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 .
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, 4, 11, 14, 15, 16, 19, 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sosulnikov et al. (US 20240340425 A1, hereinafter Sosulnikov) in view of Supikov et al. (US 20240355047 A1, hereinafter Supikov).
Regarding Claim 1, Sosulnikov teaches a method comprising: estimating distribution statistics for a 3D Gaussian [[ Splatting ]] model corresponding to one or more 3D Gaussian [[ Splatting ]] parameters, (Sosulnikov, Paragraph [0028], "entropy encoding the signal using one or more Gaussian mixture model (GMM) with determined GMM parameters"; [0036], "performing an optimization algorithm using GMM cumulative distribution functions to obtain the determined GMM parameters"; [0150], "a signal encoder comprises a GMM parameters optimization gradient decent algorithm, which may be performed in parallel"; [0015], “Embodiments of the present disclosure may be applied in the technological fields of 2D/3D Image”; [0020], "the GMM parameters comprise for each Gaussian a mean value, a standard deviation, and/or a weight"), wherein the distribution statistics comprise a distribution model (Sosulnikov, Paragraph [0068], "In case when base distributions are normal a mixture of them is called Gaussian Mixture Model (GMM)."; [0004], "Entropy coding utilizes an entropy model (prior distribution) to encode and decode a signal."), performing quantization of the one or more 3D Gaussian [[ Splatting ]] parameters; (Sosulnikov, Paragraph [0023], "The frequency tables are built on a range from minimum possible signal value −QS/2 to maximum possible signal value +QS/2, wherein QS stands for quantization step. The quantized samples have a quantization step QS."), encoding the quantized one or more 3D Gaussian [[ Splatting ]] parameters based on the estimated distribution statistics. (Sosulnikov, Paragraph [0147], "entropy encoding a signal using one or more Gaussian mixture model (GMM) with determined GMM parameters … and generating at least one bitstream comprising the entropy encoded signal and the determined GMM parameters."; [0155], "the step of building GMM entropy models with these parameters (quantized and clipped, but not scaled), 4) the step of entropy encoding signal's channels with built entropy model with corresponding parameters.").
But Sosulnikov does not explicitly disclose that the model is a 3D Gaussian Splatting model and that the parameters are 3D Gaussian Splatting parameters.
However, Supikov teaches a 3D Gaussian Splatting model corresponding to one or more 3D Gaussian Splatting parameters (Supikov, Paragraph [0017], "three-dimensional (3D) gaussian splats (3DGS) algorithms to represent geometry and directional radiance distribution in a 3D scene have been developed as an alternative to neural representations, such as NeRF, due to remarkably high rendering speed. A 3DGS algorithm generates a 3DGS representation of the 3D scene as a collection of semi-transparent blobs of different sizes and orientations with directional radiance information encoded using spherical harmonics."; [0044], "the gaussian seed creation circuitry 435 creates an initial 3D gaussian splat Gi, also referred to as a seed gaussian Gi, with mean set to (xi, yi, zi), a diffuse color (e.g., zero-order spherical harmonic coefficient) sct to (ri, gi, bi), and a post-activation opacity set to αi=1−exp (−σiδi)."; [0028], "the 3DGS initialization component 320 also initializes the covariance matrix of a given initial gaussian splat 325 based on an activation scale parameter").
Supikov and Sosulnikov are analogous since both of them are dealing with signal processing and representation of data using parameterized models. Sosulnikov provides a way of efficiently encoding and decoding signals based on estimated distribution statistics using GMM-based entropy coding. Supikov provides a 3D Gaussian Splatting model with specific parameters (geometry, color, opacity, spherical harmonics, covariance) that represent a 3D scene. 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 3D Gaussian Splatting model and its parameters taught by Supikov into the distribution-based entropy encoding method of Sosulnikov such that the 3D Gaussian Splatting parameters are encoded using the distribution-based entropy coding method. The motivation is to reduce storage and transmission requirements of 3DGS representations, since 3DGS models require a large amount of memory as acknowledged in the art, and applying proven distribution-based entropy coding techniques to compress those parameters would efficiently address this known problem.
