DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
This action is in response to the Amendment filed on 8/20/2026.
Claims 1-18 are pending. Claims 3, 10, 13, 15 have been amended. Claims 16-18 are newly added.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 2, 3, 10, 12, 14, 15, 16, 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Fang et al. (CN106447748B, hereinafter Fang) in view of Anderson (US 20080204457 A1).
Regarding Claim 1, Feng teaches a method of training one or more neural networks (NNs), comprising: (Feng, Page 3; "the step of training radial basis function neural network model", “The radial basis function network is input to after the vector that first input of trained radial basis function neural network model matches Network mode”): receiving training data comprising training input data and training output data, wherein the training input data and the training output data represent one or more training poses (Feng, Page 3; "for each sample animation segment in the sample animation fragment sequence, The kinematic parameter of sample animation segment <read on training input data > is mapped as primary vector, and according to sample animation segment in the sample animation segment sequence Order in column generates secondary vector <read on training output data>"; Page 5, "each sample animation segment in sample animation fragment sequence is Skeleton cartoon"), wherein at least one of the one or more training poses comprise an input angular component, an input linear component, an output angular component, and an output linear component (Feng, Page 7; "the bone parameters of each animation frame it may include the location parameter of root bone and the rotation parameter of each bone", "In general, in animation frame in skeleton cartoon, Location parameter <read on input linear component> can indicate that rotation parameter <read on input angular component > can be indicated with four-dimensional vector"; it is noted the output of the RBFNN-driven fusion likewise yields an animation frame with a rotation parameter and a location parameter for each bone); grouping the input angular components from each of the one or more poses into an input angular component group (Feng, Page 3, "executes spherical surface interpolation to the rotation parameter of bone each in animation frame", It is noted by collecting all bone rotation parameters for common treatment approach of applying the same spherical interpolation to all rotation parameters in each frame implies an angular component group.); grouping the input linear components from each of the one or more poses into an input linear component group (Feng, Page 4, "Linear interpolation is executed to the location parameter of root bone in animation frame"; it is noted by grouping all location (linear) parameters of bones together under a common linear interpolation treatment, can formation of an input linear component group); and providing the training input data as input to train the one or more NNs (Feng, Page 3; "Using the corresponding primary vector of sample animation segment as the radial basis function network The input of network model, and using the corresponding secondary vector of sample animation segment as the defeated of the radial basis function neural network model Out, the radial basis function neural network model is trained"), wherein the input angular component group is evaluated differently than the input linear component group, wherein the evaluation results in the output angular component and the output linear component (Feng, Page 3, “executes spherical surface interpolation to the rotation parameter <read on angular component group> of bone each in animation frame；It is right The location parameter of root bone executes linear interpolation <read on linear component group> in animation frame”);
Anderson further teaches wherein at least one of the one or more training poses comprise an input angular component, an input linear component, an output angular component, and an output linear component (Anderson, Paragraph [0014], "In a further embodiment of creating a baked component, the set of inputs includes a set of animation variables. A portion of the set of animation variables are joint rotation angles <read on angular component> expressed in a defrobulated form. Additionally, each of the set of training posing errors is at least one difference in position <read on linear component> of the point of a character model from a geometrically posed position to at least one position specified by the model component."); grouping the input angular components from each of the one or more poses into an input angular component group (Anderson, Paragraph [0056]"animation variables representing joint rotations <read on angular >are converted to a defrobulated form to prevent the appearance of visual discontinuities.", Paragraph [0058], "animation variables representing joint rotations and not already in defrobulated form are converted into a defrobulated form as described above. The regression analysis is then performed using the defrobulated form of animation variables.")
grouping the input linear components from each of the one or more poses into an input linear component group Anderson.", Paragraph [0058], "animation variables representing joint rotations and not already in defrobulated form are converted into a defrobulated form as described above. The regression analysis is then performed using the defrobulated form of animation variables”)
wherein the input angular component group is evaluated differently than the input linear component group, wherein the evaluation results in the output angular component and the output linear component (Anderson, Paragraph [0056], "In a further embodiment, animation variables representing joint rotations are converted to a defrobulated form to prevent the appearance of visual discontinuities."; "Unlike typical angle representations, the defrobulated angles do not have any angular discontinuities within the normal range of joint motion.", [0056]; Anderson reinforces the concept: angular (rotation) variables are evaluated with a transformation specific to angular space (defrobulated form), while linear variables are evaluated without this transformation — confirming different evaluation regimes for angular vs. linear groups, resulting in outputs in their respective (angular/linear) spaces.
