A data driven solution to predict the performance of 3D assets on a wide range of device

Abstract

The present invention discloses a data driven method and system to predict the performance of 3D assets on a wide range of devices The method comprises the steps of obtaining validation data from an end user device (102) and training data from a synthetic data source (101) to create an accumulated data sink (103) which will be used for correlation analysis and feature extraction (104). The extracted data is then fed to a data driven prediction engine (105). A new 3D asset (108) which is opened in a profiler (107) environment to measure the values of the output properties and compared with the expected values to send some predictions / suggestions back from the model (106) for the next iteration.

Description

TITLE OF THE INVENTIONA Data Driven Solution to Predict the Performance of 3D Assets on a Wide Range of DeviceFIELD OF INVENTION:

[0001] The present invention relates to the field of machine learning. Particularly, the present invention provides a data driven solution to predict the performance of 3D assets on a wide range of devices.BACKGROUND OF THE INVENTION:

[0002] Creation of 3D assets for distribution often involves long iterations. This is due to lack of tools that can give feedback to the artist as to how their asset is going to look and perform on a wide range of devices. Every time the asset requires optimization; the artist has to send the revised asset for performance analysis which generally is done by a different person. The performance analysis often includes manually rendering the asset on a list of select devices which might not be exhaustive. This leaves room for uncaught performance issues even after multiple iterations. Moreover, each iteration consumes a lot of man hours. Currently there is no tool that can provide real time feedback.

[0003] A U.S. document US11954112B2 discloses systems, methods, and devices for a cyberphysical (loT) software application development platform based upon a model driven architecture and derivative loT SaaS applications. The system may include concentrators to receive and forward time-series data from sensors or smart devices, message decoders to receive messages comprising the time-series data and storing the messages on message queues, a persistence component to store the time-series data in a key-value store and store the relational data in a relational database, a data services component to implement a type layer over data stores and processing component toaccess and process data in the data stores via the type layer, the processing component comprising a batch processing component and an iterative processing component.

[0004] A Japanese document JP2024147557A discloses systems, devices and methods are provided to inform the optimization of energy supply and storage resources behind the meter and local clusters of co-located or nearby resources within a community, low voltage network, feeder, neighbourhood or building. The methods include scheduled reactive and active management of local clusters of data sources and resources across objectives such as price, energy supply, renewable energy leverage, asset value, constraint or risk management, achieving local objectives such as providing offsetting resources and supporting local balancing or constraint management of larger local supplies and loads, and supporting active management of local energy demand and renewable supplies, storage resources, electric heat resources, electric vehicle charging resources or clusters of electric vehicle chargers and flexible loads within buildings.

[0005] A U.S. document US20210097449A1 discloses a memory-efficient system for decision tree machine learning. The processing circuitry allocates a tree node array in memory, where the number of array elements in the tree node array equals the number of data samples in a training dataset. The processing circuitry also obtains the training dataset, which contains data samples captured at least partially by sensor(s) and trains the decision tree model. For example, a root node is initially assigned to the data samples in the training dataset. The root node is recursively split into child nodes based on identified branch conditions, where each child node is assigned to a subset of data samples. The tree node array is continuously updated during training to identify the child nodes assigned to the data samples. The processing circuitry then stores the trained decision tree model in memory.

[0006] A U.S. document US20240127068A1 discloses a machine learning system to enhance various aspects of machine learning models. A substantially photorealistic three-dimensional (3D) graphical model of an object is accessed and a set of training images of the 3D graphical mode are generated, the set of training images generated toadd imperfections and degrade photorealistic quality of the training images. The set of training images are provided as training data to train an artificial neural network.

[0007] Francesco Morini, November 30, 2022, discloses steps to succeed with a good data-driven strategy that leverages Predictive Analytics. The document discloses about a software providers are touting products that claim to facilitate this metric-centric decision making. All-in-all, data-driven is now perceived as the right way to do business.

[0008] Joaquim A. P. Braga et al, Volume 12. June 2023, 100180, discloses data-driven decision support system for degrading assets and its application under the perspective of a railway component.

[0009] Radiant, Jul 26, 2022, Data-Driven Asset Management: Harnessing the Power of Data Analytics in Asset Management, discloses harnessing the power of data analytics, organizations can gain a greater understanding of their assets and how they are being used. This, in turn, can help to improve decision-making, optimize performance and save companies money.

[0010] However, none of the documents disclose the real time feedback and help them by suggesting minimal changes to further optimize their assets to achieve their performance targets. Currently there is no tool that can do this on the market.

[0011] In view of the above, there arises a need to develop a system which enables the artists with real time feedback and helps them by suggesting minimal changes to further optimize their assets to achieve their performance targets. The present invention provides a better solution for the deficiencies existing in the art.OBJECTIVE OF THE INVENTION:

[0012] An objective of the present invention is to provide a data driven solution to predict the performance of 3D assets on a wide range of devices.

