DYNAMIC RECONSTRUCTION OF NEW VIEWS BASED ON RIVER REMATCHING

Abstract

Several examples, systems, and methods are disclosed that involve the dynamic reconstruction of new views, at least partially based on flow rematching. A first computing system can cause an image rendering model to generate an estimated image of a scene based on at least a plurality of images of the scene. The at least one image in the plurality of images can be from at least one different time or view. The first computing system can update the image rendering model based on at least the estimated image, the plurality of images, and / or one or more motion criteria associated with the estimated image.

Description

BACKGROUND

[0001] Dynamic reconstruction systems for new views often use static scene representations, which can exhibit limited effectiveness in rendering content and objects with complex deformations or rapid changes in motion. Techniques such as voxel-based methods or static mesh models have a limited ability to represent dynamic aspects of scenes, especially when input data is sparse, temporally inconsistent, or captured from varying viewpoints. These limitations can lead to artifacts, inaccuracies in motion rendering, and increased computational load, particularly in applications requiring real-time or near-real-time scene rendering. For example, conventional methods may not adequately represent the interactions of deformable objects, resulting in visually inconsistent output and increased latency in dynamic environments such as...Augmented Reality, robot vision, or autonomous navigation systems. SUMMARY

[0002] The invention is defined by the claims. To illustrate the invention, aspects and embodiments that may or may not be within the scope of protection of the claims are described herein.

[0003] Several examples, systems, and methods are disclosed that involve the dynamic reconstruction of new views, at least partially based on flow rematching. A first computing system can cause an image rendering model to generate an estimated image of a scene based on at least a plurality of images of the scene. The at least one image from the plurality of images can be associated with at least one image from a different time or view. The first computing system can update the image rendering model based on at least the estimated image, the plurality of images, and / or one or more motion criteria associated with the estimated image.

[0004] The implementations of this disclosure relate to dynamic scene reconstruction systems and methods that incorporate flow rematching techniques to improve the modeling of complex deformations over time. Systems and methods are described that integrate models such as Gaussian splatting and neural radiance fields (NeRF) into velocity field determinations to enhance the accuracy of reconstructed views. These techniques enable the alignment of estimated motion fields with observed data, thus supporting the accurate rendering of scenes captured from multiple viewpoints or at different time intervals.For example, systems and methods according to the present disclosure can iteratively refine velocity fields to better represent interactions of dynamic objects, reducing discrepancies between the reconstructed flow and actual motion patterns. This approach can support accurate and computationally efficient scene reconstruction, such as under conditions with sparse or temporally inconsistent input data, thereby improving the performance and reliability of dynamic vision-based systems.

[0005] Some implementations refer to one or more processors containing processing circuitry. This processing circuitry causes an image rendering model to generate an estimated image of a scene based on at least a set of images of the scene. In some implementations, at least one image in the set of images is associated with at least one image from a different time or view. The processing circuitry updates the image rendering model based on at least the estimated image, the set of images, and / or one or more motion criteria associated with the estimated image.

[0006] In some implementations, one or more motion criteria include a velocity field representing a variety of deformations in space over time. In some implementations, updating the image rendering model also involves refitting a velocity field generated from the estimated image to a previous velocity field that corresponds to / matches the scene. In some implementations, one or more motion criteria include at least one of (i) a rigidity constraint that limits changes in the shape of objects over time, or (ii) a continuity constraint that assumes smooth transitions in one or more object motions within the scene.

[0007] In some implementations, the one or more criteria for motion correspond to a machine learning (ML) model that is trained / updated to generate one or more deformations of the estimated image based on parameters derived from one or more historical scenes. In other implementations, the one or more processors containing processing circuitry are used to determine the one or more criteria for motion based on at least one output of a minimization of one or more functions that satisfy at least one of (i) the rigidity constraint, (ii) the continuity constraint, or (iii) a constraint derived from the ML model.

[0008] In some implementations, the image rendering model includes at least one of (i) a Gaussian splatting model or (ii) a neural radiation field (NeRF) model. In some implementations, the multitude of images of the scene includes a multitude of multiview images. In some implementations, the multitude of multiview images corresponds to a multitude of distinct viewpoints captured at a multitude of distinct time points. In some implementations, the one or more processors containing processing circuitry are used to apply scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on the multitude of images of the scene.

[0009] In some implementations, updating the image rendering model includes minimizing reconstruction loss and / or refitting loss. In some implementations, reconstruction loss corresponds to a measure of the discrepancy between the estimated image and the multitude of images in the scene. In some implementations, refitting loss corresponds to a measure of the deviation between the estimated image and the one or more criteria for motion associated with an image flow.

[0010] Some implementations refer to a system. The system may contain one or more processors to perform operations that involve causing an image rendering model to generate an estimated image of a scene based on at least a set of images of the scene. In some implementations, at least one image in the set of images is associated with at least one from a different time or view. The system may contain one or more processors to perform operations that involve updating the image rendering model based on at least the estimated image, the set of images, and one or more motion criteria associated with the estimated image.

[0011] In some implementations, one or more motion criteria include a velocity field representing a variety of deformations in space over time. In some implementations, updating the image rendering model also involves refitting a velocity field generated from the estimated image to a previous velocity field that corresponds to / matches the scene. In some implementations, one or more motion criteria include at least one of: (i) a rigidity constraint that limits changes in the shape of objects over time, or (ii) a continuity constraint that assumes smooth transitions in one or more object motions within the scene.In some implementations, one or more criteria for motion correspond to a machine learning (ML) model that is trained / updated to generate one or more deformations of the estimated image based on parameters derived from one or more historical scenes.

[0012] In some implementations, one or more processors perform operations that include determining one or more criteria for motion, based on at least one output of a minimization of one or more functions that satisfy at least one of the following: (i) the rigidity constraint, (ii) the continuity constraint, or (iii) a constraint derived from the ML model.

[0013] In some implementations, the image rendering model includes at least one of (i) a Gaussian splatting model or (ii) a neural radiation field (NeRF) model. In some implementations, the multitude of images of the scene includes a multitude of multiview images, where the multitude of multiview images corresponds to a multitude of distinct viewpoints captured at a multitude of distinct time points. In some implementations, one or more processors perform operations that include applying scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on at least the multitude of images of the scene.

[0014] Some implementations refer to a procedure. The procedure involves inducing, using one or more processors, an image rendering model to generate an estimated image of a scene based on at least a set of images of the scene. In some implementations, at least one image in the set of images is associated with at least one from a different time or view. The procedure involves updating, using the one or more processors, the image rendering model based on at least the estimated image, the set of images, and / or one or more motion criteria associated with the estimated image.

[0015] The revelation extends to all novel aspects or features described and / or illustrated here.

[0016] Further features of the disclosure are characterized by the independent and dependent claims.

[0017] Any feature of one aspect of the disclosure can be applied to other aspects of the disclosure in any suitable combination. In particular, procedural aspects can be applied to device or system aspects, and vice versa.

[0018] Furthermore, features implemented in hardware can be implemented in software, and vice versa. Any reference to software and hardware features herein should be interpreted accordingly.

[0019] Each system or device feature described herein can also be provided as a process feature, and vice versa. System and / or device aspects that are described functionally (including means plus functional features) can alternatively be expressed in terms of their corresponding structure, such as a suitably programmed processor and associated working memory.

[0020] It is also understood that certain combinations of the various features described and defined in each aspect of the revelation can be implemented and / or provided and / or used independently of one another.

[0021] The disclosure also provides computer programs and computer program products comprising software code designed to perform one of the methods described herein when executed on a data processing device and / or to embody one of the device and system features described herein, including one or all component steps of a method.

[0022] The disclosure also provides a computer or computing system (including networked or distributed systems) with an operating system that supports a computer program for carrying out the procedures described herein and / or for embodying the device or system features described herein.

[0023] The revelation also provides a computer-readable medium on which one or more of the aforementioned computer programs are stored.

[0024] The revelation also provides a signal that transmits one or more of the aforementioned computer programs.

[0025] The disclosure extends to methods and / or devices and / or systems as described herein with reference to the accompanying drawings.

[0026] Aspects and embodiments of the disclosure will now be described purely by way of example with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The systems and methods for dynamic scene reconstruction in a reconstruction pipeline are described in detail below with reference to the accompanying figures. These show: Fig. 1 a block diagram of an example of a system according to some implementations of the present disclosure; Fig. 2 a flowchart of an example of a method for the dynamic reconstruction of new views, based at least partially on flow rematching, according to some implementations of the present disclosure; Fig. 3A an example of flow rematching using velocity fields, according to some implementations of the present disclosure; Fig. 3B an example of rendering flow rematching using velocity fields, according to some implementations of the present disclosure; Fig. 4A a block diagram of an exemplary generative language model system for use in the implementation of at least some implementations of the present disclosure; Fig. 4B a block diagram of an exemplary generative language model which includes a transformer-encoder-decoder for use in implementing at least some implementations of the present disclosure; Fig. 4C a block diagram of an exemplary generative language model that includes a decoder-transformer-only architecture for use in implementing at least some implementations of the present disclosure; Fig. 5 a block diagram of an exemplary computing device for use in the implementation of some embodiments of the present disclosure; and Fig. 6 a block diagram of an exemplary data center for use in the implementation of at least some implementations of the present disclosure. DETAILED DESCRIPTION

[0028] This revelation concerns systems and methods for the dynamic reconstruction of new views, such as systems and methods for reconstructing views based on flow rematching. Some systems may encounter technical limitations when reconstructing dynamic scenes from sparse and temporally inconsistent multi-view inputs, where images are captured at varying times or from different viewpoints. That is, the systems may fail to manage the complexity of synthesizing accurate intermediate views while preserving realistic motion and deformation patterns of objects within the scene. Some reconstruction models may rely on flow-based techniques, such as velocity fields (e.g.,These reconstruction models may rely on techniques to represent the movement of pixels or features across multiple frames to manage deformations, but such techniques can require integration, leading to increased computational overhead and limited scalability. These models may also rely on specific priors (a priori probabilities, such as learned motion patterns, shape constancy constraints, or other restrictions derived from historical scene data, physical laws, or domain-specific information), which can limit the model's effectiveness in generalizing to different types of dynamic scenes with varying motion patterns and deformation complexities.

[0029] Furthermore, reconstruction models can also be limited by technical problems in integrating multiple types of a priori probabilities (e.g., physical constraints or learned models) within a single model while maintaining reconstruction quality and generalization capabilities. That is, methods relying on predefined physical a priori probabilities or learned models may be inadequate for managing deformations in dynamic reconstruction models, especially in scenes with complex deformations or rapid changes.

[0030] The systems and methods according to this disclosure can provide a framework for the dynamic reconstruction of new views, incorporating flow rematching techniques to address these technical problems. By determining flows representing deformations and rematching these flows to the dynamic reconstruction model, the system can eliminate the need for time-step integration of velocity fields. That is, the framework allows the inclusion of various a priori probabilities (whether physical, such as rigidity or continuity, or derived from fundamental models) while maintaining computational efficiency and improving reconstruction accuracy. The flow rematching process enhances the iterative refinement of the image rendering model to align reconstructed views with realistic motion patterns and reduce errors in synthesized images.

[0031] For example, the systems and methods can update an image rendering model by combining a reconstruction loss (L REC ) and a readjustment loss (L RM) is minimized. That is, reconstruction loss can be attributed to comparing the estimated image with a multitude of images of the scene, where at least one image is captured at different times or from different viewpoints, to measure discrepancies between the estimated images and the reference images. Furthermore, refitting loss can be attributed to adapting an image flow represented in the estimated image to constraints that define the allowable motion based on criteria for motion, such as physical constraints or learned motion models. Accordingly, the combination of reconstruction loss and refitting loss improves the performance of dynamic reconstruction models in capturing realistic motion patterns and synthesizing accurate intermediate views, while reducing computational overhead.

[0032] In some implementations, the system can determine the criteria for motion based on the output of a minimization problem formulated to satisfy one or more constraints, such as a rigidity constraint, a continuity constraint, or a constraint derived from a machine learning model trained / updated using parameters extracted from historical scenes. The minimization process can be used to determine an optimal flow that satisfies the constraints. This process allows the model to adaptively update the scene representation without incorporating velocity fields at one or more (e.g., every) time steps. As a result, computational efficiency can be improved and training times reduced.

[0033] Furthermore, the criteria for motion can be represented by a velocity field, for example, to depict a variety of deformations in space over time. The velocity field can be used to update the image rendering model by refitting the velocity field generated from the estimated image with a previous velocity field corresponding to the scene. For example, the criteria for motion can include at least one rigidity constraint or continuity constraint, or correspond to a machine learning model trained to generate deformations of the estimated image based on parameters derived from historical scenes. The image rendering model can also use a combination of a Gaussian splatting model or a neural radiation field (NeRF) model to process different types of scene representations.This means the system can further improve dynamic scene reconstruction by supporting multiple types of representations, such as Gaussian splats, NeRF models, or any learnable rendering models. For example, these representations can be selected based on the scene's properties and the desired reconstruction quality. Furthermore, the multitude of images in the scene can include multi-view images captured from different viewpoints at different times.

[0034] With reference to Fig. 1 is Fig. Figure 1 is an exemplary block diagram of a System 100, according to some implementations of the present disclosure. It is understood that this and other arrangements described herein are presented only as examples. Other arrangements and elements (e.g., machines, interfaces, functions, arrays, groupings of functions, etc.) may be used in addition to or instead of those shown, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that may be implemented as single or distributed components, or in conjunction with other components, in any suitable combination and at any suitable location. Various functions described herein as being performed by entities may be executed by hardware, firmware, and / or software.Various functions can be performed, for example, by a processor executing instructions stored in main memory. In some embodiments, the systems, methods, and processes described herein can be implemented using components, features, and / or functionalities similar to those of the exemplary generative language model system 400 from [reference missing]. Fig. 4A, of the exemplary generative LM 430 from Fig. 4B-4C, the exemplary computing device 500 from Fig. 5 and / or the exemplary data center 600 from Fig. 6 are similar, to be executed.

