DETAILED ACTION
This action is in response to Applicant’s response filed on 06/24/2026. The amendments to Claims 1-10 and 12-16, and the Specification, filed on 06/24/2026, are entered. Claims 1-20 are now pending in the present application. This Action is made FINAL.
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 .
Response to Arguments
Applicant’s amendments to the Specification and Claims have overcome each and every objection previously set forth in the Office Action Mailed on 04/22/2026.
Applicant’s arguments, see section IV on pages 10-13, filed on 06/24/2026, with respect to the rejection(s) of claim(s) 1-20 under 35 U.S.C. 103 have been fully considered but are moot because the new grounds of rejection do not solely rely on Bouhnik and Tran teaching or rendering obvious the amended independent claims. Specifically, Bouhnik is no longer relied on to teach or rendering obvious generating one or more images of one or more first objects based on pose-independent information about the one or more first objects and pose-dependent information about one or more second objects, in the context of the amended claim and as previously applied. However, other aspects of Bouhnik’s teachings are relied on in the new grounds of rejection and applied differently in the context of the amended claim, including Bouhnik’s teachings of transforming a first image such that a first object matches a pose of a second object (See Bouhnik col. 9, lines 42-44). See the Claim Rejections - 35 USC § 103 section below for further detail, including the newly identified prior art by Kicanaoglu et al. (US 20180336434 A1; provided by Applicant in IDS submitted on 03/13/2026).
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1, 8, and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1; provided by Applicant in IDS submitted on 03/13/2026) in view of Bouhnik (US 10,540,757 B1).
Regarding claim 1,
Kicanaoglu teaches: One or more processors, comprising:
one or more circuits to use one or more neural networks (Kicanaoglu teaches an artificial neural-network encoder-decoder architecture implemented using one or more processors/circuits; see Figs 1-2 and 12A-B, paragraphs [0086]-[0088] describing an “encoder-decoder artificial neural network,” and paragraphs [0111]-[0115] describing implementation by “a circuit… or processor” and other hardware/software components of the system.)
to generate one or more images of one or more first objects (Kicanaouglu teaches an orbit generator that generates images of an object at different poses. In particular, paragraph [0088] teaches identity units, concatenating pose units with the identity representation, and providing the resulting representation to decoder 1212, which “generates target images.” Kicanaouglu specifically states that the decoder “reconstructs the image with the corresponding pose” (paragraph [0080]); see also paragraphs [0079]-[0080], [0088]-[0092], and Figs 12A-B.)
(Kicanaouglu teaches a two-branch architecture in which “Each branch has an encoder-decoder architecture” (paragraph [0079]). Kicanaoglu further teaches that each encoder receives respective inputs 1204, 1206, such as an image, and produces an abstract representation split into identity units and pose units (paragraph [0087]). Accordingly, the encoder of the first branch receiving first input 1204 corresponds to the claimed first encoder, and its identity units correspond to the claimed first information about the first object; see also claims 13-15.)
and one or more second encoders to encode second information about the one or more second objects (Kinaoglu teaches that each of the two branches includes an encoder receiving a respective input image, such that the encoder of the second branch receives second input 1206 and generates a representation including identity and pose units, and paragraph [0087] further teaches swapping the pose information encoded from first input 1204 (see paragraphs [0079], [0087]-[0088], FIG 12A, and claims 13-15). Thus, the encoder processing second input 1206 reasonably corresponds to the claimed second encoder, and the pose information encoded from second input 1206 reasonably corresponds to the claimed second information.),
wherein the first information is independent of one or more first poses of the one or more first objects (Kicanaoglu teaches that the encoder output is separated into identity units and pose units and describes the identity representations as “constant over time” (see paragraphs [0085] and [0087]). Because the encoder of the first branch receives the first input 1204 and encodes identity information that is maintained while the pose information is varied (paragraph [0088]), the identity information encoded from first input 1204 reasonably corresponds to the claimed first information independent of the one or more first poses (refer back to paragraphs [0079], [0085], [0087]-[0088], and Figs. 12A-B.),
and wherein the second information is dependent on the one or more second poses (Kicanaoglu teaches that the encoder output is separated into identity units and pose units, with the pose units being “relevant to the pose of the instance” (paragraph [0087]). Because the encoder of the second branch receives second input 1206 and encodes pose units corresponding to the pose represented by that second input, the pose information encoded from second input 1206 reasonably corresponds to the claimed second information that is dependent on the one or more second poses (refer back to paragraphs [0079], [0087]-[0088]j, and Fig. 12A).).
Kicanaoglu does not explicitly disclose generating the first object in the second pose exhibited by the second object. Rather, Kicanaoglu uses different object instances having the same pose to disentangle pose from identity information, explaining that the network encodes the “common pose information shared by the two images” into pose representation (paragraph [0081]), also see paragraphs [0084]-[0085], [0089]-[0090]).