[0013] Another objective of the invention is to identify performance issues in 3D assets prior to deployment.

[0014] Another objective of the invention is to generate optimization recommendations for improving asset performance on target devices.

[0015] Another objective of the invention is to reduce iterative testing by enabling early stage performance evaluation.SUMMARY OF THE INVENTION:

[0016] In an embodiment the present invention proposes a cutting-edge solution to optimize digital assets by leveraging real-time data collected from end-user devices and a synthetic source. This data may be used to train machine learning models capable of predicting whether a particular asset is too heavy for a specific device. Additionally, the system is capable of forecasting the number of frames per second a device can render for the asset and the time required for the asset to load and render its first frame.

[0017] In one embodiment, a method to predict performance of a 3D asset is disclosed. The method may comprise the steps of: obtaining training data from a synthetic source that comprises a central repository of performance data; obtaining validation data generated from execution of a 3D asset on an end user device; combining the validation data and training data to create an accumulated data sink; processing the data obtained from the accumulated data sink, by one or more processors configured to a computing device, to extract relevant features.

[0018] In an other embodiment, the method may further comprise feeding the said extracted data into a data-driven prediction engine executed within the computing device;generating a live preview of a 3D asset in a profiler environment, wherein a suite of predictive models trained using the data fed into the data-driven prediction engine, interact with the 3D asset; sharing one or more input properties of the 3D asset with the suite of predictive models in the live preview to predict performance of the 3D asset across a plurality of devices. Further the invention presents optimization recommendations, through the live preview on a display interface of the computing device, for achieving a desired performance level of the 3D asset on one or more targeted devices.

[0019] In an other embodiment, the input properties of the 3D asset may be rendered within a profiler environment and provided as inputs to the suite of predictive models. Further, the output performance properties of the 3D asset may be predicted based on input properties of the said 3D asset and predefined performance characteristics of the said asset when executed on one or more targeted devices.

[0020] In an other embodiment, the input properties of the 3D asset may be transmitted through an application programming interface to a recommendation module configured to generate optimization suggestions for the 3D asset. The input properties of the 3D asset may comprise at least one of Mesh Count, Vertex Count, Triangle Count, Light Count, Material Count, Texture Count, Texture Size, Draw cells, and Shader Complexity.

[0021] In an other embodiment, the output properties of the 3D asset may comprise at least one of contextLost / crashDetected, fps, gpuFrameTime, cpuFrameTime, and interFrameTime. Further, the device parameters may include device model, browser name, browser version, engine name, engine version, canvas height, experience type, hardware concurrency, and device memory are analysed to predict the performance of the 3D asset on the said devices.

[0022] In another embodiment, a system for predicting performance of a 3D asset is disclosed.The system may comprise a synthetic data source (101) comprising a central repositoryconfigured to store training performance data; an end user device (102) configured to execute a 3D asset and generate validation performance data.

[0023] In an other embodiment, the system may further comprise a data aggregation module operably coupled to the synthetic data source and the end user device and configured to combine the training data and the validation data to create an accumulated data sink; a data processing module configured to perform correlation analysis and feature extraction on data stored in the accumulated data sink.

[0024] In yet another embodiment, the system may further comprise a data-driven prediction engine operably coupled to the data analysis module and configured to receive extracted features to generate performance prediction; a profiler environment configured to the system to generate a live preview of the 3D asset on an application interface wherein the live preview comprises a suite of predictive models trained by the data fed into the data-driven prediction engine to predict the performance of the 3D asset across an exhaustive set of devices.

[0025] In another embodiment, the system may be configured to collect input properties of the 3D asset during execution of the live preview wherein the synthetic data source comprises historical performance data collected from executions of multiple 3D assets across a plurality of devices.

[0026] By incorporating real-time feedback, the invention offers artists minimal adjustments that can be made to an asset to meet predefined performance targets. These suggestions aim to ensure the asset runs optimally on the target device without compromising user experience. This innovation, which directly integrates into the artist’s workflow during the asset creation phase, helps streamline the production process.

[0027] The solution bridges a significant gap in the market where no existing tool offers such precise optimization capabilities. This improvement not only enables artists to produceassets that meet performance criteria but also reduces the time from concept to production, resulting in considerable economic benefits.BRIEF DESCRIPTION OF DRAWINGS:

[0028] These and other features, aspects, and advantages of the present invention will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

[0029] Fig. 1 : illustrates data flow of the modelling tool according to the present invention.

[0030] Fig. 2 : illustrates flow chart for the process according to the present invention.