[0035] System 100 can implement at least part of an artificial intelligence (AI) pipeline, such as a reconstruction AI and / or reconstruction pipeline. For example, System 100 can process data from one or more data sources representing a scene and / or a multitude of images of the scene for tasks such as synthesizing new views, dynamic scene reconstruction, and / or motion estimation. System 100 can be used to generate data for further processing by any of the systems described herein, including, but not limited to, autonomous vehicle systems, augmented reality systems, medical imaging systems, industrial automation systems, robot vision systems, virtual reality systems, computer graphics systems, and / or security monitoring systems.

[0036] In general, the reconstruction pipeline can contain operations performed by System 100. For example, the reconstruction pipeline can include one or more estimator stages, motion stages, loss stages, and / or reconstructor stages.

[0037] In some implementations, System 100, which implements the reconstruction pipeline, can be a collection of multi-view frames (F t ) received and / or received along T timestamps t ∈ {t1, ... , t T}, t1 = 0 < ... < t T = 1 are recorded, with at least one (e.g., each) frame containing M images Ft={Iit}i=1M System 100 can use the collection of multi-view frames to train and implement an image model that performs image synthesis with new views for general directions d ∈ S. 2 and the time t ∈ [t1, t T] enables. That is, System 100 can use a reconstruction pipeline for new views (hereafter also referred to as the "reconstruction pipeline"), which can be described as (Equation 1): t↦Ψt={ψt|ψ:ℝ+→V} where Ψ t the rendering model can be, ψ t can be an allocation function, the V of the functions ψ: ℝ + → V can be a vector space. V can denote a different vector space for at least one (e.g., every) ψ. That is, V can represent different properties of the scene elements (e.g., positions, densities in varying dimensions). For example, ψ can be a time-dependent position of n particles, representing a deformation geometry, V = ℝ n×d , represent. For example, ψ can model time-dependent image intensity values ​​for which V = C 1 (ℝ d ) = {f|f: ℝ d → ℝ, ∇f exists and is continuous} with d = 2.

[0038] In some implementations, a velocity field can be a time-dependent function (equation 2): v:ℝd×ℝ+→ℝd where d (e.g., d = 2 or d = 3) can be the spatial dimensionality of the scene. That is, d = 2 can correspond to 2D space and d = 3 can correspond to 3D space. The velocity fields can represent deformations (also called "flows") in space Φ. t : ℝ d → ℝ d define by an ordinary differential equation (ODE) (Equation 3): {∂∂tϕt(x)=v(ϕt(x),t)ϕ0(x)=x

[0039] System 100 can be included using dynamic reconstruction (e.g., in a dynamic reconstruction pipeline) by positioning a particle x in space along a curve, x(t) = ϕ t(x) is transported. That is, the transport can model time-dependent deformations. For example, if the underlying scene representation can be controlled particle-based (e.g., as with a geometry explicitly described by a point cloud), Φ t (x) can be applied to at least one (e.g., each) geometric element to model the deformation over time. That is, a flux can also be used to model a time-dependent function ψ. t : ℝ d → ℝ d using a reference function, ψ o (e.g. ψ) t = ϕ t * ψ t ) to define. For example, d = 2 for image intensities and / or d = 3 for volume density.

[0040] Furthermore, the inclusion of flows can imply that the system 100 ϕ t (x) determines that (equation 4) can be: ϕt(x)=x+∫0tv(ϕs(x),s)ds where x can be an initial position of the particle, v can be the velocity field representing deformations, and s can be a time variable representing the integration limits. That is, equation 4 can be used to model the trajectory of the particle over time. In some implementations, ϕ t (x) can be determined for at least one (e.g., each) evaluation of the model during training and inference. Often, determining ϕ can be t (x) prevent the training of flow-based models from converging to solutions that represent the expression of complex deformations, making it a computationally inefficient method for dynamic reconstruction.

[0041] In some implementations, at least one time-dependent reconstruction function ψ ∈ Ψ can induce a flow that is related to ψ t agrees and using ψ tand can be determined by one or more of its corresponding partial derivatives. That is, a function defined by ψ t Induced flow can be a reconstruction flow, represented by ϕ t Furthermore, the ϕ t The generating velocity field can be a reconstruction velocity field, represented by v. In some implementations, the system can use Equation 3 to impose a constraint on generating the velocity field v with respect to ψ. t to determine. That is, the restriction can ensure that the velocity field v is identical to that of ψ. t The provided dynamics match. For example, V = ℝ n×d , where ψ can contain n curves, ψ = {γ i} for at least one (e.g., each) γ i can coincide with a curve of a reconstruction flow that occurs at γ0i begins. That is, a limitation in generating the velocity field can be (Equation 5): v(γti,t)=ddtγti,∀1≤i≤n

[0042] Furthermore, for example, where ψ: ℝ + → C 1 (ℝ d ), a reconstruction flow can be determined using additional specifications for ψ. That is, equation 5 can be used to determine consistency in ψ. t to maintain the divergence constraints induced by v. In this example, ψ t can be quantified by determining that ψ t a thrust of ψ0 through ϕ t can be, for the ψ t can be reversed locally in time, where a continuity equation can be used to define a constraint for ϕ t to determine (equation 6): ∂∂tψt(x)+div(ψt(x)v(ψt(x),t))=0,∀x∈ℝd

[0043] Furthermore, div(ψ t (x)v(ψ t(x), t)) equals (∇ x ψ t (x), v(ψ t (x), t)) + div(c). Thus, a reconstruction velocity field v can contain a multitude of degrees of freedom. That is, the constraints provide several degrees of freedom for determining the velocity field.

[0044] In some implementations, System 100 can use a velocity field u: ℝ to model and / or enable flow rematching. d × ℝ + determine that v can adapt. That is, the constraint aligns u with the expected formations defined by ψ. t can be described. Furthermore, since u can also contain some a priori probabilities about the underlying possible deformations, u can be restricted to a prior class of velocity fields, denoted by P. The restrictions for v with respect to ψ tcan enable an optimization function for flow matching, using u in questions 5 and 6 shown below (equation 7): u=argminu∈P∫01ρ(u(⋅,t),ψt)dt where ρ can be either (equation 8) ρI(u(⋅,t),ψt)=∑i=1n‖u(γti,t)−ddtγti‖2 or (equation 9) can be: ρII(u(⋅,t),ψt)=∫|∂∂tψt(x)+div(ψt(x)u(x),t)|2dx

[0045] Furthermore, the solution for u can be the closest projection (e.g., for minimizing the discrepancy, best-fit projection) of the reconstruction flow P. That is, the alignment between u and the construction flow can be determined. In some implementations, the integral of equation 7 can be approximated by a sum analyzed using a random drawing of {t1}~U[0,1].

[0046] In some implementations, given u ∈ P from equations 7-9, u can be used to control the underlying reconstruction flow. That is, a flow rematching loss, L RM , a flow-matching loss (e.g., the reconstruction flow) that attempts to re-match u. L RM can be defined as (Equation 10): LRM(θ)=∫01ρ(u(⋅,t),ψt)dt where θ are the parameters of ψ t They can be. The readjustment loss L RM can be associated with a loss of reconstruction, L REC ψ parameters of θ are used to obtain a final loss (Equation 11): L(θ)=lRM(θ)+λLREC(θ) where λ > 0 can be a tuned hyperparameter.

[0047] In some implementations, System 100 can apply the equations above (e.g., equations 1–11) to a rendering model for Gaussian splats (or any neural representation). For example, an image model using Gaussian splats can be parameterized by a collection of 3D Gaussian distributions, modified by color and opacity parameters. {μi,∑i,ci,αi}i=1n are extended, where µ i ∈ ℝ 3 the Gaussian mean i th designated, Σ i ∈ ℝ 3×3 the covariance matrix can be, c i ∈ ℝ 3 the project color can be and α i ∈ ℝ the opacity can be. The 3D Gaussian distributions can be projected onto the image plane to form a collection of image-plane Gaussian distributions, which are defined by {μ2Di,∑2Di} are parameterized. Given K, where E denotes the intrinsic and extrinsic camera transformations, system 100 can determine the Gaussian parameters of the image plane using a point rendering formula. The 3D Gaussian center can be projected by (Equation 12): μ2D=KEμ(Eμ)z and the covariance matrix can be projected by (Equation 13): ∑2D=JEΣETJT where J denotes the Jacobian matrix of the affine transformation from Equation 12). That is, Equation 12 can be used to determine the position of the 2D projection, and Equation 13 can be used to determine the shape and orientation of the 2D Gaussian distribution in the image plane.

[0048] Additionally, an image pixel I(p) can be obtained by the system 100 performing alpha blending of the (depth-ordered) visible Gaussian distributions (equation 14): I(p)=∑i=1nciαiσi(p)∏j=1i−1(1−αjσj(p)) where σi(p)=exp(−12(p−μ2Di)T(Σ2Di)−1(p−μ2Di)). This means that Equation 14 can be used to model the blending of multiple Gaussian distributions based on depth order and visibility (e.g., to output a composite pixel intensity in the rendered image).

[0049] Additionally, a time-dependent Gaussian Splat can create a model of the shape {μi(t),∑i(t),ci(t),αi(t)}i=1n be. For example, {µ i (t)} with the parameterization of the framework with V = ℝ 3×3 , ψ = {µ i (t)} agree. That is, the framework can define the Gaussian parameters as time-varying functions that change within the specified vector space (e.g., representing dynamic spatial transformations of scene elements over time).

[0050] In some implementations, when Φ = µ(t) is used to model Gaussian centers, the velocity field v ∗ (·,t) can be obtained by minimizing the squared error between the velocity field and the time derivative of the Gaussian centers (Equation 15): v*(⋅,t)=arg minv∈C∫‖v(μ(t),t)−μ˙(t)‖2dμ where v ∗ v can be the optimal (or desired) velocity field, v can be any candidate velocity field within the constraint set C, C can be a set of velocity fields that satisfy specific deformation constraints, µ can be the Gaussian center positions, µ̇ can be the time derivative of the Gaussian centers.

[0051] Furthermore, the function shown below (equation 16) can be used to monitor the time-dependent reconstruction model: L=∫∫‖v*(μ,t)−μ˙‖2dμdt where L can be the loss function used to minimize the discrepancy between the predicted and actual motion of the Gaussian centers. That is, Equation 16 can model (or measure) the alignment of the velocity field with the true motion of the Gaussian centers in both space and time.

[0052] In some implementations, C can model various deformations. For example, rigid deformations can be modeled by (Equation 17): C={v:ℝ3×ℝ→ℝ3|v(x,t)=Ax+b,A∈ℝ3×3,A=−AT,b∈ℝ3} where the output can be a restricted least squares problem (e.g., the solution for A and b that best fits the rigid motion constraint for Equation 15) to solve Equation 15. That is, Equation 17 can be used to impose the rigid body constraint by restricting the shape of the velocity field to linear transformations that conserve volume.

[0053] In some implementations, multiple rigid motions can be modeled by (Equation 18): C={v|v(x,t)=∑j=1Kwj(x,t)(Ajx+bj),Aj∈ℝ3×3,Aj=−AjT,bj∈ℝ3} where w j (x,t) can be weights for at least one (e.g., each) rigid motion component, A j and b jThe transformations can be rigid, and C allows a combination of several rigid transformations. That is, Equation 18 allows the modeling of one or more complex movements (e.g., jointed movements, deformable object transformations, dynamics of multi-part objects) by superimposing several rigid body transformations, weighted by w. j (x, t).

[0054] Furthermore, u in equations 7-11 can refer to a candidate velocity field that attempts to satisfy the constraints ψ tto adapt to the defined reconstruction flow. For example, in equations 7-9, u can find the optimal u that minimizes a loss function of the flow adaptation with respect to the target flow (e.g., the reconstruction flow). The optimization process can project u onto a previous class of velocity fields P to satisfy the constraints defined by the dynamic scene reconstruction framework. In some implementations, v and v may be ∗ Equations 15-18 refer to the actual velocity field that dictates the motion of the deformation of specific elements within the reconstruction model. That is, v can be optimized to agree with the derivative of the Gaussian center positions µ̇(t), where v* can be the optimal velocity field resulting from the optimization. Thus, u can be used to provide an approximate or candidate velocity field that agrees with those of ψ.t corresponds to derived underlying constraints, and v can be used to model and control the movement of Gaussian Splats or other scene elements in space (e.g., derived and monitored by specific constraints on the object's movement, such as matching µ̇(t)).

[0055] System 100 can contain at least one image estimator 108. In some implementations, the estimator stage can refer to the stage in the reconstruction pipeline where the image estimator generates an initial estimated image of the scene based on the input images using an image rendering model. The image estimator 108 in the estimator stage can cause an image rendering model to generate an estimated image of a scene based on at least a set of images of the scene, where at least one image in the set of images is associated with at least one from a different time point (e.g., at different time steps t1 and t2 to represent temporal changes in the scene) or from a different view (e.g., captured from different camera positions to provide multiple perspectives).This means that the Image Estimator 108 can use these different images to create an estimated image that represents the scene by integrating information from varying times or viewpoints. For example, the estimation might involve combining features from the input images to generate a view of the scene that accounts for the captured spatial and temporal changes. In this example, the Image Estimator 108 can use parameters derived from the input images, such as positional data. with and transformations defined in equations 12 and 13 to allocate the scene elements in the estimated image. Furthermore, the image estimator 108 can be configured to use varying resolutions or inconsistencies in the input images to output a smooth and accurately estimated image. In the estimator stage, the initial estimate can be created as a base image for further processing and refinement. For example, the image estimator 108 can generate an estimated image that aligns with the expected scene structure without performing full flow-based deformations. For instance, the image estimator 108 can use information about the captured viewpoints to correct perspective distortions in the initially estimated image.

[0056] In some implementations, the image estimator 108 can maintain, execute, train, and / or update one or more machine learning models during the estimation stage. In some implementations, the one or more machine learning models can include all types of prediction models capable of estimating initial scene representations based on input data. For example, the one or more machine learning models can be trained and / or updated to identify correlations between images captured at different times or in different views to improve estimation accuracy. The one or more machine learning models can be a transformer-based model (e.g.,The machine learning model may be or include a generative pre-trained / updated transformer (GPT) model, a convolutional neural network (CNN) for image feature extraction, a recurrent neural network (RNN) for temporal pattern recognition, or any graph neural network (GNN) for modeling relationships between multi-view inputs. In some implementations, one or more of these machine learning models may be or include a variational autoencoder (VAE) model. The Image Estimator 108 can execute the machine learning model to generate outputs (such as initial scene estimates, set view predictions, or image corrections based on input variability).The image estimator 108 can receive data to be provided as input for one or more machine learning models, which may contain multi-view images, scene metadata and / or pre-processed feature maps.