Anderson and Feng are analogous since both deal with training neural networks on articulated character pose data where angular (rotational) and linear (translational) joint components must be evaluated differently to produce accurate pose outputs. Feng provided a way of training an RBFNN on skeletal animation data, evaluating rotation parameters via spherical interpolation (angular group) and root bone location via linear interpolation (linear group) to produce output bone parameters. Anderson provided a way of grouping joint rotation variables (angular) for specialized defrobulated-form conversion before regression, while leaving translational variables (linear) under standard regression i.e. explicitly evaluating angular and linear groups differently to produce distinct angular and linear outputs. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the differential angular/linear component group evaluation taught by Anderson into the modified invention of Feng such that training input data is explicitly grouped into angular and linear component groups evaluated with distinct methods to yield separated output angular and linear components. The motivation is to prevent angular discontinuities (e.g., gimbal lock) that arise when rotation variables are treated identically to translational variables, thereby improving the accuracy and visual quality of the trained pose model.
Regarding Claim 2, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination further teaches wherein at least one of the one or more NNs is a radial basis function neural network (RBFNN) (Feng, Page 3, "the radial basis function neural network model is by each sample cartoon in sample animation fragment sequence What section was trained", "the step of training radial basis function neural network model).
Regarding Claim 3, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination further teaches wherein the angular output component is in special orthogonal group in three dimensions (SO(3)) space (Fang, Page 7, "rotation <read on angular> parameter can be indicated with four-dimensional vector", "spherical surface interpolation method can be used for rotation parameter"; a four-dimensional unit rotation vector interpolated by spherical interpolation represents a rotation in three-dimensional space and thus reads on SO(3) space);
and wherein the output angular component is evaluated using an angular distance metric that differs from a distance metric used to evaluate the output linear component (Fang, Page 3, "executes spherical surface interpolation to the rotation parameter of bone each in animation frame ; It is right The location parameter of root bone executes linear interpolation in animation frame"; spherical surface interpolation on a rotation parameter <read on angular distance metric> inherently uses geodesic angular distance along the sphere, while linear interpolation on the location parameter <read on distance metric used to evaluate the output linear component> inherently uses straight-line/Euclidean distance in position space — these two distance measures differ from each other).
Anderson further teaches wherein the output angular component is evaluated using an angular distance metric that differs from a distance metric used to evaluate the output linear component (Anderson, Paragraph [0056], "animation variables representing joint rotations are converted to a defrobulated form to prevent the appearance of visual discontinuities."; Paragraph [0014], "each of the set of training posing errors is at least one difference in position of the point of a character model from a geometrically posed position to at least one position specified by the model component."; the defrobulated-form treatment applied to angular variables differs from the positional-difference treatment applied to linear variables, thereby applying different metrics to angular vs. linear components).
Anderson and Fang are analogous since both deal with pose systems that evaluate angular and linear components using metrics adapted to their respective spaces. Fang provided a way of measuring interpolation error along a spherical geodesic path for rotation parameters and along a straight-line/Euclidean path for location parameters. Anderson provided a way of transforming angular variables into a defrobulated form to properly measure differences without angular discontinuities, while evaluating linear differences as ordinary positional errors. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the differential distance-metric evaluation taught by Anderson into the modified invention of Fang such that the output angular component is evaluated using an angular distance metric while the output linear component is evaluated using a different (linear) distance metric. The motivation is to provide mathematically appropriate error measures for each variable type — geodesic/angular distance for rotations and Euclidean-type distance for positions — thereby avoiding angular discontinuities during evaluation, as discussed by Anderson in Paragraph [0056].
Regarding Claim 12, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination further teaches wherein evaluation includes one or more components of motion (Feng, Page 6, " By taking animation of walking as an example, the kinematic parameter for animation of walking can To include three physical quantitys such as forward speed, side velocity and turning speed, then sample animation segment be mapped to first Vector include respectively with three components corresponding to the value of forward speed, side velocity and turning speed "; Page 8, "the kinematic parameter of sample animation segment is mapped as primary vector"; It is noted that the evaluation/training input incorporates kinematic parameters, which are components of motion (forward speed, side velocity, turning speed are all motion components). The evaluation of bone parameters further includes rotation (angular component of motion) and location/translation (linear component of motion).
Regarding Claim 14, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination further teaches wherein the output angular component is in quaternion space (Feng, Page 5, "rotation parameter can be indicated with four-dimensional vector"; "spherical surface interpolation method can be used for rotation parameter"; it is noted that the angular output (rotation parameter) is represented as a four-dimensional vector and is processed using spherical surface interpolation. A four-dimensional vector used to represent a rotation, interpolated via spherical interpolation, is the standard definition of a unit quaternion in quaternion space. It is recognized that "four-dimensional vector" with SLERP as a quaternion representation).