[0031] Fig. 3: illustrates the correlation heatmap insights that justify the model's feature selection.

[0032] Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have necessarily been drawn to scale. Furthermore, the drawings may show only those specific details that are pertinent to understanding the embodiments of the present invention so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having benefit of the description herein.DETAILED DESCRIPTION:

[0033] To promote an understanding of the principles of the invention, reference will now be made to the embodiment illustrated in the drawings and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention asillustrated therein being contemplated as would normally occur to one skilled in the art to which the invention relates.

[0034] It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be restrictive thereof.

[0035] Reference throughout this specification to “an aspect”, “another aspect” or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrase “in an embodiment”, “in another embodiment”, and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.

[0036] The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or elements or structures or components proceeded by "comprises...a" does not, without more constraints, preclude the existence of other devices or other sub-systems or other elements or other structures or other components or additional devices or additional sub-systems or additional elements or additional structures or additional components.

[0037] The invention may be implemented using a computing device comprising one or more processors and a memory storing computer-readable instructions. The computer-readable instructions, when executed by the processor, cause the computing device to perform the various methods described herein. The instructions may be stored on a non-transitory computer-readable medium including, but not limited to, memory devices, storage devices, or any suitable tangible medium capable of storing program code. The computing device may be configured to receive user input through a client interface, process the input using one or more software modules, and generate outputsin the form of audio, visual, or combined responses. The described functionalities may be implemented as software, firmware, or a combination thereof, and may operate in a standalone manner or in a distributed computing environment, without departing from the scope of the present disclosure.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The system, methods, and examples provided herein are illustrative only and not intended to be limiting.

[0039] Embodiments of the present invention will be described below in detail. The present invention provides a live preview of the asset in the artist’s Tenderer of choice These are some exemplary 3D rendering environments that may be known to a user skilled in contemporary art. For example Babylon JS®, an open-source 3D engine by Microsoft, Three JS®, an open-source library, Unreal engine®, a graphics engine provided by Epic Games and Unity® by Unity Software Inc. The live preview may have a suite of proprietary ML models that may predict the performance of a given asset on a fairly exhaustive set of devices. The system helps the artist by prompting them about any necessary optimizations to achieve a desired performance level on the target devices. Having these tools directly integrated into artist’s workflows at the initial creation phase greatly reduces the overall time taken from inception to production. Therefore, having a significant economic impact.

[0040] In one embodiment, the present invention may be powered by data collected from all the live experiences. These experiences may include the run of the 3D assets on the devices of the end user, making the data diverse in terms of the device pool. All this data may be used to train a suite of ML models to accurately predict if an asset is too heavy for a device (i.e. if the given asset is going to crash on a particular device) and if not, predict the number of frames per second that the device is capable of rendering the asset and the time taken to load the asset and render the first frame.

[0041] In an alternative embodiment, the present invention may suggest minimal changes that can be made to a given asset to achieve a certain performance target that is defined by the user. The invention disclosed herein enables the artists with real time feedback and helps them by suggesting minimal changes to further optimize their assets to achieve their performance targets.

[0042] In an alternative embodiment, the present invention may solve the problem existing in the art with a live preview of the asset in the artist’s Tenderer of choice (i.e. Babylon JS®, ThreeJS®, Unreal Engine®, Unity®, etc.). The live preview may have a suite of proprietary ML models to predict the performance of a given asset on a fairly exhaustive set of devices.

[0043] In an alternative embodiment, the present invention may help the artist by prompting them about any necessary optimizations to achieve a desired performance level on the target devices. Having these tools directly integrated into artist’s workflows at the initial creation phase greatly reduces the overall time taken from inception to production. Therefore, having a significant economic impact.

[0044] In an alternative embodiment, the system may use data driven models to accurately predict the performance of a given 3D asset on a wide range of devices without having to manually test on physical devices. Additionally, also give out minimal changes, to be done to the asset, to achieve a given performance target.

[0045] In an alternative embodiment, the present invention may work by analysing the previous data to predict how a new 3D asset behaves / performs on a device. The invention may analyze the asset’s stats like vertexCount, triangleCount, materialCount, textureCount along with device specifications like logical processor count, RAM, device make, etc. to train models and give out output parameters like crash possibility, FPS and confidence score.

[0046] In an alternative embodiment, the present invention may utilize modeling tools such as Blender® by Blender Foundation to create 3D assets. From within the modeling tool, the artist can choose to preview their asset in their Tenderer of choice (i.e. BabylonJS®, Three JS®, Unreal Engine®, Unity®, etc.) and get real time feedback and suggestions to achieve their performance targets, thereby providing real time feedback and helping the artists by suggesting minimal changes to further optimize their assets to achieve their performance targets.