[0057] The Image Estimator 108 can include one or more artificial intelligence models (e.g., machine learning models, supervised models, neural network models, deep neural network models), rules, heuristics, algorithms, filters (e.g., Kalman filters), functions, or various combinations thereof to perform operations that include generating initial scene estimates, identifying image features and inconsistencies in input data, such as combining multi-view images to generate an estimated image that represents the scene structure. In some implementations, the Image Estimator 108 can be trained / updated independently of other systems or devices described herein (e.g., Motion Modeler 112, Loss Modeler 116, and / or Reconstructor 120).In some implementations, the training of the at least one image estimator 108 can be performed at least partially in conjunction with the training of the motion modeler 112, the loss modeler 116, and / or the reconstructor 120. In some implementations, the image estimator 108 can output one or more estimated images of a scene (e.g., initial estimates of the scene, corrected multi-view compositions, or preliminary reconstructed views). For example, the image estimator 108 can generate an estimated image by integrating spatial and temporal features from multiple input images to create a view of the scene. For example, the image estimator 108 can use learned models for variations in the input data, such as different lighting conditions or occlusions.In some implementations, the estimated images can be provided to the other systems or devices described herein. That is, the image estimator 108 can be a source of initial scene estimates for further refinement and processing. For example, it can provide the estimated images to the motion modeler 112, the loss modeler 116, and / or the reconstructor 120 for further processing and alignment with additional scene information.

[0058] System 100 can contain at least one motion modeler 112. In some implementations, the motion stage can refer to the stage in the reconstruction pipeline where the motion modeler 112 determines and applies criteria for motion, such as constraints or learned models, to describe the motion and deformation of scene elements over time. The motion modeler 112 in the motion stage can determine one or more criteria (e.g., a learnable model) for motion based on the output of a minimization (e.g., a minimization problem that satisfies constraints) of one or more functions (e.g., a cost, loss, or objective function) that include at least one of the following: (i) the rigidity constraint, (ii) the continuity constraint, or (iii) a constraint derived from the machine learning model.For example, the criteria can be used to guide one or more dynamic reconstructions of the scene (e.g., in the loss stage). In this example, the motion criteria can include a velocity field representing a variety of deformations in space over time (e.g., Equation 4, which is used to represent particle trajectories based on integrated velocity fields). That is, the one or more motion criteria can include at least one of the following: (i) a rigidity constraint that limits changes in the shape of objects over time, or (ii) a continuity constraint that assumes smooth transitions in one or more object motions within the scene. For example, determining criteria might involve performing a minimization problem to determine the motion parameters that satisfy the defined constraints.In this example, the motion modeler can output 112 corrected motion fields based on the calculated criteria. For example, determining criteria might involve using predefined models to predict scene dynamics and updating motion parameters accordingly.

[0059] In general, motion criteria can be implemented as a flow without requiring flow integration. That is, the Motion Modeler 112 can determine the motion criteria from defined constraints (e.g., rigidity or continuity) or learned models. For example, the Motion Modeler 112 can compute velocity fields representing deformations without needing to integrate the fields over time, thus reducing computational complexity. Furthermore, the Motion Modeler 112 can update the motion criteria in response to changes in the scene, such as varying object trajectories or deformation behavior.

[0060] Furthermore, the Motion Modeler 112 can be configured to integrate (use and / or incorporate) learned motion models that adapt to varying scene conditions. The motion stage can output motion constraints based on the expected behavior of scene elements. For example, the Motion Modeler 112 can generate velocity fields and deformation parameters for use in subsequent stages. For instance, the Motion Modeler 112 can refine motion predictions using a combination of physical and learned models.

[0061] In some implementations, the Motion Modeler 112 can maintain, execute, train, and / or update one or more machine learning models during the motion stage. In some implementations, the one or more machine learning models can include all types of prediction models capable of analyzing dynamic scene data and estimating motion parameters. For example, the one or more machine learning models can be trained and / or updated to learn motion constraints and adapt to new scene configurations. The one or more machine learning models can be a transformer-based model (e.g.,The machine learning model may be or include a generative pre-trained / updated transformer (GPT) model, a convolutional neural network (CNN) for extracting spatial features and correlations, a recurrent neural network (RNN) for capturing temporal sequences, or any graphical neural network (GNN) for analyzing relationships between scene elements. In some implementations, the one or more machine learning models may be or include a variational autoencoder (VAE) model. The Motion Modeler 112 can execute the machine learning model to generate outputs (such as velocity fields, motion constraints, or dynamic scene predictions). The Motion Modeler 112 can receive data to be provided as input to the one or more machine learning models, which may include sequences of multi-view images, a priori motion probabilities, and scene metadata.

[0062] The Motion Modeler 112 can include one or more artificial intelligence models (e.g., machine learning models, supervised models, neural network models, deep neural network models), rules, heuristics, algorithms, filters (e.g., Kalman filters), functions, or various combinations thereof to perform operations involving motion estimation, constraint application, and flow prediction, such as estimating motion fields based on scene constraints and inputs. In some implementations, the Motion Modeler 112 can be trained / updated independently of other systems or devices described herein (e.g., Image Estimator 108, Loss Modeler 116, and / or Reconstructor 120).In some implementations, the training of the at least one motion modeler 112 can be performed at least partially in conjunction with the training of the image estimator 108, the loss modeler 116, and / or the reconstructor 120. In some implementations, the motion modeler 112 can output one or more criteria for motion (e.g., corrected velocity fields, deformation models, or motion parameters). For example, the motion modeler 112 can generate velocity fields that align with the learned or predefined motion models. For example, the motion modeler 112 can dynamically update the velocity fields as new data becomes available (e.g., is received, identified, and / or accessed). In some implementations, the criteria for motion can be provided to the other systems or devices described herein.This means that the motion modeler 112 can be a source of updated motion constraints for the loss modeling or reconstruction stages. For example, the motion modeler 112 can provide the loss modeler 116 with motion criteria for updating an image rendering model.

[0063] System 100 can contain at least one Loss Modeler 116. In some implementations, the loss stage can refer to the stage in the reconstruction pipeline where the Loss Modeler 116 evaluates and improves the image rendering model by calculating and minimizing discrepancies between estimated and reference images. At the loss stage, the Loss Modeler 116 can update the image rendering model based on at least the estimated image, the plurality of images, and one or more motion criteria associated with the estimated image. For example, the Loss Modeler 116 can analyze (or model) a loss that occurs when comparing the estimated image with ground truth (e.g., L). REC ) is assigned, and a loss that results from adapting the image flow represented in the estimated image to image flow constraints (e.g., L RM) is assigned (as described with reference to Equations 10-11 and Equations 15-16). Furthermore, Equations 12-14 and / or Equations 17-18 can be used for Gaussian splats, where Loss Modeler 116 can minimize the discrepancy between the projected Gaussian parameters and the observed data. That is, Loss Modeler 116 can adjust the image model to align with the actual (e.g., real-time or near-real-time) scene data. Determining the loss might involve, for example, calculating the pixel-wise and flow-wise error between the estimated image and the input data. In this example, Loss Modeler 116 can update the rendering model parameters to reduce one or both types of error.

[0064] Furthermore, the loss modeler 116 can be configured to weight different loss components (e.g., L).REC and L RM ) to be dynamically adjusted during training. For example, the Loss Modeler 116 can determine a minimization problem to obtain optimized model parameters that best fit the scene constraints (e.g., using one or more of Equations 10-11 and / or 15-16 to minimize combined losses). In this example, one or more of Equations 10-11 and / or 15-16 can define how the Loss Modeler 116 adjusts the model based on discrepancies in image flow and reconstruction quality.

[0065] In some implementations, the loss modeler 116 can refit the velocity fields to the estimated flow to refine the final model output. That is, the reconstruction loss can correspond to a measure of discrepancy between the estimated image and the multitude of images of the scene (e.g., ground truth images and / or other reference data), and the refit loss can correspond to a measure of deviation between the estimated image and the one or more criteria for motion associated with an image flow (e.g., the criterion used to determine motion during scene reconstruction).

[0066] In some implementations, the Loss Modeler 116 can maintain, execute, train, and / or update one or more machine learning models during the loss stage. In some implementations, the one or more machine learning models can include any type of supervised or unsupervised model capable of learning loss functions for dynamic scene reconstruction. For example, the one or more machine learning models can be trained and / or updated to learn loss functions that can balance (or proportionally weight) reconstruction quality and motion orientation. The one or more machine learning models can be a transformer-based model (e.g., a transformer-based model).The Loss Modeler 116 can be or include a generative, pre-trained transformer (GPT) model, a convolutional neural network (CNN) for learning image representations, a recurrent neural network (RNN) for modeling temporal dependencies in loss evolution, or any graphical neural network (GNN) for identifying relationships in scene data. In some implementations, the one or more machine learning models can be or include a generative adversarial network (GAN) model. The Loss Modeler 116 can execute the machine learning model to generate outputs (e.g., updated loss values, updated model parameters, or refined scene reconstructions). The Loss Modeler 116 can receive data to be provided as input to the one or more machine learning models, which may include image estimates, ground-truth data, and computed motion fields.

[0067] The Loss Modeler 116 can include one or more artificial intelligence models (e.g., machine learning models, supervised models, neural network models, deep neural network models), rules, heuristics, algorithms, filters (e.g., Kalman filters), functions, or various combinations thereof to perform operations involving loss calculation, model updating, and error minimization, such as calculating combined losses (e.g., L). REC and L RM) to update the image model parameters. In some implementations, the loss modeler 116 can be trained / updated independently of other systems or devices described herein (e.g., motion modeler 112, image estimator 108, and / or reconstructor 120). In some implementations, the training of the at least one loss modeler 116 can be performed at least partially in conjunction with the training of the motion modeler 112, the image estimator 108, and / or the reconstructor 120. In some implementations, the loss modeler 116 can update the image rendering model such that the estimated image better aligns with the input data and the defined motion constraints. For example, the loss modeler 116 can refine the image parameters to reduce visual and flow errors based on the defined loss functions.For example, the Loss Modeler 116 can use the updated parameters to improve reconstruction quality in dynamic scenes.

[0068] In some implementations, the updated image rendering model can be implemented with other systems or devices described herein (e.g., used for inference). That is, the Reconstructor 120 can use the updated model parameters to generate images with new views consistent with the input scene data. For example, the Reconstructor 120 can render scene elements based on the updated model parameters provided by the Loss Modeler 116.

[0069] System 100 can contain at least one Reconstructor 120. In some implementations, the Reconstructor stage may refer to the stage in the reconstruction pipeline where Reconstructor 120 synthesizes new views and renders scene images based on the estimated parameters and motion criteria. At the Reconstructor stage, Reconstructor 120 can apply scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on the multitude of images of the scene. That is, Reconstructor 120 can use the refined parameters from the image rendering model to generate new views of the scene that conform to (or closely approximate) the defined spatial and temporal constraints.Additionally, one or more equations 1-3 can be used to model scene elements and temporal positions for the rendering process. For example, the reconstructor can synthesize 120 new views using motion criteria or fill gaps (e.g., between existing images). In this example, the reconstructor can output 120 new images based on input data and predicted motion fields.

[0070] Furthermore, the Reconstructor 120 can be configured to generate multi-view reconstructions that align with the input images and comply with scene constraints. The Reconstructor stage can produce outputs that are visually consistent and accurately represent dynamic changes in the scene. For example, the Reconstructor 120 can generate images for viewpoints not present in the input data, filling in missing information based on model predictions. In this example, one or more Equations 1-3 can be used to calculate the positions of scene elements over time for accurate rendering. For instance, the Reconstructor 120 can apply learned models to refine scene reconstructions based on observed data.

[0071] In some implementations, the Reconstructor 120 can maintain, execute, train, and / or update one or more machine learning models as part of the reconstructor stage (and / or for its implementation). For example, the image rendering model can include at least one of (i) a Gaussian splatting model or (ii) a neural radiation field (NeRF) model. It should be understood that the pipeline and modeling framework can support any learnable rendering model by incorporating a priori probabilities at the rendered image level, without being restricted to Gaussian splatting or NeRF models and without being limited to specific rendering techniques. For example, various modeling frameworks can be used, such as...However, without limitation, voxel-based models, point cloud models, implicit neural representations, surface light field models, and / or other rendering models that can utilize learnable parameters for image synthesis are supported. In some implementations, the one or more machine learning models may include all types of models capable of generating scene reconstructions based on estimated parameters. For example, the one or more machine learning models may be trained and / or updated to learn reconstruction parameters that produce realistic scene images. The one or more machine learning models may be a transformer-based model (e.g.,The machine learning model may be or include a generative, pre-trained / updated transformer (GPT) model, a convolutional neural network (CNN) for learning image synthesis, a recurrent neural network (RNN) for capturing temporal coherence in image sequences, or any graphical neural network (GNN) for learning relationships between scene elements. In some implementations, one or more of these machine learning models may be or include a variational autoencoder (VAE) model. The Reconstructor 120 can execute the machine learning model to generate outputs (e.g., images with new views, scene interpolations, or refined reconstructions). The Reconstructor 120 can receive data or a trained / updated model (e.g.,an image rendering model (trained / updated using the image estimator 108, the motion modeler 112, and / or the loss modeler 116) to perform one or more inference operations or tasks. For example, the received data (e.g., motion fields, image parameters) can be provided as input for the image rendering model (e.g., for one or more machine learning models), which may contain refined model parameters for generating new images.

[0072] The Reconstructor 120 can contain one or more artificial intelligence models (e.g., machine learning models, supervised models, neural network models, deep neural network models), rules, heuristics, algorithms, filters (e.g., Kalman filters), functions, or various combinations thereof to perform operations that include scene reconstruction, synthesis of new views, and image rendering, such as generating scene images based on the estimated model parameters. In some implementations, the Reconstructor 120 can be trained / updated independently of other systems or devices described herein (e.g., Image Estimator 108, Motion Modeler 112, and / or Loss Modeler 116).In some implementations, the training of the at least one reconstructor 120 can be performed at least partially together with the training of the image estimator 108, the motion modeler 112 and / or the loss modeler 116.