Regarding Claim 4, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches performing pruning of the 3D Gaussian Splatting model (Supikov, Paragraph [0019], "balancing densification and pruning"), the 3D Gaussian Splatting model comprises two or more Gaussians (Supikov, Paragraph [0017], "A 3DGS algorithm generates a 3DGS representation of the 3D scene as a collection of semi-transparent blobs of different sizes and orientations”; [0019], "balancing densification and pruning"; [0018], “the blobs (also referred to as gaussian splats, splats or gaussians) can be efficiently projected and rasterized”).
Supikov and Sosulnikov are analogous since both deal with processing and encoding parameterized data representations. Sosulnikov provides a distribution-based entropy encoding method for compressing signals. Supikov provides a 3DGS model that undergoes pruning during generation to reduce redundant Gaussians. 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 pruning of the 3D Gaussian Splatting model taught by Supikov into the modified invention of Sosulnikov such that pruning is performed on the 3DGS model prior to encoding. The motivation is to reduce the number of Gaussians in the 3DGS representation before compression, thereby reducing the amount of data to be encoded and improving coding efficiency, as Supikov recognizes in Paragraph [0019] that pruning is part of the process of refining the 3DGS representation.
Regarding Claim 11, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches encoding one or more of the estimated distribution statistics (Sosulnikov, Paragraph [0028], "generating at least one bitstream comprising the entropy encoded signal and the determined GMM parameters."), (Sosulnikov, Paragraph [0155], "the step of processing and writing to the bitstream parameters according to the defined procedure").
Regarding Claim 14, it recites limitations similar in scope to the limitations of claim 1, but in a system. As shown in the rejection, the combination of Sosulnikov and Supikov disclose the limitations of claims 1. Additionally, Sosulnikov discloses an apparatus that maps to Paragraph [0041]-[0042], "According to a third aspect, a decoder for decoding an encoded signal is provided, the decoder comprising processing circuitry configured to perform the method"; [0041], "According to the fourth aspect, an encoder for encoding a signal is provided, the encoder comprising processing circuitry configured to perform the method of encoding a signal"; [0042], "a computer program is provided, comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method").Thus, Claim 14 is met by Sosulnikov according to the mapping presented in the rejection of claims 1, given the method corresponds to the apparatus.
Regarding Claim 15, Sosulnikov teaches a method comprising: obtaining estimated distribution statistics for a 3D Gaussian [[ Splatting ]]model, (Sosulnikov, Paragraph [0062], "obtaining the GMM parameters based on the information from the at least one bitstream"), (Sosulnikov, Paragraph [0014], "obtaining the GMM parameters based on the information from the at least one bitstream"; [0015], “Embodiments of the present disclosure may be applied in the technological fields of 2D/3D Image”), wherein the distribution statistics comprise a distribution model; (Sosulnikov, Paragraph [0068], "In case when base distributions are normal a mixture of them is called Gaussian Mixture Model (GMM)."; [0004], "Entropy coding utilizes an entropy model (prior distribution) to encode and decode a signal."), decoding one or more quantized 3D Gaussian [[ Splatting ]] parameters based on the estimated distribution statistics, (Sosulnikov, Paragraph [0014], "entropy decoding the signal using the GMMs with the obtained GMM parameters."; [0022], "building signal frequency tables based on the decoded GMM parameters; wherein the step of entropy decoding the signal comprises using the signal frequency tables for decoding the signal."), wherein the one or more quantized 3D Gaussian [[ Splatting ]] parameters correspond to the 3D Gaussian [[ Splatting ]]model; (Sosulnikov, Paragraph [0024], "the signal includes one or more channels and the step of entropy decoding the signal comprises entropy decoding each channel with a corresponding set of GMM parameters."), and dequantizing the one or more quantized [[ 3D Gaussian Splatting ]] parameters to generate the 3D Gaussian [[ Splatting ]] parameters. (Sosulnikov, Paragraph [0023], "The frequency tables are built on a range from minimum possible signal value −QS/2 to maximum possible signal value +QS/2, wherein QS stands for quantization step. The quantized samples have a quantization step QS."; it is noted the dequantization (inverse quantization) is the inherent inverse operation of quantization at the decoder side, which would be understood by a person of ordinary skill in the art to recover the signal from its quantized form).