[0047] In various embodiments, the present invention may evaluate the performance of the 3D asset in different categories and properties as per the output metrics, the feedback is provided and suggested for changes to achieve their performance targets.

[0048] In accordance with the present invention, a data driven method and system to predict the performance of 3D assets on a wide range of devices is disclosed. The said method may be implemented through a distributed computing system, comprising one or more end user devices, a central processing system, and one or more storage units interconnected through a communication network. The end user devices may include computing hardware such as processors, GPUs, memory units, display interfaces to execute and render 3D assets and transmits the said assets to a central processing system through the communication network.

[0049] In one embodiment, the data from the end- user device (102), herein after “validation data” may relate to how a particular 3D asset / scene performs on one or more targeted devices. The validation data indicative of real world performance of a 3D asset may be obtained when an asset is executed on one or more end user devices. The data is fetched by analysing Performance metrics such as fps, frame times, drawCalls, etc., asset / scene linkage such as projectinfo, allMeshStat, allTexStat, allMatStat. The validation data may also capture signals about issues the devices face while loading or rendering 3D assets in general by analysing loadTime, contextLost. The validation data may further focus on device specifications including but not limited to device / browser / engine, resolution and canvas information. Overall, the end user devicemay provide both data, i.e., Per- asset / per- scene performance on that user’s device, and Device- level conditions and problems during loading / rendering of any 3D asset.

[0050] In other embodiment, “Frame times” may refer to the duration required to render a single frame, including GPU frame time and CPU frame time, and represent the time taken by respective processing units to complete rendering tasks per frame. “Draw calls” may refer to individual rendering commands issued by the CPU to the GPU to render assets within a scene. Further, “allMeshStaf ’ may refers to aggregated statistical data associated with all the meshes present within a scene, “allTexStat may refer to aggregated statistical data associated with all the textures present within a scene, and “allMatStaf ’ may refer to aggregated statistical data associated with all the materials present within a scene. Furthermore, “Load Time” may refer to the duration required for a device to load, initialize, and prepare a 3D asset or scene for rendering and “ContextLosf ’ may refer to a condition in which the graphics rendering context such as WebGL® or WebGPU® becomes unavailable due to device limitations.

[0051] In one embodiment, the central processing system configured to the system may comprise one or more processor units, system memory, and non-transitory storage media configured to store a central data repository herein after “synthetic data source”. The synthetic data source (101) may store structured performance data obtained from executions, and historical runs, therefore may be used as training data. The training data may comprise characteristics of 3D assets and scenes including but not limited to meshCount, textureCount, materialCount, meshSizes, texturesSize, materialSize, per-asset stats, projectinfo with AssetIDs. Along with the aforementioned, the synthetic data may also provide for performance metrics of those assets / scenes on specific devices / browsers / engines FPS, frame times, load time, draw calls, device / browser details. The validation data and the training data may be combined to form an accumulated data sink (103) to represent a comprehensive dataset reflecting both actual and synthetic performance characteristics. The data stored in the accumulated data sink may then be processed to perform correlation analysis andfeature extraction to identify relevant performance related features associated with input and output characteristics of the 3D assets.

[0052] In one embodiment, “meshCount” may refer to the total number of discrete geometric mesh objects present in a 3D scene, where each mesh represents a collection of vertices, edges, and faces defining a renderable geometric structure. “textureCount” may refer to the number of distinct texture resources applied to one or more meshes in a 3D asset. “materialCounts” may refer to the number of distinct materials definitions associated with meshes in a 3D asset, wherein each material defines rendering properties such as shading, texture and surface appearance. Further, “Pre-asset statistics” may refer to performance-relevant parameters of an individual 3D asset, including vertexCount, triangleCount, drawCall count and the run time rendering metrics associated with the said 3D asset. “Projectinfo” may refer to metadata associated with one or more 3D assets such as configuration parameters, rendering settings, and hierarchy details. “AssetIDs” may refer to unique identifiers assigned to individual 3D assets within a scene. This enables the system to track and correlate asset characteristics with corresponding performance metrics across different sets of devices.

[0053] In certain embodiments, the synthetic data source may provide data relating solely to intrinsic characteristics of 3D assets and scenes, without including device-specific performance measurements. Such data may include structural and rendering related attributes of the assets such as mesh count, material count, vertex count, triangle count, texture properties, and other scene complexity parameters. This dataset may be used to evaluate compatibility of a 3D asset by executing them on an end user device. The compatibility is decided based on if the asset is capable of being rendered within the hardware and software constraints of that device. The results obtained from this embodiment establish correlations between asset characteristics and device capabilities.