Regarding Claim 15, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination further teaches wherein the output linear component is a Euclidean distance function (Fang, Page 8, "Linear interpolation is executed to the location parameter <read on linear output> of root bone in animation frame"; Page 7, "location parameter can indicate that rotation parameter can be indicated with four-dimensional vector, therefore be melted to location parameter with three-dimensional vector When closing interpolation, linear interpolation can be used for location parameter"; linear interpolation between two three-dimensional position vectors follows the straight-line/Euclidean path and reads on a Euclidean distance function);
and wherein the output angular component is evaluated using an angular distance function different from the Euclidean distance function (Fang, Page 3, "executes spherical surface interpolation to the rotation parameter of bone each in animation frame"; spherical surface interpolation of a rotation parameter inherently operates along a geodesic (angular) path on the unit sphere <read on angular distance function> which is distinct from the straight-line Euclidean path used for the linear component).
Anderson further teaches wherein the output linear component is a Euclidean distance function (Anderson, Paragraph [0014], "each of the set of training posing errors is at least one difference in position of the point of a character model from a geometrically posed position to at least one position specified by the model component."; the positional difference in 3D space reads on a Euclidean distance function); and wherein the output angular component is evaluated using an angular distance function different from the Euclidean distance function (Anderson, Paragraph [0056], "animation variables representing joint rotations are converted to a defrobulated form to prevent the appearance of visual discontinuities."; Paragraph [0056], "Unlike typical angle representations, the defrobulated angles do not have any angular discontinuities within the normal range of joint motion."; the defrobulated-angle treatment applied to joint rotations reads on an angular distance function distinct from a Euclidean positional distance).
Anderson and Fang are analogous since both deal with pose systems that measure linear (positional) and angular (rotational) outputs using distinct mathematical functions. Fang provided a way of using linear interpolation (Euclidean distance) for location parameters and spherical interpolation (angular/geodesic distance) for rotation parameters. Anderson provided a way of computing positional errors as ordinary Euclidean position differences while transforming angular variables into a defrobulated form so that angular error is measured without angular discontinuities. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the differential linear-Euclidean vs. angular distance-function evaluation taught by Anderson into the modified invention of Fang such that the output linear component uses a Euclidean distance function and the output angular component uses a different angular distance function. The motivation is to apply a mathematically well-suited distance measure to each variable type — Euclidean for positions and an angular measure for rotations — thereby avoiding angular discontinuities and producing accurate error/loss for each output space, as discussed by Anderson in Paragraph [0056].
Regarding Claim 16, the combination of Fang and Anderson teaches the invention of Claim 1.
The combination further teaches wherein evaluating the input angular component group comprises determining an angular distance metric (Fang, Page 3, "executes spherical surface interpolation to the rotation parameter of bone each in animation frame"; spherical surface interpolation over a set of rotation parameters <read on input angular component group> operates on a geodesic path along the unit sphere and reads on determining an angular distance metric);
and evaluating the input linear component group comprises determining a linear distance metric different from the angular distance metric (Fang, Page 4, "Linear interpolation is executed to the location parameter of root bone in animation frame"; linear interpolation between root-bone location parameters <read on input linear component group> operates along the straight-line Euclidean path and reads on a linear distance metric that differs from the geodesic/angular metric used for rotation parameters).
Anderson further teaches wherein evaluating the input angular component group comprises determining an angular distance metric and evaluating the input linear component group comprises determining a linear distance metric different from the angular distance metric (Anderson, Paragraph [0056], "animation variables representing joint rotations are converted to a defrobulated form to prevent the appearance of visual discontinuities."; Paragraph [0058], "animation variables representing joint rotations and not already in defrobulated form are converted into a defrobulated form as described above. The regression analysis is then performed using the defrobulated form of animation variables."; Paragraph [0014], "each of the set of training posing errors is at least one difference in position of the point of a character model from a geometrically posed position to at least one position specified by the model component."; the defrobulated-angle treatment of joint rotations reads on an angular distance metric while positional differences of the character model points read on a distinct linear distance metric).
Anderson and Fang are analogous since both apply metric-based evaluation to grouped angular and linear variables where the metric for each group is adapted to its space. Fang provided a way of using spherical (angular) interpolation for grouped rotation parameters and linear interpolation for grouped location parameters. Anderson provided a way of grouping joint rotations under a defrobulated-form representation for measurement and grouping positional variables under ordinary positional-difference measurement. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Anderson's grouped angular vs. linear metric determination into the modified invention of Fang such that the input angular component group is evaluated using an angular distance metric and the input linear component group is evaluated using a distinct linear distance metric. The motivation is to apply a metric appropriate to each variable type and thereby avoid the angular-discontinuity problems that arise when angular variables are measured with a linear/Euclidean metric, as discussed by Anderson in Paragraph [0056].
