EP3824444A1 - Methode de monte carlo et reseau de neurones dans un procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondants - Google Patents
Methode de monte carlo et reseau de neurones dans un procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondantsInfo
- Publication number
- EP3824444A1 EP3824444A1 EP19790621.7A EP19790621A EP3824444A1 EP 3824444 A1 EP3824444 A1 EP 3824444A1 EP 19790621 A EP19790621 A EP 19790621A EP 3824444 A1 EP3824444 A1 EP 3824444A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- estimator
- neuron
- contribution function
- contribution
- parameter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T15/00—Three-dimensional [3D] image rendering
- G06T15/06—Ray-tracing
Definitions
- the present invention relates to a method for rendering a three-dimensional scene and corresponding computer programs and device.
- the invention applies more particularly to a method for rendering a three-dimensional scene, comprising, for each of at least one pixel of an image representing a view of the three-dimensional scene:
- the estimator comprising a Monte Carlo estimator using the first light paths.
- the Monte Carlo estimator shows a variance that decreases with the number of random light paths used.
- This variance is responsible for spatial and temporal instabilities during rendering. Indeed, for spatial instability, two pixels whose values should be identical or close (for example two adjacent pixels in the image) are estimated from random light paths which can be different, so that the estimates can be different, especially as the variance is high.
- temporal instability in the case of a video, the same pixel should often have the same value or values close to one image to the next. However, since its value is each time estimated from random light paths which may be different from one image to the next, its value can vary from one image to the next, the more so as the variance is high.
- a method of rendering a three-dimensional scene comprising, for each of at least one pixel of an image representing a view of the three-dimensional scene:
- the estimator comprising a Monte Carlo estimator using the first light paths
- the approximate contribution function comprising a neural network comprising:
- the presence in the estimator of the invention, denoted ⁇ , of the approximate contributions makes it possible to obtain a reduced variance compared to the variance of the Monte Carlo estimator alone.
- the use of the ⁇ estimator therefore makes it possible to obtain greater temporal and spatial stability during rendering than the use of the Monte Carlo estimator alone.
- the choice of the contribution function approximated in the form of a neural network comprising a single hidden layer and in which only the weights of the output connections (between the hidden layer and the output layer) are parameterized facilitates the determination of the parameter (s) so that the function of approximate contribution is close to the contribution function.
- the method also comprises a second random selection, independent of the first selection, of second light paths, and the parameter (s) of the approximate contribution function are determined from the second light paths.
- the determination of the parameter or parameters is carried out so as to minimize a distance between the approximate contribution function and the contribution function.
- the distance uses the second light paths.
- the distance comprises the sum of the squares of the differences between the approximate contribution function and the contribution function for the second light paths.
- the distance furthermore comprises a regularization parameter and the determination of the parameter (s) comprises the determination of the regularization parameter, then the determination of the parameter (s) using the determined regularization parameter.
- the estimator can be written as a matrix A [I m - M] B, where I m is an identity matrix of size m, A and B are vectors and M is a parameter-dependent matrix of regularization, and the parameter of regularization is determined so that the maximum eigenvalue of M presents at most a predefined deviation with 1.
- the weights associated with the connections between the neuron (s) of the input layer and the neuron (s) of the hidden layer are determined randomly.
- the estimator further includes an expectation of the approximate contribution function.
- a computer program downloadable from a communication network and / or recorded on a computer-readable medium and / or executable by a processor comprising instructions for executing the steps of a method for rendering a three-dimensional scene according to the invention, when said program is executed on a computer.
- a device for rendering a three-dimensional scene comprising, for each of at least one pixel of an image representing a view of the three-dimensional scene:
- the estimator comprising a Monte Carlo estimator using the first light paths
- the approximate contribution function means for determining at least one parameter of an approximate contribution function, so that the approximate contribution function gives a contribution of a light path to the value of the pixel (p) close to the contribution given by the contribution function, the approximate contribution function comprising a neural network comprising:
- FIG. 1 schematically represents the general structure of a three-dimensional scene rendering device, according to an embodiment of the invention
- FIG. 2 illustrates an example of a three-dimensional scene
- FIG. 3 illustrates the successive stages of a three-dimensional scene rendering method, according to an embodiment of the invention.