But Sosulnikov does not explicitly disclose that the model is a 3D Gaussian Splatting model and that the parameters are 3D Gaussian Splatting parameters.
However, Supikov teaches a 3D Gaussian Splatting model and 3D Gaussian Splatting parameters (Supikov, Paragraph [0017], "three-dimensional (3D) gaussian splats (3DGS) algorithms to represent geometry and directional radiance distribution in a 3D scene have been developed as an alternative to neural representations, such as NeRF, due to remarkably high rendering speed. A 3DGS algorithm generates a 3DGS representation of the 3D scene as a collection of semi-transparent blobs of different sizes and orientations with directional radiance information encoded using spherical harmonics."; [0044], "the gaussian seed creation circuitry 435 creates an initial 3D gaussian splat Gi, also referred to as a seed gaussian Gi, with mean set to (xi, yi, zi), a diffuse color (e.g., zero-order spherical harmonic coefficient) sct to (ri, gi, bi), and a post-activation opacity set to αi=1−exp (−σiδi).").
Supikov and Sosulnikov are analogous since both deal with signal processing and parameterized data representation. Sosulnikov provides a distribution-based entropy decoding and dequantization method. Supikov provides the 3DGS model with specific parameters. 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 3D Gaussian Splatting model and parameters taught by Supikov into the distribution-based decoding method of Sosulnikov such that the decoder decodes and dequantizes 3D Gaussian Splatting parameters. The motivation is to enable efficient decompression of compressed 3DGS representations for rendering, since the encoding of 3DGS data (as motivated in the Claim 1 analysis) necessarily requires a corresponding decoding process.
Regarding Claim 16, the combination of Sosulnikov and Supikov teaches the invention in Claim 15.
The combination further teaches obtaining the estimated distribution statistics comprises decoding one or more of the estimated distribution statistics (Sosulnikov, Paragraph [0014], "obtaining the GMM parameters based on the information from the at least one bitstream; and entropy decoding the signal using the GMMs with the obtained GMM parameters."; [0134], "Entropy decode GMM parameters: Decode μ1 j, . . . , μN j′ j—Gaussians' quantized means for j GMM."; [0111], "Entropy decode GMMs parameters range information").
Regarding Claim 19, the combination of Sosulnikov and Supikov teaches the invention in Claim 15.
The combination further teaches determining the one or more 3D Gaussian Splatting parameters [[ are one of: spherical harmonic DC coefficients, geometry parameters, shape parameters, and DC coefficients of color, ]] wherein the distribution model is a Gaussian distribution(Sosulnikov, Paragraph [0068], "In case when base distributions are normal a mixture of them is called Gaussian Mixture Model (GMM)"; [0066], "FIG. 3 illustrates respective examples for Gaussian distribution functions (upper part) and corresponding cumulative Gaussian distribution functions (lower part).").
Sosulnikov does not explicitly disclose but Supikov teaches spherical harmonic DC coefficients (Supikov, Paragraph [0050], "sets a zero order spherical harmonic parameter of the initial 3D gaussian splat based on the color value obtained from the trained neural representation"), geometry parameters (Supikov, Paragraph [0044], "an initial 3D gaussian splat Gi, also referred to as a seed gaussian Gi, with mean set to (xi, yi, zi)"), shape parameters (Supikov, Paragraph [0028], "initializes the covariance matrix of a given initial gaussian splat 325 based on an activation scale parameter") and DC coefficients of color (Supikov, Paragraph [0027], “sets the color of the given initial gaussian splat 325 based on the color of the corresponding SIM point 330”).