Regarding Claim 17, The combination of Fang and Anderson teaches the invention of Claim 16.
The combination further teaches wherein the linear distance metric comprises a Euclidean distance metric (Fang, Page 8, "Linear interpolation is executed to the location parameter of root bone in animation frame"; Page 7, "linear interpolation can be used for location parameter"; linear interpolation between two three-dimensional position vectors inherently follows the straight-line/Euclidean path and reads on a Euclidean distance metric).
Anderson further teaches wherein the linear distance metric comprises a Euclidean distance metric (Anderson, Paragraph [0014], "each of the set of training posing errors is at least one difference in position of the point of a character model from a geometrically posed position to at least one position specified by the model component."; a difference-in-position measurement in 3D space reads on a Euclidean distance metric).
Anderson and Fang are analogous since both use ordinary positional-difference (Euclidean) measurement for the linear/translational variables of a character pose. Fang provided a way of interpolating root-bone location parameters along a straight-line path in 3D space. Anderson provided a way of measuring positional posing errors as position differences in 3D space. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to incorporate Anderson's Euclidean positional-difference measurement into the modified invention of Fang such that the linear distance metric comprises a Euclidean distance metric. The motivation is to provide a mathematically well-defined and standard measure of positional differences for the linear component, as Euclidean distance is the canonical measure of separation between points in 3D space, as discussed by Anderson in Paragraph [0014].
Claim(s) 4, is/are rejected under 35 U.S.C. 103 as being unpatentable over Fang et al. (CN106447748B, hereinafter Fang) in view of Anderson (US 20080204457 A1) as applied to Claim 1 above and further in view of Lin et al. (CN108985457B, hereinafter Lin).
Regarding Claim 4, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination does not explicitly disclose but Lin teaches wherein at least one of the one or more NNs is a feed-forward neural network (FFNN) (Lin, Page 4, "For a classical feedforward network structure in which all layers share the same linear and nonlinear transformations, the forward process in this feedforward network is equivalent to using gradient descent to minimize It is an iterative process to simplify a certain function F(x)", [Summary]; "In a feedforward neural network where all layers share the same linear transformation, the relationship between the output x k of the kth layer and the output x k +1 of the k+1th layer is expressed as Equation 1: x k+1 = φ(Wx k ) (Equation 1) where φ is an activation function, such as Sigmoid or ReLU", it is noted a feedforward neural network (FFNN) where information flows layer-to-layer without feedback loops which is direct structural embodiment of a feedforward NN).
Lin and Feng are analogous since both deal with neural network architectures applied to motion and pose data processing, where the network maps structured input vectors to output parameters through layered transformations. Feng provided a way of training an RBFNN on kinematic animation parameters to learn blend weight mappings for skeletal pose generation. Lin provided a way of implementing a classical feedforward neural network (FFNN) where information flows unidirectionally layer-to-layer via x_{k+1} = φ(Wx_k), with no feedback connections, using activation functions such as ReLU or Sigmoid. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the feedforward neural network architecture taught by Lin into the modified invention of Feng such that the one or more NNs used in training are implemented as FFNNs with unidirectional, layered information flow. The motivation is to simplify the network architecture and improve training predictability by using a well-established feedforward structure that avoids the complexity of recurrent feedback connections.
Claim(s) 5, 6, 7, 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Fang et al. (CN106447748B, hereinafter Fang) in view of Anderson (US 20080204457 A1) and further in view of Lin et al. (CN108985457B, hereinafter Lin) as applied to Claim 4 above and further in view of Marsden et al. (US 11841920 B1, hereinafter Marsden).
Regarding Claim 5, the combination of Feng, Anderson and Lin teaches the invention in Claim 4.
The combination further teaches wherein the FFNN is a fully connected network (Lin, Page 9, "we use a fully connected linear transformation Wx and an activation function φ"; "use Xavier[] to initialize the fully connected layers"; it is noted the fully connected linear transformations as the basis of its FFNN formulation).
As explained in rejection of claim 4, the obviousness for combining of feedforward neural network (FFNN) of Lin into Fing is provided above.