- the device 100 is a computer device comprising a central processing unit 102, such as a microprocessor, and a memory 104 in which a computer program 106 is recorded, data 108 describing a three-dimensional scene in a virtual three-dimensional space, and data 1 10 representing parameters of a desired view of this three-dimensional scene, this view being intended to be produced by a (virtual) camera.
- View settings typically include settings extrinsic to the camera (for example, camera position and orientation in three-dimensional space), as well as intrinsic settings for the camera (for example, focal length, center pixel , optical distortion, etc. of the camera).
- the computer program 106 includes instructions for the carrying out of this step by the central processing unit 102.
- these instructions can be grouped or separated in any possible combination into one or more computer programs. Their distinction is also purely functional. Finally, their functions could also be at least partly micro-programmed or micro-wired in dedicated integrated circuits.
- the computer device 100 including the processing unit 102 could be replaced by an electronic device composed solely of digital circuits (without computer program) for carrying out the same actions.
- scene 200 has a single object, a cube 202, and a single light source 204, both having predefined positions in a virtual three-dimensional space.
- FIG. 2 An image 206 intended to represent a view of the scene 200 is also illustrated in FIG. 2.
- the image 206 comprises a plurality of pixels, among which a pixel p is represented.
- each light source in the scene emits light which can reach the image either directly or indirectly after interacting with the objects in the scene, for example after reflection on one or more objects or after deflection by one or more objects.
- light paths 208, 210 start from the light source 204 and reach the pixel p.
- a three-dimensional scene rendering method 300 will now be described.
- the method 300 renders the scene 200 in FIG. 2.
- the method 300 could be used to render the other three-dimensional scenes.
- the method 300 as described is implemented by the device 100 of FIG. 1, but it could be implemented by other means.
- the program 106 recovers the scene 200 to be rendered and the parameters of the desired view, for example by reading the data 108, 110 recorded in the memory 104.
- the program 106 renders the scene 200 in order to provide the image 206 representing a view of this scene 200.
- step 304 For each pixel p of the image 206, step 304 comprises the following steps 306 and 308.
- a function ⁇ p (x) is determined from the scene 200, from the parameters of the desired view and from the pixel p considered.
- the function f (c) is a function intended to receive as an input a light path x starting from a light source of the scene 200 (for example, the light source 204) and arriving at the pixel p, and to provide a contribution of this light path x to the value of the pixel p.
- the light paths 208, 210 of Figure 2 are examples of such light paths.
- the function f will thus be called hereinafter “contribution function”.
- ⁇ is the set of possible light paths arriving at pixel p and p (dx) a measure associated with the set ®.
- the ® set includes an infinite number of light paths.
- the contribution function ⁇ p (x) is generally not analytically integrable.
- the integral I p is very difficult to calculate exactly and it is necessary to go through an estimator of this integral I p .
- a value of the pixel p is calculated from an estimator ⁇ of the integral I p .
- This estimator ⁇ is equal to the sum, on the one hand, of a Monte Carlo estimator E of the integral I p and, on the other hand, of a stabilizing term T intended, as will be explained later to improve the temporal and spatial stability of the estimate:
- Statistical sampling uses, for example, the known method of ray tracing.
- the Monte Carlo estimator E performs an arithmetic mean of the contributions of the light paths (x x j ) and is thus given by the equation:
- stabilizer T uses at least part of the light paths ⁇ x j ) used by the Monte Carlo estimator E, so that the term stabilizer T has some correlation c (T, E) with the Monte Carlo estimator Carlo E.
- the term stabilizer T is generally given by the following equation:
- T a (-T 1 + m (T 1 ))
- c (£ ' , 7' 1 ) is the correlation between E and 7 and v (TJ is the variance of 7.
- a is equal to 10%, preferably to 1%, more preferably
- 7 has the following form:
- p m is an analytically integrable function intended to be close to the contribution function y so as to give an approximate contribution of a light path to the value of the pixel p.
- the function p m will thus be called hereinafter “approximate contribution function”.
- p ( ⁇ p m ) is the expectation of the approximate contribution function f th .
- the term 7 is equal to an average (arithmetic average in the example described) of the approximate contributions of the light paths ⁇ c ; ).