[0054] In certain embodiments, the central processing system may further be coupled with a data driven prediction engine implemented in hardware-backed execution units such as CPUs, GPUs that are configured to perform correlation analysis and feature extraction.The processing system determines which features may be relevant by focusing only on parameters that directly affect how efficiently a device's graphics hardware works, especially the GPU processing pipeline and the data transfer bandwidth between system components. Instead of collecting every possible engine parameter, the system may filter and select meaningful inputs using statistical correlation analysis, where each input metric may automatically be assigned a weight, based on how strongly it impacts performance. The selected features fall into three main categories:

[0055] “Throughput bottlenecks” which may relate to the quantity of work being sent from the CPU to the GPU to analyse how frequently the CPU must communicate rendering instructions to the GPU. This frequency may be determined based on the metrics such as the number of draw calls. For e.g., a high number of such operations increases burden on GPUs that can slow rendering performance.

[0056] “Memory footprint” may relate to how much graphics memory is being consumed.Features such as total texture size and environment map size indicate how much memory is required to load and render the scene. For e.g., larger memory usage can cause pressure on graphics memory leading to slowdowns or crashes on lower-end devices.

[0057] “Computational complexity” may relate to the amount of processing the GPU must perform. Parameters such as vertexCount and triangleCount affect the load on vertexShaders and pixelShaders, increasing the time required for rendering each frame. “VertexCount” refers to the total number of vertices used to define the geometric structure of a 3D asset, where each vertex represents a point in 3D space that defines positions, textures or other rendering attributes. “TriangleCount” refers to the total number of triangular primitives used to construct the surface geometry of a 3D asset.

[0058] In certain embodiments, to reduce unnecessary data complexity, the system may also perform statistical bucketing wherein instead of sending detailed information such as individual texture file names, the system applies dimensionality reduction techniques.For example, textures may be grouped into resolution-based categories (such as 1024x1024 or 2048x2048), allowing the model to learn performance impact patterns without being overloaded with irrelevant identifiers.

[0059] In certain embodiments, for performance-related outputs such as frames per second (FPS) and CPU / GPU frame times, the system may apply delta-time sampling using a sliding time window. This means performance data is averaged over a defined time interval so that the extracted features reflect stable performance trends rather than temporary spikes or fluctuations, resulting in more reliable and consistent prediction inputs for the models. These inputs may be then fed into a machine learning algorithm based prediction engine to develop predictive models.

[0060] As per one embodiment, the development of prediction models may be based on Supervised Machine Learning to map scene complexity to hardware- specific performance outcomes. The suite of predictive models may be trained on a "Ground Truth" dataset where input features (extracted asset stats) are labeled with the actual measured fps , gpuFrameTime , and cpuFrameTime across various device tiers.

[0061] In an other embodiment, the system may further employ Regression-based models such as Gradient Boosted Trees to interpret the non-linear relationships revealed in the correlation analysis during processing. The regression model may be configured to identify specific performance cliffs points. The cliff points may refer to specific points where a small increase in any parameter pertaining to the 3D asset can cause sudden performance mishaps.

[0062] In further embodiments, a profiler environment (107) may be executed on a computing device that may comprise at least one processor, memory, and GPU. The profiler environment may be configured to load the 3D asset and generate a live preview of the 3D asset. This live preview allows real time evaluation of the 3D asset during its creation or modification. The profiler may perform recursive scene traversal by programmatically crawling through the entire scene graph wherein it iterates throughthe active arrays for meshes , materials , and textures , querying the engine's internal state to retrieve live properties of a 3D asset (e.g., vertexCount , materialCount ) rather than relying on static source files.

[0063] In further embodiments, the profiler may use hardware- level instrumentation by interfacing with a graphics rendering interface such as WebGL® / WebGPU® API via 3D graphics engine such as BabylonJS®, Unity®, Unreal Engine®, etc., to allow the tool to isolate and extract low-level timing data, such as gpuFrameTime and drawCalls, directly from the GPU's execution queue during the render loop. After collecting both scene complexity data and hardware performance data, the profiler may serialize the data into a structured format such as JSON, XML, YAML that are compatible with the machine learning model's input dimensions.

[0064] In certain embodiments, the profiler may operatively be coupled to the data driven prediction engine (105) and the predictive models (106) to transmit the extracted input properties of the 3D asset to the trained predictive models. The said transmission may occur during the live preview wherein a suite of predictive models trained through the data fed into the machine learning based prediction engine interacts with the input properties of the 3D assets. Based on the received input properties, the predictive models estimate the performance of the 3D asset across an exhaustive set of devices, including variations in hardware, software, and execution environments.