[0073] In some implementations, Reconstructor 120 can output a scene reconstruction to render an estimated image for one or more viewpoints and time intervals. For example, Reconstructor 120 can generate new views of the scene based on parameters and motion fields provided by other systems in the pipeline. For example, Reconstructor 120 can generate interpolated images for missing or obscured views. In some implementations, the scene reconstruction can be provided to the other systems or devices described herein. That is, Application 124 can use the rendered scene images for further processing or visualization in various contexts. For example, Application 124 can apply the rendered images in virtual reality systems, autonomous navigation, or any other context where dynamic scene representations are implemented.

[0074] System 100 can contain at least one application 124. The application 124 can be a downstream device or system that can utilize the rendered scene images generated by the reconstructor 120 for various tasks or functionalities. In some implementations, the application 124 can process or analyze the rendered images to extract information, generate visualizations, and / or support decision-making processes. For example, the application 124 can be configured to operate within an autonomous navigation system that uses the reconstructed scenes to determine object trajectories or environmental changes for path planning. For example, the application 124 can be integrated into an augmented reality system to overlay dynamic scene elements based on the reconstructed images onto a user's view.Application 124 can also integrate the rendered scenes into simulation environments, virtual reality (VR) platforms, and / or any other system that outputs or inputs high-quality visual data of dynamic scenes. That is, Application 124 can be or include an interface for using the output of the image rendering pipeline in operation. In some implementations, Application 124 can receive updates from Reconstructor 120 and adjust its processing based on refined scene data. For example, Application 124 can apply the rendered images in virtual reality systems, autonomous navigation, robot vision, augmented reality systems, medical imaging, industrial automation, security monitoring, digital content creation, gaming environments, cinematic visual effects, and / or any other implementation, including dynamic scene rendering.

[0075] With reference to Fig. Section 2 shows an exemplary flowchart illustrating a method for dynamically reconstructing new views based at least partially on flow rematching, according to some implementations of this disclosure. It is understood that this and other arrangements described herein are presented only as examples. Other arrangements and elements (e.g., machines, interfaces, functions, arrays, groupings of functions, etc.) may be used in addition to or instead of those shown, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that can be implemented as single or distributed components, or in conjunction with other components, in any suitable combination and location.Various functions described herein as being performed by entities can be executed by hardware, firmware, and / or software. For example, various functions can be performed using one or more processors executing instructions stored in one or more memory locations. For example, in some implementations, the system and procedures described herein can be implemented using one or more generative language models (e.g., as in ). Fig. 4A-4C), one or more computing devices or components thereof (e.g., as described in Fig. 5 described) and / or one or more data centers or components thereof (e.g. as described in Fig. 6) will be implemented.

[0076] Now, with reference to Fig. 2 Each block of the procedure 200 described herein contains a computational process that can be performed using any combination of hardware, firmware, and / or software. For example, various functions can be performed using one or more processors executing instructions stored in one or more memory locations. The procedure can also be embodied as computer-usable instructions stored on computer storage media. The procedure can be provided by a standalone application, service, or hosted service (alone or in combination with another hosted service), as a microservice via an application programming interface (API), or as a plugin for another product, to name just a few. Furthermore, the procedure 200 is described, for example, in relation to the system of Fig. 1 described. However, this procedure can additionally or alternatively be performed by any system or any combination of systems, including, but not limited to, the systems described herein.

[0077] Fig. Figure 2 is a flowchart illustrating Method 200 for generating, updating, and applying an image rendering model based on estimated images and motion criteria, according to some implementations of the present disclosure. Various operations of Method 200 aim to improve the accuracy and computational efficiency of dynamic scene reconstruction pipelines by integrating flow rematching techniques and iterative refinement of motion fields. Existing systems typically use static representations or predefined deformation models, which can lead to inaccuracies in managing complex scenarios involving variable motion patterns and dynamic object interactions.Such technical limitations can arise when conventional systems are unable to accurately model motion or deformation fields, leading to visual distortions, increased latency, and inefficient use of computing resources. Method 200 and the systems described in... Fig. As described in section 2, these problems are addressed by implementing a dynamic reconstruction process that evaluates and updates the image rendering model based on learned motion criteria, such as velocity fields, to iteratively refine scene representations in real time (or near real time). This improves the output of the rendering model when representing complex deformations and motion dynamics, thereby enhancing the accuracy and computational efficiency of the scene reconstruction pipeline.

[0078] Procedure 200, in block 210, involves causing an image rendering model to generate an estimated image of a scene based on at least one image of the scene. That is, at least one image from the multitude of images can be associated with at least one image from a different time point (e.g., t1, representing an initial state of the scene; t2, representing a subsequent state showing motion) or a different viewpoint (e.g., different camera positions d1 and d2 to capture varying viewpoints). In some implementations, the multitude of images of the scene can include a multitude of multi-view images. For example, the multitude of multi-view images can correspond to a multitude of different viewpoints captured at a multitude of different times.In this example, capturing the multi-view images at different times provides the data used to define the velocity fields that describe movements between the views.

[0079] Procedure 200, in block 220, includes the determination of one or more criteria for motion based on the output of a minimization. That is, the processing circuits can determine the one or more criteria for motion based on the output of a minimization. In general, the one or more criteria for motion can correspond to a machine learning (ML) model that is trained / updated to generate one or more deformations of the estimated image based on parameters derived from one or more historical scenes. This means the processing circuits can determine a velocity field directly from the constraints or learned models (e.g., without performing flow integration), thus avoiding the computational overhead associated with integrating velocity fields over time.In some implementations, the learnable model can be trained / updated using parameters from previously captured dynamic scenes to learn how to generate realistic deformations of an estimated image. During the training phase, for example, the processing circuits can extract features such as object trajectories, velocity fields, and deformation patterns from labeled datasets of dynamic scenes to build a predictive model. Furthermore, during the inference phase in Block 220, the processing circuits can use the trained / updated model to generate expected motion criteria for captured (e.g., recently captured and / or historically captured) scenes based on visual input. For example, the processing circuits can input multi-view images of the scene into the model, which can then output predicted velocity fields that describe the expected motion.

[0080] The criteria for motion can include one or more velocity fields representing a variety of deformations in space over time (e.g., defining the deformations in space over time, which can be constraints or criteria for how motion is represented in scene reconstruction). For example, the processing circuits can determine a velocity field representing a rigid transformation of an object (e.g., rotation without deformation). Alternatively, the processing circuits can determine a velocity field capturing a smooth deformation of a flexible object. In some implementations, the processing circuits can use a minimization problem that may satisfy constraints on one or more functions (e.g., cost, loss, or objective function).In this example, one or more of equations 5 and 6 can be used to define and constrain the velocity fields according to the rigidity and continuity conditions. The constraints of one or more functions can satisfy at least one of the following conditions (e.g., criteria that guide the dynamic reconstruction of the scene): (i) the rigidity constraint, (ii) the continuity constraint, and / or (iii) a constraint derived from the machine learning model. That is, the rigidity constraint can restrict changes in the shape of objects over time. For example, the rigidity constraint (e.g., by limiting nonlinear deformations, imposing only rotation and translation) can influence the interpretation of velocity fields, such as by specifying that the motion of objects must not result in a change in shape.Furthermore, the continuity constraint (e.g., by ensuring gradual changes in velocity fields) assumes smooth transitions in one or more object movements within the scene. For example, the continuity constraint can influence the interpretation of velocity fields, such as specifying that the velocity field must represent smooth transitions in one or more object movements.

[0081] Procedure 200, in block 230, includes updating the image rendering model based on at least the estimated image (e.g., visual data) and one or more criteria for motion (e.g., velocity fields and / or flows). In some implementations, the image rendering model may include at least one of (i) a Gaussian splatting model and / or (ii) a neural radiation field (NeRF) model. For example, the Gaussian splatting model may be trained / updated to represent the scene by projecting three-dimensional (3D) points onto a two-dimensional (2D) plane using one or more Gaussian functions. In this example, the processing circuits may compute parameters of the Gaussian splat, such as center positions and covariances, using equations 12 and 13 to update the image rendering model.For example, the NeRF model can be trained / updated to synthesize new views of the scene from multiple 2D images by learning a continuous volumetric scene function. In this example, the processing circuits can adjust the density and color fields within the NeRF model to match the input images and motion criteria.

[0082] In some implementations, updating the image rendering model may involve minimizing reconstruction loss and rematching loss. For example, the processing circuits (e.g., using and / or applying one or more equations 10 and 11) may formalize a minimization problem to v ∗ to obtain (e.g., the optimal velocity field aligned with the Gaussian center motion u(t)). In this example, the processing circuits v ∗re-adjust to a reconstruction model to adjust the rate to obtain a re-adjustment loss. That is, v ∗ can represent the optimal deformation field, and the readjustment of v ∗ It can minimize discrepancies between predicted and observed motion paths. One or more equations 7-9 can be used to adjust u to the v. ∗ to adapt the derived flow. For example, the processing circuits can use equations 7-9 to optimize u such that it matches the expected dynamics of the scene based on v. ∗This example illustrates the reconstruction loss. In this example, the reconstruction loss can be evaluated to ensure that u (e.g., with consistent velocity patterns, while maintaining rigidity and / or continuity) aligns with the reconstructed motion model parameters. In some implementations, the reconstruction loss can correspond to a measure of discrepancy between the estimated image and the multitude of images of the scene (e.g., ground truths). In other implementations, the reconstruction loss can correspond to a measure of deviation between the estimated image and one or more criteria for motion associated with an image flow (e.g., used as a criterion to determine motion in scene reconstruction).

[0083] In some implementations, the processing circuitry can update the image rendering model based on at least the estimated image, the plurality of images, and / or one or more motion criteria associated with the estimated image. That is, the motion criteria can be represented by a velocity field (containing, for example, a physical a priori probability (rigidity), a continuity constraint, a learnable model, piecewise rigid motion, volume-preserving deformation, or any learned dynamic flow model). In some implementations, updating the image rendering model can include determining and analyzing an initial loss (e.g., L). REC ) contained by the processing circuits, specifically in assignment to analyze (e.g., compare) estimated images with ground truth information and / or a second loss (e.g., L RM) in assignment to analyze (e.g., fit) image flows represented in estimated images, with constraints on an image flow. That is, to update the image rendering model, the processing circuits can perform gradient-based optimization to minimize both reconstruction and refitting losses simultaneously (or sequentially). For example, the first loss can be determined based on the pixel-wise difference between the estimated and ground-truth images. In this example, the second loss can be based on the difference between the reconstructed flow and the expected flow, which is determined by the velocity fields v. ∗ Given, it can be determined. That is, the processing circuits can evaluate the first loss and the second loss, or combine them to iteratively update the model parameters.

[0084] In some implementations, updating the image rendering model may involve rematching a velocity field generated from the estimated image to a previous velocity field corresponding to the scene. That is, the rematching operation can occur between the generated velocity field (e.g., a portion of the estimated image) and a previous velocity field. For rematching, the processing circuits may, for example, use Equation 15 to determine the velocity field. ∗to obtain this by minimizing the squared difference between the predicted and true Gaussian center velocities. Furthermore, the motion criteria can include comparing and adjusting the velocity fields to refine the dynamic reconstruction. For example, the processing circuits can iteratively adjust (or update) the velocity field v to ensure consistency with motion constraints, such as rigidity or continuity. In this example, the updated fields can be used to refine the rendered scene representation for improved accuracy in dynamic reconstructions. For example, the processing circuits can optimize the v ∗ use to readjust the predicted movement to the observed scene dynamics.

[0085] In some implementations, L RECIn Block 230, the loss can be determined by modeling the difference between the estimated image generated by the image rendering model and the ground-truth images of the scene. The loss can quantify how accurately the rendering model represents the visual appearance of the scene, taking into account various factors (e.g., color, brightness, and / or spatial structure). In some implementations, if the rendering model is based on a Gaussian splatting representation, the reconstruction loss can be expressed as the pixel-wise mean squared error (MSE) between the rendered image I est (p) and the ground truth image I gt (p) are calculated: LREC=∑p‖Iest(p)−Igt(p)‖2 where p can be the pixel positions in the image. The loss function can penalize discrepancies between the rendered images and the ground-truth images, guiding the model to generate images that closely resemble the real-world scene.

[0086] In a NeRF model, reconstruction loss can be modeled by comparing the predicted color and density values ​​along the rays sampled from the scene with the observed color values ​​in the input images. For example, the processing circuitry can integrate the predicted scene function along at least one (e.g., each) ray to generate a synthesized image and then determine the MSE between the synthesized and ground-truth images.

[0087] In some implementations, the following processing can be used to determine the NeRF reconstruction loss: LREC=∑r∈R‖Cest(p)−Cgt(p)‖2 where r can be a ray passing through a pixel in the image, R can be a set of rays corresponding to the pixels in the image, C est (r) can be a color of the pixel, estimated by the NeRF model along the ray r, and C gt (r) can be the ground truth color of the pixel along the ray r.

[0088] In some implementations, C est (r) can be determined by integrating the colors and densities along the ray: Cest(r)=∑i=1NTi(1−exp(−σiδi))ci where Ti=exp(−∑j=1i−1σiδi) the accumulated permeability up to point i can be, σ i the density at the i-th point can be, δ i the distance between adjacent sampled points and c i the RGB color at the i-th point.

[0089] In block 230, the processing circuits can enable gradient-based optimization to reduce the reconstruction loss L. REC with the readjustment loss L RM to minimize and refine the image rendering model to output visually accurate images that follow the expected motion dynamics. In some implementations, v ∗ If ground-truth motion is available, it can be used to directly compare the reconstructed motion with the actual dynamics, minimizing the re-matching loss using known reference data. In some implementations, if model-constrained motion is used, u can be optimized to align with predefined constraints, such as constraints derived from physical a priori probabilities or learned models. It should be understood that v ∗can be determined (or modeled) if the true motion or deformation path of the scene elements is known or can be accurately estimated from external data sources. That is to say, v ∗ It can be used if it represents the known motion or deformation path of the scene elements, which the model can adapt to. The readjustment loss can be calculated as follows: LRM=∬‖v∗(μ,t)−μ˙‖2dμdt where the loss is the alignment of the candidate field v (from the model) with the true velocity field v ∗ can measure.