- stabilizer T therefore has zero expectation:
- the expectation of the estimator ⁇ is equal to the expectation of the Monte Carlo estimator E, so that the estimate given by the estimator ⁇ is, on average, a good estimate of the integral I p .
- the approximate contribution function f th is a parameterized function comprising at least one parameter, this or these parameters being intended to be determined so as to minimize a distance between the approximate contribution function f th and the contribution function f.
- the approximate contribution function f th comprises a neural network, comprising neurons and connections between these neurons, these connections having respective weights, at least part of which depends on the parameter (s) of the approximate contribution function f th .
- the approximate contribution function f th is formed by a neural network called “random characteristic maps” (from the English “random feature maps”).
- a neural network includes an input layer of neuron (s), a single hidden layer of neuron (s) and an output layer of neuron (s).
- the neurons of the input layer and those of the hidden layer are connected by connections to which weights are respectively associated.
- the neurons of the hidden layer and of the output layer are connected by connections to which weights are respectively associated, the latter depending on the parameter or parameters of the approximate contribution function f gh .
- the activation function s can be any derivable function. In the example described, the activation function s is given by the equation:
- the number N of neurons in the hidden layer is predefined and chosen as a function of the targeted performance.
- Step 308 includes the following steps 310 to 320.
- the weights are calculated randomly.
- the m light paths are
- the light paths ⁇ 3 ⁇ 4] are independent and identically distributed variables.
- the light paths ⁇ 3 ⁇ 4] are for example selected by the method, known per se, of ray-tracing.
- the parameters X ⁇ are determined so as to minimize a distance between the approximate contribution function f th and the contribution function f according to a predefined distance concept.
- the method of least squares is used, so that the distance used comprises the sum of the squares of the differences between the approximate contribution function f th and the contribution function y for the light paths [x j ] , and a regularization term.
- the parameters X j are determined so as to minimize the distance S given by the following equation:
- F-. A ® - (F ⁇ - - Z T ZA) (f ( TM ) - - Z T ZA) + gZ t A t ZA is a convex and differentiable function in A. Its minimum can therefore be obtained by canceling its gradient with respect to to L:
- step 314 of determining the parameters Z j first comprises the determination of the regularization parameter g.
- the regularization parameter g can for example be found using the method, known per se, of cross-validation (“cross-validation” in English).
- estimator ⁇ can be written as a matrix:
- the determined regularization parameter g is that for which the maximum eigenvalue of M is as close as possible to 1.
- a regularization parameter g for which the maximum eigenvalue of M has at most one predefined deviation with 1 is chosen.
- the predefined difference is for example at most 10 -5 , preferably 10 -7 , for example 10 -9 .
- This alternative method makes it possible to find a regularization parameter y close to that which would have been found by the cross validation method, but at a lower cost.
- the parameters ⁇ j are determined using the regularization parameter y determined, using the formula (explained above):
- the matrix ⁇ Z T ZA + m 2 yI m is invertible for y> 0.
- the inversion of the matrix M would have a complexity in o (m 3 ) where m is the number of prints.
- N is less than m, it is possible thanks to the invention to exploit the presence of Z T Z to reduce the complexity of the inversion of the matrix M, and therefore to facilitate the determination of the parameters Z j .
- the expectation m (p m ) is calculated.
- the weights ⁇ 1 ⁇ 2 follow a normal law of zero mean and matrix of is given
- f m is the characteristic function associated with the expectation m.
- the paths of light ⁇ x are for example selected by the method, known per se, of ray-tracing.
- the estimator ⁇ is evaluated from the light paths ⁇ ;) selected in step 318:
- the method 300 then returns to step 306 for a new pixel p of the image 206, until the values of all the pixels of the image 206 have been calculated.
