image (I) and the enclosing volume (PBB) of the projection of a polygon of the first set (EP1) in the space of the images (I). The method of claim 7, wherein a step of comparing the fragment (Fr) with the edges of the polygon of interest is performed in the case where the comparison between a fragment (Fr) and the projection of the bounding volume (PBB). a polygon of the first set indicates that the fragment (Fr) is at least partially in the enclosing volume (PBB) of the projection of the polygon in the space of the images (I). 9. Method according to one of the preceding claims, comprising, between the first rasterization step (S2) and the second rasterization step (S4) an intermediate step (S3) for filling the image intended to eliminate the zones of the image without visible geometry. 10. The method of claim 9, wherein the filling step (S3) corresponds to an expansion of the polygons of the second set (EP2) in the space of the images (I). 25 11. Computer program product comprising code instructions arranged to implement the steps of a method according to one of claims 1 to 10. A data carrier comprising the code instructions of a computer program product according to claim 11. A computer system comprising a central processor (CPU) and a graphics processor (GPU) in which the central processor and / or the graphics processor (GPU) are arranged to implement instructions of a computer program product according to the claim 11.