Marsden further teaches the same limitation (Marsden, Column 6, Line 53-56, "Both, master networks and expert networks, are fully connected neural networks that take a feature vector of an input hand image and produce a prediction of the hand pose."; Column 21, Line 30-32, "In a fully connected layer of a neural network, all the elements of all the features of the previous layer get used in the calculation of each element of each output feature")
Marsden and Feng are analogous since both deal with training layered neural networks on articulated character pose data to learn mappings between kinematic input parameters and pose output parameters. Feng provided a way of training a structured RBFNN with input, hidden, and output layers on animation kinematic parameters to generate blend weight outputs for skeletal pose blending. Marsden provided a way of implementing master and expert networks as fully connected neural networks where all elements of all features from the previous layer contribute to each element of each output feature which confirming a fully connected FFNN architecture. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the fully connected network structure taught by Marsden into the modified invention of Feng such that the FFNN is implemented as a fully connected network where all neurons between adjacent layers are interconnected. The motivation is to ensure comprehensive feature integration across all input dimensions during pose regression, as fully connected layers are proven effective at learning complex non-linear pose mappings.
Regarding Claim 6, the combination of Feng, Anderson and Lin teaches the invention in Claim 4.
The combination does not explicitly disclose but Marsden teaches wherein the FFNN comprises a single hidden layer (Marsden, Column 12, Line 9-11, "Neural network 300 has three inputs, two neurons in the hidden layer and two neurons in the output layer."; it noted the system a fully connected neural network with a single hidden layer, i.e. the network has an input layer, one hidden layer (with two neurons), and an output layer. This is a direct structural embodiment of an FFNN with a single hidden layer; Column 18, Line 9-13, "A CNN of seven layers includes an input layer, a convolution layer C1, a sub-sampling layer S1, another convolution layer C2, another sub-sampling layer S2, a hidden layer H and an output layer F."; Column 18, Line 37-40, "The hidden layer H includes one hundred and seventy neurons, each neuron connected to four hundred neurons of layer S2.").
Marsden and Feng are analogous since both deal with training layered neural networks on articulated character pose data, where the network architecture comprises defined input, hidden, and output layers performing non-linear transformations on pose feature vectors. Feng provided a way of training an RBFNN structured with one input layer, one Gaussian kernel hidden layer, and one output layer — effectively a three-layer network with a single hidden layer — to learn blend weight mappings from skeletal animation kinematic parameters. Marsden provided a way of implementing a fully connected neural network explicitly described with a single hidden layer (e.g., Neural network 300 with an input layer, one hidden layer, and an output layer), directly embodying the FFNN single hidden layer architecture for pose estimation tasks. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the single hidden layer FFNN architecture taught by Marsden into the modified invention of Feng such that the FFNN comprises exactly one hidden layer between the input and output layers. The motivation is to balance model expressiveness and computational efficiency, as a single hidden layer provides sufficient capacity for pose regression while minimizing parameter count and training time, as demonstrated by Marsden's successful single hidden layer network for hand pose estimation.
Regarding Claim 7, the combination of Feng, Anderson, Lin and Marsden teaches the invention in Claim 4.
The combination further teaches wherein the single hidden layer is a residual NN block (Lin, Page 7, "x k+1 = φ(Wx k )+β(x k -x k-1 ) (Equation 4)",; "x k+1 =T(x k )+β 1 x k +β 2 x k-1 , (Equation 7)"; Lin explicitly teaches a network layer/block whose output is the sum of a transformed input T(x_k) and a skip-connected term β₁x_k , it directly corresponding to a residual block structure where the output = transformation + skip connection from the input. This is the defining characteristic of a residual NN block (as in ResNet). Page 9, "The classical network structures RreNet and DenseNet are special cases of the network structure (Equation 7 and Equation 8), respectively."; Page 2, "the former adds side branches to the traditional feedforward network structure"; It is noted that acknowledgement of ResNet (which is the canonical residual NN block architecture) is a special case of its proposed framework (Equation 7), and describes ResNet's side branches (skip connections) as the feature that distinguishes it. Lin's HB-Net block (with β₁=1, β₂=−1: x_{k+1} = T(x_k) + x_k − x_{k-1}) is a generalized residual block with momentum.
Lin and Feng are analogous since both deal with training feedforward neural networks with layered architectures on pose and motion data, where individual network blocks/layers perform structured transformations on input feature vectors. Feng provided a way of training an RBFNN with a single hidden Gaussian kernel layer on skeletal animation kinematic parameters to generate blend weight outputs for bone pose generation. Lin provided a way of implementing a residual NN block structure where the output is the sum of a transformed input T(x_k) and a skip-connected term x_k — the defining characteristic of a residual block — and explicitly subsumes ResNet as a special case of its HB-Net framework, confirming the residual block as a valid hidden layer structure within an FFNN. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the residual NN block structure taught by Lin into the modified invention of Feng such that the single hidden layer of the FFNN is implemented as a residual NN block with a skip connection. The motivation is to improve gradient flow and training stability during backpropagation by using skip connections that allow gradients to bypass transformation layers, a well-established benefit of residual architectures as recognized in Lin.