[0065] In certain other embodiments, the profiler environment (107) may further present optimization recommendations through the live preview, guiding an artist or developer toward modifications required to achieve a desired performance level on one or more targeted devices. The profiler environment may be embedded into the user's device or may function on a cloud to communicate the optimization recommendations to the user. By embedding performance prediction and optimization feedback directly into the asset creation workflow, the method enables iterative refinement of the 3D asset at an early stage, thereby minimizing performance related failures on target devices, and shortening the overall time from asset inception to production deployment.

[0066] In another embodiment of the present invention a system for predicting performance of a 3D asset is disclosed. The system may comprise a synthetic data source (101) that has a central repository configured to store training performance data, an end user device (102) configured to execute a 3D asset and generate validation performance data. The system may further comprise a data aggregation module that may be operably coupled to the synthetic data source and the end user device and configured to combine the training data and the validation data to create an accumulated data sink (103).

[0067] In yet another embodiment, the system may further comprise a data processing module to perform correlation analysis and feature extraction on data stored in the accumulated data sink (103). The extracted features are fed into a data-driven prediction engine (105) that is operably coupled to the data analysis module to generate performance prediction. The profiler environment (107) configured to the system may generate a live preview of the 3D asset on an application interface wherein the live preview comprises a suite of predictive models (106) trained by the data fed into the data-driven prediction engine (105) to predict the performance of the 3D asset across an exhaustive set of devices.

[0068] In an embodiment, a computer-implemented method for predicting performance of a three-dimensional (3D) asset across a plurality of devices prior to deployment is disclosed. The method may utilize both historical performance data and real-time execution data to train and operate predictive models capable of estimating how a 3D asset will perform under different hardware and software environments. Further, training data may be obtained from a synthetic data source comprising a centralized repository of performance information associated with various 3D assets and device conditions. Validation data may be additionally collected from execution of a 3D asset on an end user device. The training data and validation data may be aggregated into an accumulated data sink and processed to extract performance-relevant features. These extracted features are provided to a data-driven prediction engine that trains and updates a suite of predictive models. A profiler environment that may be integrated togenerate a live preview of the 3D asset, within which the predictive models may interact with the asset by receiving its input properties and estimating corresponding performance outcomes across multiple devices. Based on these predictions, the system may further present optimization recommendations through a display interface, enabling users to modify asset characteristics to achieve a desired performance level on one or more targeted devices.

[0069] Fig. 1 illustrates data flow (100) of the modelling tool according to the present invention. The flow may comprise the steps of obtaining validation data from an end user device (102) and training data from a synthetic data source (101) e.g. a central data repository. Both data may be used to create an accumulated data sink (103) which will be used for correlation analysis and feature extraction (104). The extracted data may be then fed to a machine learning algorithm (105).

[0070] In one embodiment, an individual 3D artist (109) starts executing a 3D asset (108)which may be opened in a profiler (107) environment. The input properties of the 3D assets are used to measure the values of the output properties and compared with the expected values. In case of any issues or performance improvements needed, the machine learning algorithm (105) may send some predictions / suggestions back from the model (106) for the next iteration.

[0071] In another embodiment, the characteristic input properties of the 3D assets obtained from the synthetic data source may be correlated with corresponding characteristic properties extracted from new 3D assets (108) loaded into the profiler environment by the user for evaluation. The assets in the synthetic data source may serve as a reference data for model training and correlation analysis. This correlation enables the system to generate a coherent predictive output for the user-loaded asset.

[0072] Fig. 2 illustrates flow chart (200) for the process according to the present invention. In the flowchart, the process is started by opening assets in a modelling tool (201) and choosing a suitable environment (202). The asset’s live preview may be opened inbrowser or application (204) by clicking (203). After opening the preview (204), the asset properties may be collected (205) and sent to ML API (206). All the properties are validated to obtain a list of devices where the asset might have performance issues (207). Check if there is any performance problem (208). If there are no performance issues observed, then the process is stopped and in case of issues, need help with how to optimise (209). The properties may again be sent to ML API (210) and along with a list of optimization suggestions for the assets (211). Further, optimized assets in a modelling tool (212) may again be opened the live preview (204) by clicking the live preview (203) to execute the whole cycle for next iterations until there are no performance issues observed.