[0090] It should also be understood that u can be determined (or modeled) when attempting to satisfy constraints derived from the dynamic model (e.g., instead of fitting to a true reference flow). That is, u can be used when it represents the candidate flow that conforms to a priori probabilities or constraints defined by the scene reconstruction model. The refitting loss can be calculated as follows: LRM(u)=∫01ρ(u(⋅,t),ψt)dt where p measures how well u fits into the constraints ψ defined by the dynamic model t fits (e.g., instead of fitting to a true flow or reference flow).

[0091] Procedure 200 includes, in Block 240, the application of scene reconstruction using the image rendering model. In some implementations, the processing circuits can apply scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on the multitude of images of the scene. That is, the processing circuits can synthesize new views using motion criteria to fill gaps (e.g., between existing images, such as missing perspectives or obscured views). To apply the scene reconstruction, the processing circuits can use the updated image rendering model to generate synthesized views of the scene. That is, the processing circuits can apply the learned motion criteria, such as velocity fields. ∗The processing circuits can use these parameters to refine the scene representation. They can optimize the image rendering model by adjusting parameters (e.g., Gaussian splat centers, NeRF density fields) to reduce discrepancies between the synthesized images and the scene's input images. For example, a downstream application, such as autonomous vehicle navigation, can use the reconstructed scenes to improve path planning and obstacle detection based on the dynamic environment. Furthermore, the scene can be reconstructed using the updated image rendering model to output a temporally consistent sequence of images representing the dynamic scene evolution. For example, the reconstructed sequence can be used to create experiences in virtual or augmented reality applications by providing realistic cues for motion and depth.

[0092] Disclosed implementations may be found in a wide variety of different systems, such as automotive systems (e.g., a control system for an autonomous or semi-autonomous machine, a perception system for an autonomous or semi-autonomous machine), systems implemented using a robot or robotic platform, aviation systems, media systems, boat systems, intelligent area monitoring systems, systems for performing deep learning operations, systems for performing simulation operations (e.g., in a driving or vehicle simulation, in a robotics simulation, in a smart city or surveillance simulation, etc.), systems for performing digital twin operations (e.g., in conjunction with a platform or system for collaborative content creation, such as...).NVIDIA's OMNIVERSE and / or any other platform, system, or service that uses USD or OpenUSD data types), systems implemented using an edge device, systems that include one or more virtual machines (VMs), systems that perform operations to generate synthetic data (e.g., using one or more neural rendering fields (NERFs), Gaussian splat techniques, diffusion models, transformer models, etc.), systems that are at least partially implemented in a data center, systems that perform conversational AI operations, systems that implement one or more language models, such as one or more large language models (LLMs), one or more vision language models (VLMs), one or more multimodal language models, etc., systems for performing light transport simulations, systems for performing collaborative content creation for 3D assets (e.g., using data from a universal scene descriptor (USD), such as OpenUSD, computer-aided design (CAD) data, 2D and / or 3D graphics or design data and / or other data types), systems that are implemented at least partially using cloud computing resources and / or other types of systems.

[0093] With reference to Fig. 3A is Fig. Figure 3A presents an example of flow rematching using velocity fields, according to some implementations of the present disclosure. The flow rematching 300 can represent the deformation of an object over time, where ϕ(x, 0) can be the initial position of the object at time t = 0. As the object moves along the path defined by ϕ(x, t), the velocity field v(x, t) can specify the instantaneous velocity (e.g., direction and magnitude of motion) of the object at time t. At a specific point along the trajectory, denoted by 302, the instantaneous rate of change of position or the tangent to the curve can be determined by the derivative ddt (ϕ(x, t)). That is, vector 304 can represent the predicted direction and magnitude of the motion described by v(x, t). The velocity field v(x, t) can describe the continuous transformation of the object's position in space over time. The rematching process can be used to align the modeled velocity field v(x, t) with the observed flow of the object, ensuring that the trajectory ϕ(x, t) accurately captures the scene's dynamic motion. At time t = 1, the object's position can be represented by ϕ(x, 1).

[0094] With reference to Fig. 3B is Fig. Figure 3B is an example of flow rematching rendering using velocity fields, according to some implementations of this disclosure. 3D Box 310 represents a scene reconstructed using a typical Gaussian splatting technique, which, due to its limited ability to capture complex deformations and motion dynamics, can result in pixelated and poorly rendered output. In contrast, 3D Box 312 represents the same scene rendered using the reconstruction pipeline described above, incorporating dynamic flow rematching techniques that can optimize the velocity fields and Gaussian parameters over time. EXEMPLARY LANGUAGE MODELS

[0095] In at least some implementations, language models, such as large language models (LLMs), vision language models (VLMs), multimodal language models (MMLMs), and / or other types of generative artificial intelligence (AI), can be implemented. Generally, the language models can be used to implement and train the image estimator 108, motion modeler 112, loss modeler 116, and / or reconstructor 120 models described above. That is, the reconstruction pipeline can include one or more models to generate, modify, and interpret scene representations based on the estimated parameters and criteria for motion. These models may be capable of interpreting text (e.g., natural language text, code, etc.), images, videos, computer-aided design (CAD) assets, OMNIVERSE and / or METAVERSE file information (e.g.,in USD format (such as OpenUSD) and / or the like, based on the context provided in input prompts or queries. These language models can be considered "large" in implementations, based on the fact that the models are trained / updated on massive datasets and have architectures with a large number of learning network parameters (weights and biases), such as millions or billions of parameters. The LLMs / VLMs / MMLMs / etc. can be implemented to summarize text data, analyze data and derive insights from it (e.g., text, image, video, etc.) and generate new text / image / video / etc. in user-specified styles, tones, and / or formats. The LLMs / VLMs / MMLMs / etc.The present disclosure may be used in implementations solely for text processing, while in some implementations multimodal LLMs may be implemented to accept, understand, and / or generate text and / or other types of content such as images, audio, 2D and / or 3D data (e.g., in USD formats), and / or video. For example, vision language models (VLMs) or, more generally, multimodal language models (MMLMs) may be implemented to accept input data types such as image, video, audio, text, 3D design (e.g., CAD), and / or others, and / or to generate or output data types such as image, video, audio, text, 3D design, and / or others.

[0096] Different types of architectures for LLMs / VLMs / MMLMs / etc. can be implemented in various implementations. For example, different architectures can be implemented that use different techniques for understanding and generating outputs, such as text, audio, video, image, 2D and / or 3D design or asset data, etc. In some implementations, architectures for LLMs / VLMs / MMLMs / etc., such as recurrent neural networks (RNNs) or long-short-term memory (LSTM) networks, may be used, while in other implementations, transformer architectures, such as those based on self-attention and / or cross-attention mechanisms (e.g., between context data and text data), are used to understand and recognize relationships between words or tokens and / or context data (e.g., other text, video, image, design data, USD, etc.).One or more generative processing pipelines containing LLMs / VLMs / MMLMs / etc. may also contain one or more diffusion blocks (e.g., denoisers). The LLMs / VLMs / MMLMs / etc. of this disclosure may contain one or more encoder and / or decoder blocks. For example, discriminative or encoder-only models such as BERT (Bidirectional Encoder Representations from Transformers) may be implemented for tasks involving language understanding, such as classification, sentiment analysis, question answering, and named entity recognition. As another example, generative or decoder-only models such as GPT (Generative Pretrained Transformer) may be used for tasks involving language and content generation, such as... B. Text completion, story generation, and dialogue generation. LLMs / VLMs / MMLMs / etc.Architectures containing both encoder and decoder components, such as the T5 (Text-to-Text Transformer), can be implemented to understand and generate content, for example, for translation and summarization. These examples are not intended as a limitation, and any architecture type, including but not limited to those described herein, can be implemented depending on the specific implementation and the one or more tasks performed using the LLMs / VLMs / MMLMs / etc.

[0097] In various implementations, the LLMSs / VLMs / MMLMs / etc. can be trained / updated using unsupervised learning, where an LLM / VLM / MMLM / etc. learns patterns from large amounts of unlabeled text / audio / video / image / design / USD / etc. data. Due to the extensive training, the models in some implementations may not require task-specific or domain-specific training. LLMs / VLMs / MMLMs / etc. that have undergone extensive pre-training with enormous amounts of unlabeled data can be considered foundational models and may be suitable for a variety of tasks, such as answering questions, summarizing, filling in missing information, translating, and generating image / video / design / USD / data. Some LLMs / VLMs / MMLMs / etc.They can be tailored for a specific use case using techniques such as prompt tuning, fine-tuning, retrieval-enhanced generation (RAG), adding adapters (e.g., custom-tailored neural networks and / or layers of neural networks that tune or adapt prompts or tokens to align the language model with a particular task or domain), and / or using other fine-tuning or tailoring techniques that optimize the models for use in specific tasks and / or domains.

[0098] In some implementations, the LLMs / VLMs / MMLMs / etc. of this disclosure can be implemented using various model alignment techniques. For example, guardrails can be implemented in some implementations to identify impermissible or unwanted inputs (e.g., prompts) and / or outputs of the models. The system can then use the guardrails and / or other model alignment techniques to either prevent a specific unwanted input from being processed using the LLMs / VLMs / MMLMs / etc., and / or to prevent the output or presentation (e.g., display, audio output, etc.) of information generated using the LLMs / VLMs / MMLMs / etc. In some implementations, one or more additional models—or layers thereof—can be implemented to identify problems with the inputs and / or outputs of the models.For example, these "security models" can be trained / updated to identify inputs and / or outputs that are "safe" or otherwise acceptable or desirable, and / or that are "unsafe" or otherwise undesirable for the particular application / implementation. As a result, the LLMs / VLMs / MMLMs / etc. of this disclosure are less likely to output speech / text / audio / video / design data / USD data / etc. that is offensive, vulgar, inappropriate, unsafe, non-technical, and / or otherwise undesirable for the particular application / implementation.

[0099] In some implementations, the LLMs / VLMs / etc. may be configured or able to access or use one or more plugins, application programming interfaces (APIs), databases, data stores, repositories, etc. For example, for certain tasks or operations for which the model is not ideally suited, instructions (e.g., as a result of training and / or based on instructions in a given prompt) may be provided to access one or more plugins (e.g., third-party plugins) to assist in processing the current input. In such an example, where at least part of a prompt is related to restaurants or the weather, the model may access one or more restaurant or weather plugins (e.g., via one or more APIs) to retrieve the relevant information.As another example where at least part of a response requires a mathematical calculation, the model can access one or more mathematical plugins or APIs to assist in solving the one or more problems and then use the plugin's and / or API's response in the model's output. This process can be repeated, for example recursively, for any number of iterations and using any number of plugins and / or APIs until a response to the input prompt can be generated that addresses each question / request / requirement / process / operation, etc. Therefore, the one or more models can rely not only on their own knowledge gained from training on one or more large datasets but also on the expertise or optimized nature of one or more external resources, such as APIs, plugins, and / or the like.

[0100] In some implementations, multiple language models (e.g., LLMs / VLMs / MMLMs / etc.), multiple instances of the same language model, and / or multiple prompts served to the same language model or instance of the same language model can be implemented, executed, or accessed (e.g., using one or more plugins, user interfaces, APIs, databases, data stores, repositories, etc.) to provide output in response to the same query or in response to separate parts of a query. In at least one implementation, multiple language models, such as language models with different architectures or language models trained / updated on different (e.g., updated) datasets, can be served with the same input query and the same prompt (e.g., a set of constraints, conditioners, etc.).In one or more implementations, the language models can be different versions of the same base model. In one or more implementations, at least one language model can be instantiated as multiple agents; for example, more than one prompt can be provided to constrain, direct, or otherwise influence the style, content, or character of the output provided. In one or more exemplary, non-constraintive implementations, the same language model can be prompted to provide output corresponding to a different role, perspective, character, or with a different knowledge base, as defined by a provided prompt.

[0101] In each of these implementations, the output of two or more (e.g., each) language models, two or more versions of at least one language model, two or more instantiated agents of at least one language model, and / or two additional prompts provided to at least one language model can be further processed, e.g., aggregated, compared, or filtered, or used to determine (and provide) a consensus response. In one or more implementations, the output of a language model—or a version, instance, or agent—can be provided as input to another language model for further processing and / or validation. In one or more implementations, a language model can be instructed to generate or otherwise obtain output with respect to input source material, with the output being associated with the input source material.Such an assignment might involve, for example, generating a label or a portion of text that is embedded (e.g., as metadata) in input source text or image. In one or more implementations, an output from a language model can be used to determine the validity of input source material for further processing or insertion into a dataset. For example, a language model can be used to evaluate the presence (or absence) of a target word in a portion of text or an object in an image, annotating the text or image to indicate such presence (or absence). Alternatively, the determination from the language model can be used to determine whether the source material should be included in a maintained dataset, for example, with or without restrictions.

[0102] Fig. Figure 4A is a block diagram of an exemplary generative language modeling system 400 suitable for use in implementing at least some implementations of the present disclosure. In general, the exemplary generative language modeling system 400 may include components such as the image estimator 108, the motion modeler 112, the loss modeler 116, and / or the reconstructor 120, which can implement and train models of a dynamic scene reconstruction pipeline. That is, the system 400 can be used to deploy and manage the training of generative models by these components. In the Fig. 4A illustrated example contains the generative language model system 400 a Retrieval Augmented Generation (RAG) component 492, an input processor 405, a tokenizer 410, an embedding component 420, plug-ins / APIs 495 and a generative language model (LM) 430 (which may contain an LLM, a VLM, a multimodal LM, etc.).