Regarding Claim 8, the combination of Feng, Anderson and Lin teaches the invention in Claim 4.
The combination further teaches wherein the FFNN comprises rectified linear unit activation functions Lin, Page 4, "x k+1 = φ(Wx k ) (Equation 1) where φ is an activation function, such as Sigmoid or ReLU"; Page 9,"the first layer network is a composite of convolution, BN and ReLU"; "T (x) is the composite of BN, ReLU and convolution"; it is noted uses ReLU as the activation function within its feedforward network layers).
Lin and Feng are analogous since both deal with training feedforward neural networks on pose and motion-related data, where activation functions are applied within network layers to introduce non-linearity into the learned input-to-output mappings. Feng provided a way of training an RBFNN using Gaussian kernel functions in the hidden layer to process kinematic animation parameter vectors and generate blend weight outputs for skeletal pose blending. Lin provided a way of implementing FFNN layers with explicit ReLU activation functions — expressed as φ in x_{k+1} = φ(Wx_k) with ReLU listed as a candidate alongside Sigmoid — and further confirming ReLU use in its composite layer structure (convolution + BN + ReLU). Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the ReLU activation functions taught by Lin into the modified invention of Feng such that the FFNN used in training comprises rectified linear unit activation functions within its layers. The motivation is to improve gradient flow and training convergence by using ReLU, which avoids the vanishing gradient problem associated with sigmoid-type activations and is computationally efficient for pose regression tasks.
Marsden further teaches the same limitation (Marsden, Column 16, Line 60-67 "Non-linear layers 108 use a variety of specific functions to implement the non-linear triggering, including the rectified linear units (ReLUs), hyperbolic tangent, absolute of hyperbolic tangent, sigmoid and continuous trigger (non-linear) functions. In one implementation, a ReLU activation implements the function y=max(x, 0) and keeps the input and output sizes of a layer the same"; Column 21, Line 48-50, “(LRelLU) activation function is used by the master networks 110 and expert networks 112”).
Marsden and Feng are analogous since both deal with training multi-layer neural networks on articulated pose data where activation functions are applied within network layers to introduce non-linearity into the learned mappings. Feng provided a way of training an RBFNN using Gaussian kernel functions in the hidden layer to process kinematic animation parameter vectors and generate blend weight outputs. Marsden provided a way of training fully connected neural networks with explicit ReLU activations — implementing y = max(x, 0) — and confirming LReLU use in both master and expert networks for hand pose estimation. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the ReLU activation functions taught by Marsden into the modified invention of Feng such that the FFNN used in training comprises rectified linear unit activation functions. The motivation is to improve gradient flow and training convergence, as ReLU avoids the vanishing gradient problem associated with sigmoid activations and is computationally efficient for pose regression tasks.
Claim(s) 9, 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Fang et al. (CN106447748B, hereinafter Fang) in view of Anderson (US 20080204457 A1) as applied to Claim 1 above and further in view of Marsden et al. (US 11841920 B1, hereinafter Marsden).
Regarding Claim 9, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination does not explicitly disclose but Marsden teaches wherein the training input data includes data that describes a pose of a digital character associated with an alternative reality (AR) or virtual reality (VR) system (Marsden, Column 1, Line 60-66, "Particularly in the Augmented reality (AR) and virtual reality (VR) context, such traditional paradigms greatly diminish the user experience. Accordingly, the technology disclosed allows users to interact with the virtual interfaces generated in AR/VR environment using free-form in-air gestures."; Column 58, Line 25-30, "training a first set of atemporal generalist neural networks with the hand position images to produce estimated hand position parameters"; Column 58, Line 63-67, "the hand position parameters are a plurality of joint angles in three-dimensional (3D) space"; Column 57, Line 23-26, "the neural network module 5726 stores the convolutional neural network 101, which operates in conjunction with the AR/VR environment 5600").
Marsden and Feng are analogous since both deal with training neural networks on pose data of articulated digital models — skeletal characters and hand joints respectively — where training input data encodes the spatial configuration of a digital character in a computing environment. Feng provided a way of training an RBFNN on kinematic animation parameters of skeletal digital characters to produce bone pose outputs (rotation and position) for animation generation. Marsden provided a way of training neural networks specifically in an AR/VR context, where input hand pose images and joint angle parameters describe the pose of a digital hand character interacting with virtual interfaces in an AR/VR environment 5600. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the AR/VR digital character pose training input taught by Marsden into the modified invention of Feng such that the training input data includes data describing the pose of a digital character associated with an AR/VR system. The motivation is to extend the trained pose network to support digital character interaction in immersive AR/VR environments, leveraging Marsden's established practice of training neural networks on structured joint pose inputs within AR/VR systems.