Supikov and Sosulnikov are analogous since both deal with processing parameterized data for encoding and decoding. Sosulnikov provides a Gaussian distribution model for entropy decoding of signals. Supikov provides specific 3DGS parameter types including DC spherical harmonics, geometry, and shape parameters that characterize the 3DGS model. 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 specific 3DGS parameter types (spherical harmonic DC coefficients, geometry parameters, shape parameters) taught by Supikov into the modified decoding method of Sosulnikov such that the Gaussian distribution model is used to decode these specific types of 3DGS parameters. The motivation is to apply a well-suited statistical distribution (Gaussian) when decoding position, shape, and DC color parameters of the 3DGS model, since these parameter types typically exhibit symmetric, bell-shaped distributions that are naturally modeled by a Gaussian distribution, thereby improving entropy decoding accuracy.
Regarding Claim 20, the combination of Sosulnikov and Supikov teaches the invention in Claim 15.
The combination further teaches obtaining the estimated distribution statistics comprises using one or more pre-determined and fixed parameters (Sosulnikov, Paragraph [0195], "scale_mu_coding_mode_flag (predetermined or signaled)"; [0195], "If scale_mu_coding_mode_flag==0: nothing to encode, scaleμ=predefined_scaleμ").
Claim(s) 2, 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sosulnikov et al. (US 20240340425 A1, hereinafter Sosulnikov) in view of Supikov et al. (US 20240355047 A1, hereinafter Supikov) as applied to Claim 1, 15 above respectively and further in view of Kreis et al. (US 20240005604 A1, hereinafter Kreis).
Regarding Claim 2, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches determining the one or more 3D Gaussian Splatting parameters are spherical harmonic AC coefficients [[, wherein the distribution model is a Laplace distribution ]]. (Supikov, Paragraph [0047], "the gaussian seed creation circuitry 435 computes higher order spherical harmonics for the color of a given 3D gaussian splat Gi by sampling radiance at the sample Xi in multiple directions in a hemi-sphere around an estimated surface normal or in the sample's ray direction"; [0003], "directional radiance information encoded using spherical harmonics").
The combination of Sosulnikov nor Supikov does not explicitly disclose but Kreis teaches wherein the distribution model is a Laplace distribution for the spherical harmonic AC coefficients (Kreis, Paragraph [0043], “Furthermore, pξ(x0|h0, z0) denotes the decoder, parametrized as a factorial Laplace distribution with predicted means and fixed unit scale parameter”; [0050], “a probability flow ODE used to further encode into the latent DDMs' Gaussian priors, where spherical interpolation can be performed with valid shapes expected along the interpolation path”).
Kreis and Sosulnikov are analogous since both deal with entropy coding of quantized data using estimated distribution models. Sosulnikov provides GMM-based distribution modeling for entropy encoding. Kreis provides Laplacian distribution modeling for entropy coding of quantized transform coefficients. 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 Laplacian distribution model taught by Kreis into the modified invention of Sosulnikov and Supikov such that a Laplace distribution is used as the distribution model when encoding the spherical harmonic AC coefficients of the 3DGS model. The motivation is to use a distribution model that better fits the statistical characteristics of AC coefficients, which typically exhibit a peaked distribution around zero with heavy tails—a characteristic well-modeled by the Laplacian distribution—thereby improving entropy coding efficiency, as discussed by Kreis in Step (2330).
Regarding Claim 17, it recites limitations similar in scope to the limitations of Claim 2 and therefore is rejected under the same rationale.
Claim(s) 5-8, 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sosulnikov et al. (US 20240340425 A1, hereinafter Sosulnikov) in view of Supikov et al. (US 20240355047 A1, hereinafter Supikov) as applied to Claim 1 above and further in view of Looper et al. (US 12380647 B1, hereinafter Looper).
Regarding Claim 5, the combination of Sosulnikov and Supikov teaches the invention in Claim 4.