[0103] Generally speaking, the 405 input processor can receive a 401 input containing text and / or other types of input data (e.g., audio, video, image, sensor data (e.g., LiDAR, RADAR, ultrasound, etc.), 3D design data, CAD data, Universal Scene Descriptor (USD) data, OpenUSD, etc.), depending on the architecture of the generative LM 430 (e.g., LLM / VLM / MMLM / etc.). In some implementations, the 401 input contains plain text in the form of one or more sentences, paragraphs, and / or documents. Additionally or alternatively, the 401 input can contain numeric sequences, pre-computed embeddings (e.g., word or sentence embeddings), and / or structured data (e.g., in tabular formats, JSON, or XML).In some implementations where the generative LM 430 is capable of processing multimodal inputs, the 401 input can combine text (or omit text) with image data, audio data, video data, design data, USD data, and / or other types of input data, such as, but not limited to, those described herein. Using raw input text as an example, the 405 input processor can prepare raw input text in various ways. For instance, the 405 input processor can perform various types of text filtering to remove noise (such as special characters, punctuation, HTML markup, stop words, portions of one or more images, portions of audio, etc.) from relevant text content.In an example involving stop words (frequent words that tend to convey little semantic meaning), the 405 input processor can remove stop words to reduce noise and allow the generative LM 430 to focus on more meaningful content. The 405 input processor can also apply text normalization, such as converting all characters to lowercase, removing accents, and / or handling special cases like contractions or abbreviations to ensure consistency. These are just a few examples, and other types of input processing can be applied as well.

[0104] In some implementations, a RAG component 492 (which may contain one or more RAG models and / or be performed using the generative LM 430 itself) can be used to retrieve additional information to be used as part of the input 401 or the prompt. The RAG can be used to enhance the input to the LLM / VLM / MMLM / etc. with external knowledge, making the answers to specific questions, queries, or requirements more relevant, such as in a case where specialized knowledge is needed. The RAG component 492 can retrieve this additional information (e.g., grounding information such as grounding text / image / video / audio / USD / CAD / etc.) from one or more external sources, which can then be fed into the LLM / VLM / MMLM / etc. along with the prompt to improve the accuracy of the model's responses or outputs.

[0105] For example, in some implementations, input 401 can be generated using the query or input to the model (e.g., a question, a request, etc.) in addition to the data retrieved using RAG component 492. In some implementations, input processor 405 can parse input 401 and communicate with RAG component 492 (or RAG component 492 can be part of input processor 405 in some implementations) to identify relevant text and / or other data to provide to generative LM 430 as additional context or information from which to identify the reaction, response, or output 490 in general.For example, if the input indicates that the user is interested in a desired tire pressure for a specific make and model of vehicle, the RAG component 492, using a RAG model that performs a vector search in an embedding space, can retrieve, for example, the tire pressure information or the relevant text from a digital (embedded) version of the owner's manual for that specific vehicle make and model. Similarly, if a user revisits a chatbot in connection with a specific product offering or service, the RAG component 492 can retrieve a previously stored conversation history—or at least a summary of it—and include the previous conversation history, along with the current question / request, as part of the 401 input in the generative LM 430.

[0106] The RAG component 492 can employ various RAG techniques. For example, naive RAG can be used when documents are indexed, split into pieces, and applied to an embedding model to generate embeddings that correspond to the pieces. A user request can also be applied to the embedding model and / or another embedding model of the RAG component 492, and the piece embeddings can be compared with the request's embeddings to identify the most similar embeddings that can be fed to the generative LM 430 to generate output.

[0107] In some implementations, more advanced RAG techniques can be used. For example, the pieces can undergo pre-fetching processes (e.g., redirection, rewriting, metadata analysis, extension, etc.) before being passed to the embedding model. Furthermore, post-fetching processes (e.g., re-ranking, prompt compression, etc.) can be performed on the outputs of the embedding model before the final embeddings are generated and used as a comparison to an input query.

[0108] As another example, modular RAG techniques can be used, such as those that resemble naive and / or extended RAG, but may also include features such as hybrid search, recursive retrieval and query engines, step-back approaches, subqueries, and hypothetical document embedding.

[0109] As another example, Graph-RAG can use knowledge graphs as a source of contextual or factual information. Graph-RAG can be implemented using a graph database as a source of contextual information, which is then sent to the LLM / VLM / MMLM / etc. Instead of providing the model with snippets of data extracted from (or in addition to) larger documents, which can result in a lack of context, factual accuracy, linguistic precision, etc., Graph-RAG can also provide structured entity information to the LLM / VLM / MMLM / etc. by combining the structured text description of the entity with its many properties and relationships, allowing the model deeper insights. When implementing Graph-RAG, the systems and procedures described herein use a graph as a content store, extract relevant snippets from documents, and request them from the LLM / VLM / MMLM / etc.to respond using this. In such implementations, the knowledge graph can contain relevant text content and metadata about the knowledge graph and be integrated into a vector database. In some implementations, the graph RAG can use a graph as a domain expert, extracting descriptions of concepts and entities relevant to a query / prompt and passing them to the model as semantic context. These descriptions can include relationships between the concepts. For example, the graph can be used as a database, where part of a query / prompt can be allocated to a graph query, the graph query can be executed, and the LLM / VLM / MMLM / etc. can summarize the results.In such an example, the graph can store relevant factual information, and a query (natural language (NL) query) to a graph query tool (NL-to-graph query tool) and an entity join can be used. In some implementations, graph RAG (e.g., using a graph database) can be combined with standard RAG (e.g., vector database) and / or other RAG types to benefit from multiple approaches.

[0110] In all implementations, the RAG component 492 can implement a plug-in, an API, a user interface, and / or other functionality to perform RAG. For example, a graph RAG plug-in can be used by the LLM / VLM / MMLM / etc. to query the knowledge graph to extract relevant information for feeding into the model, and a standard or vector RAG plug-in can be used to query a vector database. For example, the graph database can interact with a plug-in's REST interface, thus decoupling the graph database from the vector database and / or the embedding models.

[0111] The Tokenizer 410 can segment (e.g., processed) text data into smaller units (tokens) for subsequent analysis and processing. Depending on the implementation, the tokens can represent individual words, partial words, characters, parts of audio / video images, etc. Word-based tokenization divides the text into individual words, with each word treated as a separate token. Partial word tokenization breaks words down into smaller meaning-bearing units (e.g., prefixes, suffixes, stems), enabling the generative LM 430 to understand morphological variations and more effectively handle words not contained in the vocabulary. Character-based tokenization represents each character as a separate token, allowing the generative LM 430 to process text at a fine-grained level.The choice of tokenization strategy can depend on factors such as the language being processed, the task at hand, and / or the characteristics of the training dataset. Therefore, Tokenizer 410 can convert the (e.g., processed) text into a structured format according to the tokenization scheme implemented in the specific implementation.

[0112] The embedding component 420 can use any known embedding technique to convert discrete tokens into (e.g., dense, continuous vector) representations of semantic meaning. For example, the embedding component 420 can use pretrained / updated word embeddings (e.g., Word2Vec, GloVe, or FastText), one-hot coding, term frequency-inverse document frequency (TF-IDF) coding, one or more neural network embedding layers, and / or other techniques.

[0113] In some implementations where the input 401 contains image data / video data / etc., the input processor 401 can resize the data to a standard size compatible with the format of a corresponding input channel and / or normalize pixel values ​​to a common range (e.g., 0 to 1) to ensure uniform representation. The embedding component 420 can encode the image data using any known technique (e.g., using one or more convolutional neural networks (CNNs) to extract visual features). In some implementations where the input 401 contains audio data, the input processor 401 can resample an audio file to a uniform sampling rate for consistent processing, and the embedding component 420 can use any known technique to extract and encode audio features, such as in the form of a spectrogram (e.g., a spectrogram).(of a Mel spectrogram). In some implementations where the input contains video data, the input processor can extract frames or apply resizing to extracted frames, and the embedding component can extract features such as optical flow or video embeddings and / or encode temporal information or sequences of frames. In some implementations where the input contains multimodal data, the embedding component can fuse representations of the different data types (e.g., text, image, audio, USD, video, design, etc.) using techniques such as early fusion (concatenation), late fusion (sequential processing), attention-based fusion (e.g., self-attention, cross-attention), etc.

[0114] The generative LM 430 and / or other components of the generative LM System 400 can use different types of neural network architectures, depending on the implementation. For example, transformer-based architectures, such as those used in models like GPT, can be implemented, incorporating self-attention mechanisms that weigh the importance of different words or tokens in the input sequence, and / or feedforward networks that process the output of the self-attention layers, applying nonlinear transformations to the input representations and extracting higher-level features. Some non-restrictive example architectures include transformers (e.g.,Encoder-decoder, decoder-only, multimodal), RNNs, LSTMs, fusion models, diffusion models, crossmodal embedding models that learn shared embedding spaces, graph neural networks (GNNs), hybrid architectures that combine different types of architectures, adversarial networks such as generative adversarial networks (GANs) or adversarial autoencoders (AAEs) for joint distributional learning, and others. Therefore, depending on the implementation and architecture, the embedding component 420 can apply a coded representation of the input 401 to the generative LM 430, and the generative LM 430 can process the coded representation of the input 401 to generate an output 490 that may contain response text and / or other types of data.

[0115] As described herein, in some implementations, the generative LM 430 may be configured to access or use plug-ins / APIs 495—or have the capability to access or use them—(which may include one or more plug-ins, application programming interfaces (APIs), databases, data stores, repositories, etc.). For example, for certain tasks or operations for which the generative LM 430 is not ideally suited, the model may have instructions (e.g., as a result of training and / or based on instructions in a given prompt, such as those retrieved using the RAG component 492) to access one or more plug-ins / APIs 495 (e.g., third-party plug-ins) to obtain assistance in processing the current input.In such an example, where at least part of a prompt is related to restaurants or weather, the model can access one or more restaurant or weather plugins (e.g., via one or more APIs), send at least part of the prompt related to the specific plugin / API 495 to the plugin / API 495, the plugin / API 495 can process the information and return a response to the generative LM 430, and the generative LM 430 can use the response to generate the output 490. This process can be repeated, e.g., recursively, for any number of iterations and using any number of plugins / APIs 495 until an output 490 can be generated that addresses every question / request / request / process / operation / etc. from the input 401.Therefore, one or more models can rely not only on their own knowledge from training with one or more large datasets and / or from data retrieved using the RAG component 492, but also on the expertise or optimized nature of one or more external resources, such as the plug-ins / APIs 495.

[0116] Fig. Figure 4B is a block diagram of an exemplary implementation in which the generative LM 430 incorporates a transformer encoder decoder. In general, the generative LM 430 can provide the implementation and training of models for the image estimator 108, the motion modeler 112, the loss modeler 116, and / or the reconstructor 120. That is, the generative LM 430 can be used to develop and train models that can execute the aforementioned components to process input data, estimate motion, and perform dynamic scene reconstructions based on the generated outputs. For example, suppose an input text, such as "Who discovered gravity," is tokenized (e.g., by the Tokenizer 410). Fig. 4A) into tokens, such as words, and each token is encoded into a corresponding embedding (e.g., of size 512) (e.g., by the embedding component 420 of Fig. 4A). Since these token embeddings typically do not represent the token's position in the input sequence, any known technique can be used to add positional encoding to each token embedding to encode the sequential relationships and context of the tokens in the input sequence. Therefore, the (e.g., resulting) embeddings can be applied to one or more encoders 435 of the generative LM 430.

[0117] In an exemplary implementation, the one or more Encoders 435 form an encoder stack, with each encoder containing a self-attention layer and a feedforward network. In an exemplary transformer architecture, each token (e.g., a word) flows through a separate path. Therefore, each encoder can accept a sequence of vectors, with each vector passing through the self-attention layer, then through the feedforward network, and then up to the next encoder in the stack. Any known self-attention technique can be used.For example, to calculate a self-attention score for each token (word), a query vector, a key vector, and a value vector can be created for each token. A self-attention score can be calculated for pairs of tokens by taking the dot product of the query vector with the corresponding key vectors, normalizing the resulting numerical values, multiplying them by the corresponding value vectors, and summing the weighted value vectors. The encoder can employ multi-head attention, where the attention mechanism is applied multiple times in parallel with different learned weight matrices. Any number of encoders can be cascaded to generate a context vector that encodes the input. An attention projection layer 440 can convert the context vector into attention vectors (keys and values) for one or more decoders 445.

[0118] In an exemplary implementation, the one or more decoders 445 form a decoder stack, with each decoder containing a self-attention layer, an encoder-decoder self-attention layer that uses the encoder's attention vectors (keys and values) to focus on relevant parts of the input sequence, and a feedforward network. As with the one or more encoders 435, in an exemplary transformer architecture, each token (e.g., word) flows through a separate path in the one or more decoders 445. During a first pass, the one or more decoders 445, a classifier 450, and a generation mechanism 455 can generate an initial token, and the generation mechanism 455 can apply the generated token as input during a second pass. The process can repeat in a loop, successively passing tokens (e.g.,words) are generated and added to the output of the previous pass, and the token embeddings of the compound sequence with positional encodings are applied as an input to one or more Decoder 445s during a subsequent pass, generating one token at a time (known as autoregression) until a symbol or token is predicted that represents the end of the response. Within each decoder, the self-attention layer is typically restricted to focusing only on previous positions in the output sequence by applying a masking technique (e.g., setting future positions to negative infinity) before the softmax operation. In an exemplary implementation, the encoder-decoder attention layer works similarly to the (e.g.,Multi-head) self-attention in the one or more encoders 435, except that it creates its queries from the layer below and takes the keys and values ​​(e.g. matrix) from the output of the one or more encoders 435.

[0119] Therefore, the one or more decoders 445 can output a decoded (e.g., vector) representation of the input applied during a given iteration. The classifier 450 can include a multi-class classifier containing one or more layers of a neural network that project the decoded (e.g., vector) representation into an appropriate dimensionality (e.g., one dimension for each supported word or token in the output vocabulary) and a softmax operation that converts logits into probabilities. Therefore, the generation mechanism 455 can select or sample a word or token based on an appropriate predicted probability (e.g., selecting the word with the highest predicted probability) and append it to the output of a previous iteration, generating each word or token sequentially.The generation mechanism 455 can repeat the process, triggering successive decoder inputs and corresponding predictions until a symbol or token is selected or sampled that represents the end of the response, after which the generation mechanism 455 can output the generated response.