Regarding Claim 11, the combination of Feng and Anderson teaches the invention in Claim 1.
The combination does not explicitly disclose but Marsden teaches wherein the output angular component and the output linear component includes data that describes a pose of a digital character associated with an alternative reality (AR) or virtual reality (VR) system (Marsden, Column 6, Line 53-56, "Both, master networks and expert networks, are fully connected neural networks that take a feature vector of an input hand image and produce a prediction of the hand pose."; Column 22, Line 54-58, "These hand position parameters are in the form of joint location models, joint angel models, capsule models, skeleton models, volumetric models and/or mesh models, muscle hand models, each in 2D and/or 3D space."; Column 4, Line 34-40 "FIGS. 40A, 40B and 40C show one implementation of skeleton hand models fitted to estimated joint covariances interacting with and manipulating virtual objects (e.g., depicted boxes) in an augmented reality (AR)/virtual reality (VR) environment 5600."; Column 1, Line 60-66, "Particularly in the Augmented reality (AR) and virtual reality (VR) context, such traditional paradigms greatly diminish the user experience. Accordingly, the technology disclosed allows users to interact with the virtual interfaces generated in AR/VR environment using free-form in-air gestures."; Column 58, Line 25-30, "training a first set of atemporal generalist neural networks with the hand position images to produce estimated hand position parameters"; Column 58, Line 63-67, "the hand position parameters are a plurality of joint angles in three-dimensional (3D) space"; Column 57, Line 23-26, "the neural network module 5726 stores the convolutional neural network 101, which operates in conjunction with the AR/VR environment 5600").
; Marsden teaches that the trained NNs produce as output a prediction of a hand pose in the form of joint angles (angular component) and joint locations (linear component), which together describe the pose of a digital hand character (virtual hand) in an AR/VR environment. The output angular (joint angles) and linear (3D joint location coordinates) components directly correspond to the claimed output angular and linear components describing a digital character pose in AR/VR. Taught by Marsden.)
Marsden and Feng are analogous since both deal with training neural networks whose outputs describe the full pose of an articulated digital character through angular (rotational) and linear (positional) components in 3D space. Feng provided a way of training an RBFNN to output bone rotation parameters (output angular component) and root bone location parameters (output linear component) for each animation frame of a skeletal digital character. Marsden provided a way of training fully connected networks that produce hand pose predictions in the form of joint angles (output angular component) and 3D joint locations (output linear component) of a digital hand character operating within an AR/VR environment 5600. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention was made to incorporate the AR/VR digital character pose output framework taught by Marsden into the modified invention of Feng such that the output angular component and the output linear component describe the pose of a digital character associated with an AR/VR system. The motivation is to enable the trained network to produce pose outputs directly usable for digital character rendering in AR/VR environments, where both joint rotation (angular) and joint location (linear) outputs are required for accurate and responsive character animation.
Allowable Subject Matter
Claim 10 is allowed.
Claim 13, 18 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
The following is an examiner’s statement of reasons for allowance:
Regarding Claim 10, the prior art of record, specifically the prior art Fang et al. (CN106447748B) teaches a way of training an RBFNN on skeletal animation data, evaluating rotation parameters via spherical interpolation and root bone location via linear interpolation to produce output bone parameters. Anderson (US 20080204457 A1) teaches a way of grouping joint rotation variables for specialized defrobulated-form conversion before regression. However, none of the prior art cited alone or in combination provides motivation to teach "wherein the training input data includes data that represents a low-order skeleton of a digital character, and the training output data represents a high-order skeleton of a digital character associated with an alternative reality (AR) or virtual reality (VR) system" as recited in independent claim 10. Therefore, applicant's amendment rewriting claim 10 in independent form is persuasive.
Regarding Claim 13, the prior art of record, specifically Fang et al. (CN106447748B) teaches applying spherical surface interpolation to the rotation parameter of bones and linear interpolation to the location parameter of the root bone in an animation frame, and Anderson (US 20080204457 A1) teaches that geometric operations include translation, scaling, and rotation as manipulations of a character model. However, none of the prior art cited alone or in combination provides motivation to teach "wherein the rotational motion, translational motion, and scale are evaluated as separate components" as required by amended claim 13. While the cited references may separately mention bone size or scaling as a general geometric operation, no reference teaches or suggests evaluating scale as a distinct, separately-evaluated component alongside rotational and translational motion within the recited training framework. Therefore, applicant's amendment to claim 13 is persuasive.