The combination does not explicitly disclose but Looper teaches assigning an importance level to each of the two or more Gaussians (Looper, Column 6, Line 20-25 “splat generation system 100 assigns (at 406) the different priority values directly to the 3D primitives of the 3D model that are included in the selected region or part”; Column 6, Line 35-40 “The AI/ML techniques may also dynamically determine the importance of a 3D model based on a popularity measure derived from Internet traffic”). eliminating, from the 3D Gaussian Splatting model, one or more of the two or more Gaussians with a corresponding importance level below a threshold (Looper, Column 13, Line 17-21, “Training (at 906) the 3D model includes replacing the set of two or more splats representing a not commonly or consistently viewed part or element of the 3D model with a single lower fidelity splat”; Column 4, Line 24-26, “the viewing path includes generating priority values <read on importance level > for different volumes or regions of the 3D model with the priority value for a given volume or region being”; Column 13, Line 40-45, “two reductive iterations may be performed to replace the 10 splats with 4 splats and the 4 splats with 1 splats if the particular element is a threshold distance outside the field-of-view or remains outside the field-of-view for the entirety of the converged viewing path”).
Looper and Sosulnikov are analogous since both deal with processing and optimizing 3D data representations for efficient handling. Sosulnikov provides distribution-based entropy encoding for compressing data. Looper provides a method of assigning importance levels to 3D splat primitives and eliminating lower-importance splats to reduce data volume. 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 importance-based pruning taught by Looper into the modified invention of Sosulnikov and Supikov such that an importance level is assigned to each Gaussian and Gaussians with importance below a threshold are eliminated from the 3DGS model before compression. The motivation is to reduce the number of splats to be encoded while preserving quality for important regions, as Looper teaches that this produces “a splat 3D model with a reduced file size” (Column, Line 19-21).
Regarding Claim 6, the combination of Sosulnikov and Supikov teaches the invention in Claim 4.
The combination does not explicitly disclose but Looper teaches determining a mean-based value for each of the two or more Gaussians and eliminating, from the 3D Gaussian Splatting model, one or more of the two or more Gaussians with a corresponding mean-based value outside a set range (Looper, Column 13, Line 30-45, “a single reductive iteration may be performed to replace the 10 splats with 4 splats if the particular element is just outside the field-of-view or is in the field-of-view for one part of the converged viewing path, and two reductive iterations may be performed to replace the 10 splats with 4 splats and the 4 splats with 1 splats if the particular element is a threshold distance outside the field-of-view or remains outside the field-of-view for the entirety of the converged viewing path”).
Looper and Sosulnikov are analogous since both deal with processing 3D data representations for efficient handling. Sosulnikov provides distribution-based entropy encoding. Looper provides a method of pruning splats whose positions fall outside a defined spatial range. 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 position-based pruning taught by Looper into the modified invention of Sosulnikov and Supikov such that Gaussians with mean positions outside a set range are eliminated from the 3DGS model. The motivation is to remove Gaussians that are positioned outside the relevant viewing area, reducing the data to be encoded while maintaining visual quality for the important viewing regions.
Regarding Claim 7, the combination of Sosulnikov and Supikov teaches the invention in Claim 4.
The combination does not explicitly disclose but Looper teaches using rate control to adjust a percentage of pruning of the 3D Gaussian Splatting model (Looper, Column 5, Line 65-67-Column 6, Line 1-6, “the splat generation system may perform more splat generation iterations and define the loss function with a lower level of acceptable loss when generating the splats for regions or parts of the 3D model that are assigned a higher priority value, and may perform fewer splat generation iterations and define the loss function with a high level of acceptable loss when generating the splats for regions or parts of the 3D model that are a lower priority value”; Column 2, Line 18-22“The adaptive density control produces a splat 3D model with a reduced file size that is optimized for streaming across a data network and for faster rendering with less resources by a viewing device”; it is noted Looper’s adaptive density control that adjusts the number of splats (and thus the effective percentage of pruning) based on priority values and loss function thresholds reads on “rate control to adjust a percentage of pruning.”