- the ⁇ estimator is of the form near
Landscapes
- Engineering & Computer Science (AREA)
- Computer Graphics (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Generation (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| FR1856732A FR3084192B1 (fr) | 2018-07-19 | 2018-07-19 | Procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondants |
| PCT/FR2019/051691 WO2020016501A1 (fr) | 2018-07-19 | 2019-07-08 | Methode de monte carlo et reseau de neurones dans un procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondants |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP3824444A1 true EP3824444A1 (fr) | 2021-05-26 |
Family
ID=63491778
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP19790621.7A Withdrawn EP3824444A1 (fr) | 2018-07-19 | 2019-07-08 | Methode de monte carlo et reseau de neurones dans un procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondants |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP3824444A1 (fr) |
| FR (1) | FR3084192B1 (fr) |
| WO (1) | WO2020016501A1 (fr) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP7511230B2 (ja) * | 2020-08-20 | 2024-07-05 | 国立大学法人 東京大学 | 量子回路生成装置、量子回路生成方法及び量子回路生成プログラム |
| CN118484666B (zh) * | 2024-07-16 | 2024-10-01 | 浙江兴创新能源有限公司 | 面向源网荷多元应用的储能电站评估方法及系统 |
-
2018
- 2018-07-19 FR FR1856732A patent/FR3084192B1/fr not_active Expired - Fee Related
-
2019
- 2019-07-08 WO PCT/FR2019/051691 patent/WO2020016501A1/fr not_active Ceased
- 2019-07-08 EP EP19790621.7A patent/EP3824444A1/fr not_active Withdrawn
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020016501A8 (fr) | 2021-10-07 |
| FR3084192A1 (fr) | 2020-01-24 |
| FR3084192B1 (fr) | 2020-10-02 |
| WO2020016501A1 (fr) | 2020-01-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11398015B2 (en) | Iterative image inpainting with confidence feedback | |
| EP1410331B1 (fr) | Procede et systeme pour modifier une image numerique en prenant en compte son bruit | |
| US8755628B2 (en) | Image de-hazing by solving transmission value | |
| US20160098820A1 (en) | System for robust denoising of images | |
| EP3646121B1 (fr) | Procede de generation numerique d'un hologramme, dispositif, equipement terminal, systeme et programme d'ordinateur associes | |
| WO2001099052A1 (fr) | Raffinement d'un maillage triangulaire en trois dimensions | |
| FR3047586A1 (fr) | Procede et dispositif d'anonymisation de donnees stockees dans une base de donnees | |
| WO2020016501A1 (fr) | Methode de monte carlo et reseau de neurones dans un procede de rendu d'une scene tridimensionnelle, programme d'ordinateur et dispositif correspondants | |
| EP2294552B1 (fr) | Dispositif de traitement d'images amélioré | |
| EP2391989B1 (fr) | Procede de representation d'un materiau | |
| EP0410826B1 (fr) | Procédé itératif d'estimation de mouvement, entre une image de référence et une image courante, et dispositif pour la mise en oeuvre de ce procédé | |
| EP2980529B1 (fr) | Procede d'estimation de la distance d'un objet a un systeme d'imagerie | |
| FR3067482A1 (fr) | Dispositif de caracterisation et/ou de modelisation de temps d'execution pire-cas | |
| FR3156963A1 (fr) | Méthode de super-résolution d’image présentant une robustesse universelle | |
| Ranjan et al. | VAEWGAN-NCO in image deblurring framework using variational autoencoders and Wasserstein generative adversarial network | |
| EP3918576A1 (fr) | Procédé dynamique d'imagerie tridimensionnelle | |
| EP3797509B1 (fr) | Traitement d'un bruit impulsionnel dans une séquence vidéo | |
| EP4150574B1 (fr) | Procédé de traitement d'images | |
| EP4300129B1 (fr) | Procédé de regroupement de descriptions de formes d'ondes | |
| FR3110268A1 (fr) | Procédés d’utilisation sécurisée d’un premier réseau de neurones sur une donnée d’entrée, et d’apprentissage de paramètres d’un deuxième réseau de neurones | |
| EP4078959B1 (fr) | Procede d'encodage d'une image numerique en vue de sa compression | |
| EP4287122B1 (fr) | Procede de generation de modele de representation de relief par photogrammetrie | |
| WO2026087114A1 (fr) | Procédé et dispositif de représentation de contenu volumétrique | |
| EP4163866A1 (fr) | Procédé, programme d'ordinateur et dispositif de traitement d'images par extraction de composante(s) principale(s) d'une représentation d'images | |
| WO2026087113A1 (fr) | Procédé et dispositif de représentation par couche de contenu volumétrique |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20210111 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
| 17Q | First examination report despatched |
Effective date: 20230614 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN |
|
| 18D | Application deemed to be withdrawn |
Effective date: 20231025 |