[0120] Fig. 4C is a block diagram of an exemplary implementation where the generative LM 430 incorporates a decoder-transformer-only architecture. For example, one or more 460 decoders can be used. Fig. 4C similar to one or more 445 decoders from Fig. 4B work, except that each of the one or more decoders 460 of Fig. 4C omits the encoder-decoder self-attention layer (since there is no encoder in this implementation). Therefore, the one or more Decoders 460 can form a decoder stack, with each decoder containing a self-attention layer and a feedforward network. Furthermore, instead of encoding the input sequence, a symbol or token representing the end of the input sequence (or the beginning of the output sequence) can be appended to the input sequence, and the resulting sequence (e.g., corresponding embeddings with positional encodings) can be applied to the one or more Decoders 460. As with the one or more Decoders 445 of Fig. 4B allows each token (e.g., word) to flow through a separate path in the one or more decoders 460, and the one or more decoders 460, a classifier 465, and a generation mechanism 470 can use autoregression to sequentially generate one token after another until a symbol or token is predicted that represents the end of the response. The classifier 465 and the generation mechanism 470 can be used similarly to the classifier 450 and the generation mechanism 455 of Fig. 4B operates wherein the generation mechanism 470 selects or samples each successive output token based on a corresponding predicted probability and appends it to the output of a previous pass, each token being generated sequentially until a symbol or token is selected or sampled that represents the end of the response. This and other architectures described herein are intended only as examples, and other suitable architectures may be implemented within the scope of protection of this disclosure. EXAMPLE CALCULATION DEVICE

[0121] Fig. Figure 5 is a block diagram of an exemplary computing device 500 suitable for use in implementing some implementations of the present disclosure. In general, the one or more exemplary computing devices 500 can perform operations of the system 100, such as managing data processing and performing calculations for dynamic scene reconstruction. That is, the one or more computing devices 500 can execute instructions to process multi-view images, determine criteria for motion, and update the image rendering model based on input data and constraints.The computing device 500 can include an interlink system 502 that directly or indirectly couples the following devices: memory 504, one or more central processing units (CPUs) 506, one or more graphics processing units (GPUs) 508, a communication interface 510, input / output (I / O) ports 512, input / output components 514, a power supply 516, one or more presentation components 518 (e.g., display(s)), and one or more logic units 520. In at least one implementation, the one or more computing devices 500 can comprise one or more virtual machines (VMs), and / or each of the components thereof can comprise virtual components (e.g., virtual hardware components).As non-restrictive examples, one or more of the GPUs 508 can comprise one or more vGPUs, one or more of the CPUs 506 can comprise one or more vCPUs, and / or one or more of the logic units 520 can comprise one or more virtual logic units. Thus, one or more computing devices 500 can contain discrete components (e.g., a complete GPU allocated to computing device 500), virtual components (e.g., a portion of a GPU allocated to computing device 500), or a combination thereof.

[0122] Although the various blocks of Fig. Where components 5 are shown as being connected via the intermediate interconnect system 502, this is not intended as a limitation and is only for clarity. In some implementations, for example, a presentation component 518, such as a display device, may be considered an I / O component 514 (e.g., if the display is a touchscreen). As another example, the CPUs 506 and / or GPUs 508 may contain memory (e.g., the memory 504 may represent a memory device in addition to the memory of the GPUs 508, the CPUs 506, and / or other components). In itself, the computing device of Fig. 5 is for illustrative purposes only. No distinction is made between categories such as "workstation", "server", "laptop", "desktop", "tablet", "client device", "mobile device", "handheld device", "game console", "electronic control unit (ECU)", "virtual reality system" and / or other device or system types, as all are within the scope of protection of the computing device of Fig. 5 are being considered.

[0123] The 502 interconnect system can represent one or more connections or buses, such as an address bus, a data bus, a control bus, or a combination thereof. The 502 interconnect system can include one or more bus or connection types, such as an Industry Standard Architecture (ISA) bus, an Extended ISA bus, a Video Electronics Standards Association (VESA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI Express (PCIe) bus, and / or another type of bus or connection. In some implementations, there are direct connections between components. For example, the CPU (506) can be directly connected to the memory (504). Furthermore, the CPU (506) can be directly connected to the GPU (508).In a direct or point-to-point connection between components, the 502 interconnection system can include a PCIe link to establish the connection. In these examples, a PCI bus does not need to be included in the 500 computing device.

[0124] The 504 main memory can contain a variety of computer-readable media. Computer-readable media can be any available media that the 500 computer can access. Computer-readable media can include both volatile and non-volatile media, as well as removable and non-removable media. For example, and without limitation, computer-readable media can include computer storage media and communication media.

[0125] Computer storage media can include both volatile and non-volatile media, and / or removable and non-removable media, implemented using any method or technology for storing information, such as computer-readable instructions, data structures, program modules, and / or other data types. For example, main memory can store 504 computer-readable instructions (e.g., representing one or more programs and / or one or more program elements, such as an operating system).Computer storage media may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other storage technologies, CD-ROM, Digital Versatile Discs (DVDs) or other optical disk storage, magnetic cartridges, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and that the Computing Device 500 can access. As used herein, computer storage media do not inherently contain signals.

[0126] Computer storage media can embody computer-readable instructions, data structures, program modules, and / or other data types in a modulated data signal, such as a carrier wave or other transport mechanism, and include any media for transmitting information. The term "modulated data signal" can refer to a signal in which one or more of its properties are set or modified to encode information within the signal. Computer storage media can include, but are not limited to, wired media, such as a wired network or a direct-wired connection, and wireless media, such as acoustic, RF, infrared, and other wireless media. Combinations of the above should also be included in the scope of protection of the computer-readable media.

[0127] The one or more CPUs 506 can be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 500 to perform one or more of the procedures and / or processes described herein. The one or more CPUs 506 can each contain one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of processing a multitude of software threads simultaneously. The CPUs 506 can contain any type of processor and can contain different types of processors depending on the type of computing device 500 implemented (e.g., processors with fewer cores for mobile devices and processors with more cores for servers).Depending on the type of computing device 500, the processor can be, for example, an Advanced RISC Machine (ARM) processor implemented with Reduced Instruction Set Computing (RISC), or an x86 processor implemented with Complex Instruction Set Computing (CISC). The computing device 500 can contain one or more CPUs 506, in addition to one or more microprocessors or additional coprocessors, such as mathematical coprocessors.

[0128] In addition to or as an alternative to the one or more CPUs 506, the one or more GPUs 508 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 500 to perform one or more of the procedures and / or processes described herein. One or more of the GPUs 508 may be an integrated GPU (e.g., with one or more of the CPUs 506) and / or one or more of the GPUs 508 may be a discrete GPU. In implementations, one or more of the GPUs 508 may be a coprocessor of one or more of the CPUs 506. The one or more GPUs 508 may be used by the computing device 500 to render graphics (e.g., 3D graphics) or to perform general-purpose computing. For example, the one or more GPUs 508 may be used for general-purpose computing on GPUs (GPGPU).The one or more GPUs 508 can contain hundreds or thousands of cores capable of processing hundreds or thousands of software threads simultaneously. The one or more GPUs 508 can generate pixel data for output images in response to rendering commands (e.g., rendering commands from the one or more CPUs 506 received through a host interface). The one or more GPUs 508 can include graphics memory, such as display memory, for storing pixel data or other suitable data, such as GPGPU data. The display memory can be included as part of the 504 main memory. The one or more GPUs 508 can contain two or more GPUs operating in parallel (e.g., via a link). The link can connect the GPUs directly (e.g., using NVLINK) or connect them via a switch (e.g., using NVSwitch).When combined, each GPU can generate 508 pixel data or GPGPU data for different sections of an output or for different outputs (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU can have its own dedicated memory or share memory with other GPUs.

[0129] In addition to or as an alternative to the one or more CPUs 506 and / or the one or more GPUs 508, the one or more logic units 520 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 500 to perform one or more of the procedures and / or processes described herein. In implementations, the one or more CPUs 506, the one or more GPUs 508, and / or the one or more logic units 520 may discretely or collectively execute any combination of the procedures, processes, and / or sections thereof. One or more of the logic units 520 may be part of and / or integrated within one or more of the CPUs 506 and / or one or more of the GPUs 508, and / or one or more of the logic units 520 may be discrete components or otherwise separate from the CPUs 506 and / or the GPUs 508.In implementations, one or more of the 520 logic units can be a co-processor of one or more of the 506 CPUs and / or one or more of the 508 GPUs.

[0130] Examples of the one or more Logic Units 520 include one or more processor cores and / or components thereof, such as Data Processing Units (DPUs), Tensor Cores (TCs), Tensor Processing Units (TPUs), Pixel Visual Cores (PVCs), Vision Processing Units (VPUs), Graphics Processing Clusters (GPCs), Texture Processing Clusters (TPCs), Streaming Multiprocessors (SMs), Tree Traversal Units (TTUs), Artificial Intelligence Accelerators (AIAs), Deep Learning Accelerators (DLAs), Programmable Vision Accelerators (PVAs), one or more Direct Memory Access (DMA) systems, one or more vision or Vector processing units (Vision or Vector Processing Units, VPUs),one or more pixel processing engines (PPEs), e.g. B. including a 2D array of processing elements, each communicating with one or more other processing elements in the array, one or more decoupled accelerators or units (e.g., decoupled lookup table (DLUT) accelerators or units), etc., vision processing units (VPUs), optical flow accelerators (OFAs), field programmable gate arrays (FPGAs), neuromorphic chips, quantum processing units (QPUs), associative processing units (APUs), arithmetic logic units (ALUs), application-specific integrated circuits (ASICs), floating-point units (FPUs), input / output (I / O) elements,Peripheral Component Connection (PCI) or Peripheral Component Connection Express (PCIe) elements and / or the like.

[0131] The Communications Interface 510 can include one or more receivers, transmitters, and / or transceivers that enable the Computing Device 500 to communicate with other computers over an electronic network, including wired and / or wireless communication. The Communications Interface 510 can include components and functions that enable communication over a variety of different networks, such as wireless networks (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, etc.), wired networks (e.g., Ethernet or InfiniBand communication), low-power wide-area networks (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.In one or more implementations, the one or more logic units 520 and / or the communication interface 510 can contain one or more data processing units (DPUs) to transfer data received via a network and / or the interconnection system 502 directly to one or more GPUs 508 (e.g., a memory thereof).

[0132] The I / O ports 512 enable the computing device 500 to be logically coupled with other devices, including the I / O components 514, one or more presentation components 518, and / or other components, some of which may be built into (e.g., integrated with) the computing device 500. Illustrative I / O components 514 include a microphone, mouse, keyboard, joystick, gamepad, game controller, satellite dish, scanner, printer, wireless device, etc. The I / O components 514 can provide a natural user interface (NUI) that processes user-generated air gestures, speech, or other physiological inputs. In some cases, the inputs can be transmitted to a suitable network element for further processing.A NUI can implement any combination of speech capture, stylus capture, face capture, biometric capture, gesture capture (both on-screen and off-screen), air gestures, head and eye tracking, and touch capture (as further described below) associated with a display on the Computing Device 500. The Computing Device 500 can include depth cameras, such as stereoscopic camera systems, infrared camera systems, RGB camera systems, touchscreen technology, and combinations thereof, for gesture capture and recognition. Additionally, the Computing Device 500 can include accelerometers or gyroscopes (e.g., as part of an inertial measurement unit (IMU)) that enable motion detection. The output from the accelerometers or gyroscopes can be used, for example, by the Computing Device 500 to render immersive augmented reality or virtual reality.

[0133] The power supply 516 can include a hardwired power supply, a battery power supply, or a combination thereof. The power supply 516 can power the computing device 500 to enable the operation of the computing device 500's components.

[0134] The one or more presentation components 518 can include a display (e.g., a monitor, a touchscreen, a television screen, a heads-up display (HUD), other display types, or a combination thereof), speakers, and / or other presentation components. The one or more presentation components 518 can receive data from other components (e.g., the one or more GPUs 508, the one or more CPUs 506, DPUs, etc.) and output the data (e.g., as an image, video, sound, etc.). EXEMPLARY DATA CENTER

[0135] Fig. Figure 6 illustrates an exemplary data center 600 that can be used in at least one implementation of the present disclosure. In general, the exemplary data center 600 can include hardware and software resources for processing, storing, and managing data for dynamic scene reconstruction. That is, the data center 600 can perform large-scale computations, manage data storage, and provide network access to support the training and inference operations of the image estimator 108, the motion modeler 112, the loss modeler 116, and the reconstructor 120. The data center 600 can include a data center infrastructure layer 610, a framework layer 620, a software layer 630, and / or an application layer 640.

[0136] As in Fig. As shown in Figure 6, the infrastructure layer 610 of the data center can contain a resource orchestrator 612, clustered compute resources 614 and node compute resources (“node RRs”) 616(1)-616(N), where “N” is any positive integer. In at least one implementation, the Node RRs 616(1)-616(N) may include, but are not limited to, any number of central processing units (CPUs) or other processors (including DPUs, accelerators, field-programmable gate arrays (FPGAs), graphics processing units (GPUs), etc.), memory devices (e.g., dynamic solid-state memory), storage devices (e.g., solid-state or disk drives), network input / output (NW I / O) devices, network switches, virtual machines (VMs), power supply modules, and / or cooling modules, etc. In some implementations, one or more Node RRs may be included.The node RRs 616(1)-616(N) correspond to a server that has one or more of the aforementioned computing resources. Furthermore, in some implementations, node RRs 616(1)-616(N) may include one or more virtual components, such as vGPUs, vCPUs, and / or the like, and / or one or more of the node RRs 616(1)-616(N) may correspond to a virtual machine (VM).

[0137] In at least one implementation, the grouped compute resources 614 can contain separate groupings of Node RRs 616, which are housed in one or more racks (not shown) or in many racks in data centers at different geographic locations (also not shown). Separate groupings of Node RRs 616 within grouped compute resources 614 can contain grouped compute, network, memory, or storage resources that can be configured or allocated to support one or more workloads. In at least one implementation, multiple Node RRs 616, including CPUs, GPUs, DPUs, and / or other processors, can be grouped in one or more racks to provide compute resources to support one or more workloads.The one or more racks can also contain any number of power supply modules, cooling modules and / or network switches in any combination.