Looper and Sosulnikov are analogous since both deal with controlling the amount of data to be processed or encoded. Sosulnikov provides entropy encoding with adjustable parameters. Looper provides adaptive density control that adjusts the degree of splat reduction. 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 adaptive density control taught by Looper into the modified invention of Sosulnikov and Supikov such that rate control is used to adjust the percentage of pruning of the 3DGS model. The motivation is to control the trade-off between the number of Gaussians retained and the resulting file size, enabling optimization for streaming and rendering efficiency
Regarding Claim 8, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination does not explicitly disclose but Looper teaches performing finetuning of the 3D Gaussian Splatting model (Looper, Column 2, Line 54-58, “the splat generation system retrains or refines <read on performing finetuning >the splat representations for the regions about the converged viewing path by generating more-and-more increasingly smaller sized and higher fidelity splats for the regions that are increasingly closer to the converged viewing path”; Column 9, Line 34-36, “Refining (at 512) the particular 3D model includes performing additional splat generation iterations to account for changes in the converged viewing path.”).
Looper and Sosulnikov are analogous since both deal with iteratively optimizing data representations. Sosulnikov provides iterative optimization of GMM parameters for encoding. Looper provides iterative refinement (finetuning) of splat representations to improve quality. 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 finetuning (refining) of the 3DGS model taught by Looper into the modified invention of Sosulnikov and Supikov such that finetuning is performed on the 3DGS model to improve representation quality before compression. The motivation is to refine the 3DGS model by performing additional optimization iterations to improve quality and detail, as discussed by Looper in refining at Step (512).
Regarding Claim 10, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches using rate control to adjust a percentage of finetuning of the 3D Gaussian Splatting model (Looper, Column 5, Line 65-67-Column 6, Line 1-6, “the splat generation system may perform more splat generation iterations and define the loss function with a lower level of acceptable loss when generating the splats for regions or parts of the 3D model that are assigned a higher priority value, and may perform fewer splat generation iterations and define the loss function with a high level of acceptable loss when generating the splats for regions or parts of the 3D model that are a lower priority value”; Column 8, Line 30-35 “training (at 508) the particular 3D model involves performing multiple splat generation iterations”).
As explained in rejection of claim 8, the obviousness for combining of Looper into Sosulnikov is provided above.
Claim(s) 12, 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sosulnikov et al. (US 20240340425 A1, hereinafter Sosulnikov) in view of Supikov et al. (US 20240355047 A1, hereinafter Supikov) as applied to Claim 1 above and further in view of Minnen et al. (US 20230419555 A1, hereinafter Minnen).
Regarding Claim 12, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches performing a GS optimization (Supikov, Paragraph [0019], “the 3DGS algorithm moves the initial gaussians towards the scene geometry by balancing densification and pruning”; [0021], “disclosed examples can solve the problem of slow training of a 3DGS representation by providing an initial distribution of gaussian splats that enables a 3DGS algorithm to converge quickly”).
The combination of Sosulnikov nor Supikov does not explicitly disclose but Minnen teaches performing an adaptive loss computation (Minnen, Paragraph [0062], “The compression system and the decompression system can be jointly trained using machine learning training techniques (e.g., stochastic gradient descent) to optimize a rate-distortion objective function. The training process can user error metric such as mean squared error (MSE), mean absolute error (MAD), and multiscale structural similarity (MS-SSIM) or any differentiable loss function. The training process can also use adversarial loss (the adversary part of a GAN) or a perceptual metric (i.e., a learned metric) loss.”).
Minnen and Sosulnikov are analogous since both deal with encoding data using distribution-based models where encoding parameters are optimized. Sosulnikov provides distribution-based entropy encoding for signal compression. Minnen provides a learned compression framework with adaptive loss computation using multiple loss functions for optimizing encoding quality. 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 adaptive loss computation taught by Minnen into the modified invention of Sosulnikov and Supikov such that a GS optimization is performed with an adaptive loss computation to optimize the 3DGS compression. The motivation is to improve the trade-off between reconstruction quality and compressed bitstream size by using adaptive loss functions during optimization, as discussed by Minnen in its Paragraph [0013] regarding rate-distortion objective optimization.