Regarding Claim 18, the prior art of record, specifically Fang et al. (CN106447748B) teaches an RBFNN trained on primary-vector kinematic parameters that outputs secondary-vector fusion weight coefficients, where the differential handling of rotation parameters (spherical interpolation) and location parameters (linear interpolation) occurs during post-network animation fusion. Anderson (US 20080204457 A1) teaches defrobulated-form preprocessing of joint rotation variables prior to statistical regression analysis. However, none of the prior art cited alone or in combination provides motivation to teach "wherein the angular component group and the linear component group are maintained separately during training of the one or more neural networks" as recited by new claim 18. The cited references at most disclose differential treatment of angular and linear variables in a preprocessing or post-network animation-blending step, not the maintenance of separate angular and linear groupings throughout the training process of the neural network itself. Therefore, new claim 18 is persuasive over the prior art of record.
Response to Arguments
The rejection of Claims 1-15 under Nonstatutory Double Patenting are withdrawn in view of Applicant filed eTerminal/Terminal Disclaimer on 8/20/2026.
Applicant’s arguments with respect to claim 1, 2, 3, 12, 13, 14, and 15, filed on 8/20/2026, with respect to rejection under 35 USC § 103 have been considered but they are not persuasive.
In Regard to Claim 1, Applicant argues that Fang's differential treatment of rotational versus positional parameters occurs only during animation blending/interpolation of output data, not during neural-network training, and that Anderson's defrobulation preprocessing fails to disclose grouping angular and linear training data into separate groups evaluated differently.
In response to the argument, claim 1 does not recite that the differential evaluation must occur within the internal computations of the NN or within a training-time loss function. Under the broadest reasonable interpretation, "evaluated differently" encompasses any differential processing of grouped angular versus linear parameters within the pose-training pipeline that yields distinct output angular and linear components. Fang expressly trains an RBFNN on skeletal-animation data (Fang, Page 3) in which grouped rotation parameters are subjected to spherical surface interpolation and grouped location parameters are subjected to linear interpolation (Fang, Pages 3-4), yielding the recited output angular and linear components. Anderson further teaches grouping joint rotations under defrobulated-form conversion prior to regression while treating positional variables under ordinary positional-difference measurement (Anderson, Paragraphs [0056], [0058]), confirming a differential evaluation regime for grouped angular versus linear variables. Therefore, Applicant's argument is not persuasive.
in regard to Claim 3 (as amended), Applicant argues that neither reference discloses an SO(3)-based differential distance-metric framework.
In response to the argument, Fang teaches that the rotation parameter is a four-dimensional vector processed via spherical surface interpolation (Fang, Page 7), which reads on SO(3) space. Fang's spherical interpolation inherently determines an angular (geodesic) distance along the unit sphere, while Fang's linear interpolation on the location parameter inherently determines a straight-line distance in position space (Fang, Pages 3-4) — two mathematically distinct metrics. Anderson reinforces this by measuring angular differences via defrobulated form (Anderson, Paragraph [0056]) while measuring positional errors as ordinary positional differences (Anderson, Paragraph [0014]). Therefore, Applicant's argument is not persuasive.
In regard to Claim 15 (as amended), Applicant argues that neither reference establishes a Euclidean distance function for the linear output and a different angular distance function for the angular output.
In response to the argument, Fang's linear interpolation between three-dimensional position vectors (Fang, Pages 7-8) follows the straight-line path — precisely the Euclidean distance path — and reads on a Euclidean distance function. Fang's spherical surface interpolation on the rotation parameter (Fang, Page 3) operates along the geodesic path on the unit sphere, an angular distance function fundamentally distinct from the Euclidean path. Anderson reinforces the distinction by measuring positional errors as position differences (Anderson, Paragraph [0014]) while separately treating joint rotations via defrobulated form (Anderson, Paragraph [0056]). Therefore, Applicant's argument is not persuasive.
In regard to Claims 2, 12, 24, they directly/indirectly depends on independent Claim 1. Applicant does not argue anything other than the independent claim 1. The limitations in those claims in conjunction with combination previously established as explained.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
US 20180259608 A1 AUTOMATED CARDIAC VOLUME SEGMENTATION
US 20180218502 A1 AUTOMATED SEGMENTATION UTILIZING FULLY CONVOLUTIONAL NETWORKS
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to YUJANG TSWEI whose telephone number is (571)272-6669. The examiner can normally be reached 8:30am-5:30pm EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kent Chang can be reached at (571)272-7667. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/YuJang Tswei/Primary Examiner, Art Unit 2614