[0138] The Resource Orchestrator 612 can configure or otherwise control one or more Node RRs 616(1)-616(N) and / or grouped compute resources 614. In at least one implementation, the Resource Orchestrator 612 can include a Software Design Infrastructure (SDI) management entity for the Data Center 600. The Resource Orchestrator 612 can include hardware, software, or a combination thereof.

[0139] In at least one implementation, as in Fig. As shown in Figure 6, the framework layer 620 can contain a job scheduler 628, a configuration manager 634, a resource manager 636, and / or a distributed file system 638. The framework layer 620 can contain a framework that supports the software 632 of the software layer 630 and / or one or more applications 642 of the application layer 640. The software 632 or the one or more applications 642 can each contain web-based service software or applications, such as those provided by Amazon Web Services, Google Cloud, and Microsoft Azure. The framework layer 620 can be a type of free and open-source software web application framework, such as Apache Spark™ (hereinafter "Spark"), which can use a distributed file system 638 for processing large amounts of data (e.g., "Big Data"), without being limited to it.In at least one implementation, the job scheduler 628 can include a Spark driver to facilitate the scheduling of workloads supported by different layers of the data center 600. The configuration manager 634 can be able to configure different layers, such as the software layer 630 and the framework layer 620, which includes Spark and the distributed file system 638, to support the processing of large amounts of data. The resource manager 636 can be able to manage clustered or grouped compute resources allocated or assigned to support the distributed file system 638 and the job scheduler 628. In at least one implementation, the clustered or grouped compute resources can include the grouped compute resources 614 on the infrastructure layer 610 of the data center.The Resource Manager 636 can coordinate with the Resource Orchestrator 612 to manage these allocated or assigned computing resources.

[0140] In at least one implementation, the software contained in software layer 630 may include software 632 that is used by at least sections of the node RRs 616(1)-616(N), the grouped compute resources 614, and / or the distributed file system 638 of framework layer 620. One or more types of software may include, among others, web page search software, email virus scanning software, database software, and streaming video content software.

[0141] In at least one implementation, the applications contained in application layer 640 can include one or more types of applications used by at least sections of the node RRs 616(1)-616(N), the clustered compute resources 614, and / or the distributed file system 638 of framework layer 620. One or more types of applications can include, but are not limited to, any number of genome applications, cognitive computations, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), and / or other machine learning applications used in conjunction with one or more implementations.

[0142] In at least one implementation, a configuration manager 634, resource manager 636, and resource orchestrator 612 can implement any number and type of self-modifying actions based on any amount and type of data collected in any technically feasible way. Self-modifying actions can relieve a data center operator of the data center 600 from potentially making poor configuration decisions and potentially avoiding underutilized and / or poorly functioning sections of a data center.

[0143] The Data Center 600 may contain tools, services, software, or other resources to train one or more machine learning models or to predict or infer information using one or more machine learning models according to one or more implementations described herein. For example, one or more machine learning models may be trained / updated by calculating weighting parameters according to a neural network architecture, using software and / or computing resources described above with reference to Data Center 600.In at least one implementation, trained / updated or deployed machine learning models corresponding to one or more neural networks may be used to infer or predict information using the resources described above with reference to Computing Center 600 by using weighting parameters computed by one or more training techniques such as, but not limited to, those described herein.

[0144] In at least one implementation, the data center can utilize 600 CPUs, application-specific integrated circuits (ASICs), GPUs, FPGAs, and / or other hardware (or equivalent virtual computing resources) to perform training and / or inference using the resources described above. Additionally, one or more of the software and / or hardware resources described above can be configured as a service to allow users to train or infer information, such as image capture, speech capture, or other artificial intelligence services. EXEMPLARY NETWORK ENVIRONMENTS

[0145] Network environments suitable for use in implementing the disclosure may include one or more client devices, servers, network-attached storage (NAS), other backend devices, and / or other device types. The client devices, servers, and / or other device types (e.g., each device) may run on one or more instances of the one or more computing devices. Fig. 5. For example, each device may contain similar components, features, and / or functionality to one or more computing devices 500. If backend devices (e.g., servers, NAS, etc.) are implemented, the backend devices may also be included as part of a data center 600, an example of which is given herein with reference to Fig. 6 is described in more detail.

[0146] The components of a network environment can communicate with each other over one or more networks, which can be wired, wireless, or both. The network can contain multiple networks or a network of networks. For example, the network can contain one or more wide area networks (WANs), one or more local area networks (LANs), one or more public networks, such as the internet and / or a public switched telephone network (PSTN), and / or one or more private networks. If the network includes a wireless telecommunications network, components such as a base station, a communications tower, or even access points (as well as other components) can provide wireless connectivity.

[0147] Compatible network environments can include one or more peer-to-peer network environments—in which case a server cannot be included in a network environment—and one or more client-server network environments—in which case one or more servers can be included in a network environment. In peer-to-peer network environments, the functionality described herein can be implemented with reference to one or more servers on any number of client devices.

[0148] In at least one implementation, a network environment can include one or more cloud-based network environments, a distributed computing environment, a combination thereof, etc. A cloud-based network environment can include a framework layer, a job scheduler, a resource manager, and a distributed file system implemented on one or more servers, which may include one or more core network servers and / or edge servers. A framework layer can include a framework to support software of a software layer and / or one or more applications of an application layer. The software or the one or more applications may each include web-based service software or applications. In implementations, one or more of the client devices can use the web-based service software or applications (e.g.,by accessing the service software and / or applications via one or more application programming interfaces (APIs). The framework layer can be a type of free and open-source software web application framework that, for example, uses a distributed file system for processing large amounts of data (e.g., "Big Data"), but is not limited to this.

[0149] A cloud-based network environment can provide cloud computing and / or cloud storage, performing any combination (or parts thereof) of the computing and / or data storage functions described herein. Each of these different functions can be distributed across multiple locations of central or core servers (e.g., one or more data centers, which may be distributed across a state, region, country, the globe, etc.). If a connection to a user (e.g., a client device) is relatively close to one or more edge servers, one or more core servers can offload at least some functionality to the one or more edge servers. A cloud-based network environment can be private (e.g., restricted to a single organization), public (e.g., available to many organizations), and / or a combination thereof (e.g., a hybrid cloud environment).

[0150] The one or more client devices can incorporate at least some of the components, features, and functionality of the one or more mentioned here with respect to Fig.The exemplary computing devices described in Section 5 include 500. By way of example, and not as a limitation, a client device may be a personal computer (PC), a laptop, a mobile device, a smartphone, a tablet computer, a smartwatch, a portable computer, a personal digital assistant (PDA), an MP3 player, a virtual reality headset, a global positioning system (GPS) or global positioning device, a video player, a video camera, a surveillance device or surveillance system, a vehicle, a boat, a hydrofoil, a virtual machine, a drone, a robot, a handheld communication device, a hospital device, a gaming device or gaming system, an entertainment system, a vehicle computer system, an embedded system controller, a remote control, an apparatus, a consumer electronics device, a workstation, an edge device,any combination of these described devices or any other suitable device may be embodied.

[0151] The revelation can be described in the general context of computer code or machine-usable instructions, including computer-executable instructions such as program modules that are executed by a computer or other machine, such as a personal data assistant or other handheld device. Generally, program modules, which contain routines, programs, objects, components, data structures, etc., refer to code that performs specific tasks or implements certain abstract data types. The revelation can be practiced in a variety of system configurations, including handheld devices, consumer electronics, general-purpose computers, more specialized computing devices, etc.The revelation can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are connected to each other via a network for communication.

[0152] As used herein, any mention of "and / or" in relation to two or more elements should be interpreted as referring to only one element or a combination of elements. For example, "element A, element B and / or element C" may include only element A, only element B, only element C, element A and element B, element A and element C, element B and element C, or elements A, B, and C. Furthermore, "at least one of element A or element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. Additionally, "at least one of element A and element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B.

[0153] The subject matter of this disclosure is specifically described herein to satisfy legal requirements. However, the description itself is not intended to limit the scope of protection afforded by this disclosure. Rather, the inventors have considered that the claimed subject matter may also be embodied in a way that includes various steps or combinations of steps similar to those described herein, in conjunction with other present or future technologies. Although the terms “step” and / or “block” may be used herein to denote various elements of the methods employed, these terms should not be interpreted as implying any particular sequence among or between the various steps disclosed herein, except where the sequence of each step is expressly described.

[0154] The disclosure of this application also contains the following numbered clauses: Clause 1 One or more processors, comprising processing circuits, for: Causing an image rendering model to generate an estimated image of a scene based on at least a plurality of images of the scene, wherein at least one image of the plurality of images is associated with at least one from a different time or view; and Updating the image rendering model based on at least the estimated image, the multitude of images, and one or more criteria for motion associated with the estimated image. Clause 2 The one or more processors according to Clause 1, wherein the one or more criteria for motion comprise a velocity field representing a plurality of deformations in space over time. Clause 3 The one or more processors according to Clause 1 or 2, wherein updating the image rendering model further includes rematching a velocity field generated from the estimated image to a previous velocity field corresponding to the scene. Clause 4 The one or more processors according to any of the preceding clauses, wherein the one or more criteria for motion include at least one of the following: (i) a rigidity constraint that restricts changes in the shape of objects over time, or (ii) a continuity constraint that assumes smooth transitions in object motion within the scene. Clause 5 The one or more processors according to Clause 4, wherein the one or more criteria for motion correspond to a machine learning (ML) model that is updated to generate one or more deformations of the estimated image based on parameters derived from one or more historical scenes. Clause 6 The one or more processors according to Clause 5, wherein the one or more processors comprising processing circuits shall be used for the following: Determining one or more criteria for motion, based on at least one output of a minimization of one or more functions that satisfy at least one of the following: (i) the rigidity constraint, (ii) the continuity constraint, or (iii) a constraint derived from the ML model. Clause 7 The one or more processors according to any of the preceding clauses, wherein the image rendering model comprises at least one of the following: (i) a Gaussian splatting model or (ii) a neural radiation field (NeRF) model. Clause 8 The one or more processors according to any of the preceding clauses, wherein the plurality of images of the scene comprises a plurality of multi-view images, wherein the plurality of multi-view images corresponds to a plurality of different viewpoints captured at a plurality of different times. Clause 9 The one or more processors according to any of the preceding clauses, wherein the processing circuitry is used for the following: Applying scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on the multitude of images of the scene. Clause 10 The one or more processors according to any of the preceding clauses, wherein updating the image rendering model includes minimizing a reconstruction loss and a refitting loss, and wherein the reconstruction loss corresponds to a measure of discrepancy between the estimated image and the plurality of images of the scene, and wherein the refitting loss corresponds to a measure of deviation between the estimated image and the one or more criteria for motion associated with an image flow. Clause 11 The one or more processors according to any of the preceding clauses, wherein the one or more processors comprise at least one of the following: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for performing digital twin operations; a system for performing light transport simulations; a system for conducting collaborative content creation for 3D assets; a system for performing deep learning operations; a system for performing remote operations; a system for performing real-time streaming; a system for generating or presenting one or more augmented reality, virtual reality, or mixed reality content; a system that is implemented using an edge device; a system that is implemented using a robot; a system for performing operations using conversational AI; a system that implements one or more multimodal language models; a system that implements one or more large language models (LLMs); a system that implements one or more Vision Language Models (VLMs); a system for generating synthetic data; a system for generating synthetic data using AI; a system that includes one or more virtual machines (VMs); a system that is at least partially implemented in a data center; or a system that is implemented at least partially using cloud computing resources. Clause 12 System, comprehensive: one or more processors for performing operations, including: Causing an image rendering model to generate an estimated image of a scene based on at least a plurality of images of the scene, wherein at least one image of the plurality of images is associated with at least one from a different time or view; and Updating the image rendering model based on at least the estimated image, the multitude of images, and one or more criteria for motion associated with the estimated image. Clause 13 System according to Clause 12, wherein one or more criteria for motion comprise a velocity field representing a multitude of deformations in space over time. Clause 14 System according to Clause 12 or 13, wherein updating the image rendering model further includes rematching a velocity field generated from the estimated image to a previous velocity field corresponding to the scene. Clause 15 System according to one of Clauses 12 to 14, wherein the one or more criteria for motion include at least one of (i) a rigidity constraint that restricts changes in the shape of objects over time or (ii) a continuity constraint that assumes smooth transitions in object motion within the scene, and wherein the one or more criteria for motion correspond to a machine learning (ML) model trained to generate one or more deformations of the estimated image based on parameters derived from one or more historical scenes. Clause 16 System according to Clause 15, wherein the one or more processors are used to perform operations, comprising: Determining one or more criteria for motion, based on at least one output of a minimization of one or more functions that satisfy at least one of the following: (i) the rigidity constraint, (ii) the continuity constraint, or (iii) a constraint derived from the ML model. Clause 17 System according to one of Clauses 12 to 16, wherein the image rendering model includes at least one of (i) a Gaussian Splatting model or (ii) a neural radiation field (NeRF) model. Clause 18 System according to one of Clauses 12 to 17, wherein the multitude of images of the scene comprises a multitude of multi-view images, wherein the multitude of multi-view images corresponds to a multitude of different viewpoints captured at a multitude of different times. Clause 19 System according to any one of Clauses 12 to 18, wherein the one or more processors are used to perform operations, comprising: Applying scene reconstruction using the image rendering model to render the estimated image for one or more viewpoints and one or more time intervals based on the multitude of images of the scene. Clause 20 Procedure, comprehensive: To cause, using one or more processors, an image rendering model to generate an estimated image of a scene based on at least a plurality of images of the scene, wherein at least one image of the plurality of images is associated with at least one from a different time or view; and to update, using the one or more processors, the image rendering model based on at least the estimated image, the plurality of images, and one or more criteria for motion associated with the estimated image.

[0155] It is understood that the aspects and embodiments described above are purely exemplary and that modifications of details may be made within the scope of protection of the claims.

[0156] Each device, each method and each feature disclosed in the description, and (where applicable) the claims and drawings, may be provided independently or in any suitable combination.

[0157] Reference numerals appearing in the claims are for illustrative purposes only and do not restrict the scope of protection of the claims.

Images