[0073] Fig. 3 illustrates the heatmap as the mathematical justification for the model's feature selection. The correlation observation demonstrates a high positive totalVertices & triangleCount which confirms the data integrity offered by the invention wherein these redundant features can be collapsed to reduce model noise. Further, strong negative drawCalls and fps validates that CPU-side management of scene objects is a primary bottleneck for frame rate. Furthermore, positive meshCount and cpuFrameTime Confirms the direct impact of scene hierarchy complexity on the engine's processing overhead.The categories, properties and description of the 3D asset, rendering environment and the output metrics are tabulated as below:Category Property Description3D Asset Mesh Count The total number of mesh objects present in a 3D scene or mesh.(ex: “meshCount”)Vertex Count The total number of vertices used to define the geometric structure of (ex: “vertexCount”)a mesh.Largest mesh vertex count The highest number of vertices in the largest mesh.(ex: “largestMesh VertexCount”)Triangle count The total number of triangular primitives used to construct the (ex: “triangleCount”)geometry of a 3D asset.Light Count The total number of active light sources in a 3D scene(ex: “lightCount”)Material Count The total number of distinct materials used across meshes in a (ex: “materialCount”)3D scene.Texture Count The total number of textures being used in the scene.(ex: “textureCount”)Total texture size The cumulative size of all textures.(height * width * No of channels) (ex: “totalTextur eSize”)Largest texture size The size of the largest texture of a mesh in a 3D scene.(ex: “largestTextureSize”)Environment size Resolution of env(HDRI) used for image-based lighting(ex: “envSize”)Resolution specific te xture A breakdown of texture counts by count resolution: x64, xl28, x256, x512, xl024, x2048, x4096, x8192: (ex: “textureCountResSpec fic”)Number of textures for each resolutionDraw calls (ex: “drawCalk ”) The number of rendering commands issued by CPU to GPU.Complexity of shader Number of shader instructions being used to calculate each pixel (ex: “shaderComplexity”)drawn on screen (for a given material)Rendering Model of device The model of the device being environment used(ex: “deviceModel”)Name of browser The name of the browser (e.g., Chrome,Safari)(ex: “browserName”)Version of browser The version of the browser used for rendering(ex: “browserVersion”)Rendering engine The name of the rendering engine (Three, js, Baby Ion. js)(ex: “engineName”)Rendering engine The version of the rendering engine version(ex: “engine Version”)Height of canvas The height of the rendering canvas in pixels(ex: “canvasHeight”)Width of canvas The width of the rendering canvas in pixels(ex: “canvasWidth”)Rendering experience Type of experience being rendered (e.g., (ex: “experienceType”) "VR", "AR", "3D")Concurrency of hardware Number of logical processors available to (ex: “hardwareConcurren run threads on the targeted device. cy”)Memory of device Approximate amount of device memory in gigabytes(ex: “deviceMemory”)Output Rendering output Whether the experience ran out of metrics memory or crashed.(ex: “contextLost / crashDetected”)Frames per second The frames per second recorded during a rendering session.(ex: “fps”)GPU frame rendering The time spent by the GPU to render time each frame (milliseconds)(ex: “gpuFrameTime”)CPU frame rendering The time spent by the CPU to process time each frame (milliseconds)(ex: “cpuFrameTime”)Frame transition time The time gap between two consecutive frames (milliseconds)(ex: “interFrameTime”)

[0074] While the description herein provides various processes that a user may follow, an exemplary workflow is being described for ease of reference as per one embodiment. The steps may be:Receiving asset characteristic data to analyse if they can perform smoothly on one or more targeted devices;Defining one or more performance indicators representing smoothness of a rendering experience, the indicators including at least average frames per second (FPS) and variation in FPS over time;Identifying performance limitations based on input properties of the assets and pre-defined performance of those asset on targeted device;Feeding the 3D asset, that is required to be analysed into a prediction engine wherein the input properties of 3D asset are shared with the trained prediction modules;Receiving optimization feedback from the prediction module to provide an insight to the user regarding the modifications required in the asset to be rendered on targeted devices, thereby reducing iterative QA cycles.

[0075] If not this invention, these iterations would keep happening until the Asset clears QA. Since there are 2 parties involved (Artists and QA team), there will be time wasted in communicating and there will be time wasted in retesting the asset multiple times. Hence, a ML route where the model learns the relation (both linear and non-linear relation) between the parameters and the output (FPS) has been disclosed in the present invention This ensures any nonlinear relation between parameters and the output.

[0076] In various embodiments, the present invention may reduce the overall time involved in iterative asset optimization and quality assurance processes. The invention may address the technical problem of minimizing the likelihood of asset rejection during performance validation stages by enabling early prediction of rendering performance. Instead of relying solely on generalized design guidelines, which may fail to account for complex interactions among asset parameters and device capabilities, the present invention may employ a machine learning model configured to learn relationships between multiple asset characteristics and performance outcomes. Such a model may capture both linear and non-linear dependencies among parameters, thereby enabling more accurate and scalable performance assessment across diverse hardware environments.