Regarding Claim 13, the combination of Sosulnikov and Supikov teaches the invention in Claim 1.
The combination further teaches wherein performing the GS optimization comprises performing a joint optimization of GS reconstruction quality [[ and compressed bitstream size ]] (Supikov, Paragraph [0021], “disclosed examples can solve the problem of slow training of a 3DGS representation by providing an initial distribution of gaussian splats that enables a 3DGS algorithm to converge quickly (e.g., reducing overall training time from ˜1 hour to 5-8 minutes for high quality representations”).
The combination of Sosulnikov nor Supikov does not explicitly disclose but Minnen teaches performing a joint optimization of GS reconstruction quality and compressed bitstream size (Minnen, Paragraph [0013], “In learning-based image compression, image codecs are developed by optimizing a computational model to minimize a rate-distortion objective.”; [0028], “The networks used by the compression system 100 are jointly trained (along with neural networks used by the decompression system) using a rate-distortion objective function.”; [0062], “the first encoder neural network 110, the second-encoder neural network 120, the multiple convolutional neural network blocks 225, 230, 315, 320, 325, 415, 420, 425 and the decoder neural network 155 can be jointly trained to optimize the rate distortion objective function.”).
Minnen and Sosulnikov are analogous since both deal with optimizing encoding parameters to achieve efficient data compression. Sosulnikov provides distribution-based entropy encoding. Minnen provides a framework for jointly optimizing reconstruction quality (distortion) and compressed data size (rate) through a rate-distortion objective function. 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 joint rate-distortion optimization framework taught by Minnen into the modified invention of Sosulnikov and Supikov such that a joint optimization of GS reconstruction quality and compressed bitstream size is performed. The motivation is to achieve an optimal trade-off between the quality of the reconstructed 3DGS model and the size of the compressed bitstream, as the rate-distortion objective function in Minnen is designed to minimize the compressed data size while maintaining reconstruction quality.
Allowable Subject Matter
Claims 3, 9, 18 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.
Regarding Claim 3, the prior art of record, specifically the prior art Sosulnikov et al. (US 20240340425 A1) teaches distribution-based entropy encoding and decoding of quantized parameters using Gaussian Mixture Model (GMM) distribution models, with distribution parameter estimation and signaling. The prior art Supikov et al. (US 20240355047 A1) teaches a 3D Gaussian Splatting model with opacity as one of its parameters, including setting a post-activation opacity parameter for each 3D Gaussian splat. However, none of the prior art cited alone or in combination provides motivation to teach determining the one or more 3D Gaussian Splatting parameters are opacity parameters, wherein the distribution model is a Poisson distribution. Therefore, Claim 3 is allowable over the prior art of record.
Regarding Claim 9, The prior art of record, specifically the prior art Sosulnikov et al. (US 20240340425 A1) teaches distribution-based entropy encoding with quantization of parameters. The prior art Supikov et al. (US 20240355047 A1) teaches a 3D Gaussian Splatting model with various parameters including opacity, spherical harmonics, and covariance. The prior art (US 12380647 B1) teaches refining and retraining splat representations including performing additional splat generation iterations. However, none of the prior art cited alone or in combination provides motivation to teach finetuning of the 3D Gaussian Splatting model comprises performing one of a parameter clip process, a normalization process, and a resetting process on the 3D Gaussian Splatting model. Therefore, Claim 9 is allowable over the prior art of record.
Regarding Claim 18, it recites limitations similar in scope to the limitations of Claim 3 and therefore is with the same reason for allowable specified above.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
From Chaos to Clarity 3DGS in the Dark - 20240612 - Li et al
US20250148678A1 Human subject gaussian splatting using machine learning
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 on (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
Prosecution Insights