Claims

1. CLAIMSWe claim:

1. A method to predict performance of a 3D asset, comprising the steps of:a. obtaining training data from a synthetic data source (101) that comprises a central repository of performance data;b. obtaining validation data generated from execution of a 3D asset on an end user device (102);c. combining the validation data and training data to create an accumulated data sink (103);d. processing the data obtained from the accumulated data sink (103), by one or more processors configured to a computing device, to extract relevant features; e. feeding the said extracted data into a data-driven prediction engine (105) executed within the computing device;f. generating a live preview of a 3D asset (108) in a profiler environment (107), wherein a suite of predictive models (106) trained using the data fed into the data-driven prediction engine (105), interact with the 3D asset; andg. sharing one or more input properties of the 3D asset with the suite of predictive models (106) in the live preview to predict performance of the 3D asset across a plurality of devices.

2. The method to predict performance of a 3D asset as claimed in claim 1, further comprises presenting optimization recommendations, through the live preview on a display interface of the computing device, for achieving a desired performance level of the 3D asset on one or more targeted devices.

3. The method to predict performance of a 3D asset as claimed in claim 1 , wherein the input properties of the 3D asset are rendered within the profiler environment (107) and provided as inputs to the suite of predictive models (106).

4. The method to predict performance of 3D asset as claimed in claim 1 wherein output performance properties of the 3D asset is predicted based on input properties of the said 3D asset and predefined performance characteristics of the said asset when executed on one or more targeted devices.

5. The method to predict performance of 3D asset as claimed in claim 1 further comprises, transmitting the input properties of the 3D asset through an application programming interface (206) to a recommendation module configured to generate optimization suggestions for the 3D asset.

6. The method to predict performance of 3D asset as claimed in claim 1, wherein the input properties of the 3D asset comprise at least one of Mesh Count, Vertex Count, Triangle Count, Light Count, Material Count, Texture Count, Texture Size, Draw cells, and Shader Complexity.

7. The method to predict performance of 3D asset as claimed in claim 1, wherein the output properties of the 3D asset comprise at least one of contextLost / crashDetected, fps, gpuFrameTime, cpuFrameTime, and interFrameTime.

8. The method to predict performance of 3D asset as claimed in claim 1, wherein the device parameters including device model, browser name, browser version, engine name, engine version, canvas height, experience type, hardware concurrency, and device memory are analysed to predict the performance of the 3D asset on the said devices.

9. A non-transitory computer readable medium storing instructions which, when executed by a processor, cause a computing device to perform the steps comprising:a. obtaining training data from a synthetic data source (101) that comprises a central repository of performance data;b. obtaining validation data generated from execution of a 3D asset on an end user device (102);c. combining the validation data and training data to create an accumulated data sink (103);d. processing the data obtained from the accumulated data sink (103), by one or more processors configured to a computing device, to extract relevant features; e. feeding the said extracted data into a data-driven prediction engine (105) executed within the computing device;f. generating a live preview of a 3D asset in a profiler environment (107), wherein a suite of predictive models (106) trained using the data fed into the data-driven prediction engine (105), interact with the 3D asset; andg. sharing input properties of the 3D asset (108) with the suite of predictive models (106) in the live preview to predict performance of the 3D asset across an exhaustive set of devices.

10. The non-transitory computer-readable medium as claimed in claim 9, further comprises presenting optimization recommendations, through the live preview on a display interface of the computing device, for achieving a desired performance level of the 3D asset on one or more targeted devices.

11. A system for predicting performance of a 3D asset, comprising:a. a synthetic data source (101) comprising a central repository configured to store training performance data;b. an end user device (102) configured to execute a 3D asset and generate validation performance data;c. a data aggregation module operably coupled to the synthetic data source and the end user device and configured to combine the training data and the validation data to create an accumulated data sink;d. a data processing module configured to perform correlation analysis and feature extraction on data stored in the accumulated data sink;e. a data-driven prediction engine (105) operably coupled to the data analysis module and configured to receive extracted features to generate performance prediction; andf. a profiler environment (107) configured to the system to generate a live preview of the 3D asset on an application interface wherein the live preview comprises a suite of predictive models (106) trained by the data fed into the data-driven prediction engine (105) to predict the performance of the 3D asset across an exhaustive set of devices.

12. The system for predicting performance of a 3D asset as claimed in claim 11, wherein the system is configured to collect input properties of the 3D asset during execution of the live preview.

13. The system for predicting performance of a 3D asset as claimed in claim 11, wherein the synthetic data source (101) comprises historical performance data collected from executions of multiple 3D assets across a plurality of devices.

14. The system for predicting performance of a 3D asset as claimed in claim 11, wherein the input properties of the 3D asset comprise at least one of Mesh Count, Vertex Count, Triangle Count, Light Count, Material Count, Texture Count, Texture Size, Draw cells, and Shader Complexity.

15. The system for predicting performance of a 3D asset as claimed in claim 11, wherein the output properties of the 3D asset comprise at least one of contextLost / crashDetected, fps, gpuFrameTime, cpuFrameTime, and interFrameTime.

Images