In a related art, Bouhnik teaches the missing relationship. Bouhnik expressly teaches the system “receives and/or generates first and a second image, respectively” (Bouhnik Fig. 3, step 310, and column 5 lines 29-31), and explains that the pose data represents the orientation of the subject’s head, arms, torso, legs, and other body portions (Fig. 4A and col. 5 lines 45-48). Bouhnik further teaches pose alignment processing between the two subjects and expressly provides that “the first image can be transformed to match the pose of the second subject” (col. 9 lines 33-50; also see Figs. 5A-5C). Thus, Bouhnik teaches using pose information obtained from a second object image to transform a first object into the pose of the second object.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Kicanaoglu such that the pose information encoded from the second-object image is used to generate the first object in the pose of the second object, as taught by Bouhnik. Kicanaoglu expressly teaches “sharing the pose information across different object instances, which improves pose recognition and also improves the purity of identity information” (Kicanaoglu paragraph [0088]), while Bouhnik teaches transforming a first subject to “match the pose of the second subject” and explains that improved pose matching leads to an “improved mixed image” (Bouhnik col. 9 lines 34-50). Thus, both reference concern using pose information associated with different object instances to improve image generation or alignment. One of ordinary skill therefore would have had reason to use Bouhnik’s direct pose-matching technique with Kicanaoglu’s identity/pose encoder-decoder architecture to permit direct transfer of the second object’s observed pose to the first object, thereby improving pose alignment while retaining the first object’s identity information.
Regarding claim 8, Kicanaoglu and Bouhnik teach a system (see Kicanaoglu Figs 1-2, 12A-B, and paragraphs [0002] and [0035]; see Bouhnik, column 1, lines 17-20, FIG. 1 “a system diagram of system for generating combined images utilizing image processing of multiple-images according to examples of the present disclosure.”). The limitations found in claim 8 equally mirror the limitations established in claim 1. Claim 8 is similarly rejected based on the prior art taught by Kicanaoglu and Bouhnik in claim 1.
Regarding claim 15, Kicanaoglu and Bouhnik teach a method (see Kicanaoglu Abstract and paragraph [0009]; see Bouhnik, Abstract). The limitations found in claim 15 equally mirror the limitations established in claim 1. Claim 15 is similarly rejected based on the prior art taught by Kicanaoglu and Bouhnik in claim 1.
Claim(s) 2 and 9 are rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1) in view of Bouhnik (US 10,540,757 B1), and in further view of Tran (“Disentangled Representation Learning GAN for Pose-Invariant Face Recognition”; provided by applicant in IDs; examiner provided a copy of the version used by examiner in the Office Action mailed on 04/22/2026, with corresponding citations mentioned in the present office action).
Regarding claim 2, Kicanaoglu and Bouhnik teach the one or more processors of claim 1, including a first encoder to encode first information about the one or more first objects.
Kicanaoglu and Bouhnik fail to explicitly teach: wherein the one or more first encoders are to encode the first information at least by generating a class-specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects.
In a related art, Tran teaches: generating a class-specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects (Tran p. 1285, right column, last paragraph through p. 1286, left column, first and second paragraphs, discloses an encoder and decoder aimed to learn an identity representation from a face image (i.e. target image of one or more objects), including through the use of identity classification. Specifically, Tran teaches learning “pose-invariant identity representation for PIFR” (p. 1285, right column, third paragraph), a well-known technique based on pose-independent information. In particular, tran teaches that DR-GAN can take “one or multiple images as the input” and generate “one unified representation,” and that the learned identity representation is “exclusive or invariant to pose” (Abstract). Tran further teaches that, when multiple images are available, the encoder produces representations that are combined into a single identity representation (p. 1284 left column second paragraph). Tran’s identity representation reasonably corresponds to the claimed “class-specific representation” because it represents the particular subject/identity depicted by the target image(s), while being disentangled from pose.).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to further modify the one or more processors of Kicanaoglu as modified by Bouhnik to generate the pose-independent information from one or more target images using Tran’s unified identity representation technique. Kicanaoglu and Tran both aim to separate identity information from pose information, and Tran explains that using multiple images provides additional constraints for learning the encoder and result in a better identity representation (Tran p. 1288 left column first paragraph of 4.2). One of ordinary skill therefore would have had reason to use Tran’s technique in the combined processor to obtain a more robust pose-independent representation of the first object form the available target image(s).
Regarding claim 9, Kicanaoglu and Bouhnik teach the system of claim 9, including a first encoder to encode first information about the one or more first objects.
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more first encoders are to encode the first information at least by generating a class- specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects, and wherein the one or more processors are to generate the one or more images based, at least in part, on the class-specific representation.
In a related art, Tran teaches: generating a class-specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects (These limitations equally mirror the limitations found in claim 2. For sake of brevity, refer to claim 2’s 103 rejection, found above, for Tran’s teachings of the limitation.), and wherein the one or more processors are to generate the one or more images based, at least in part, on the class-specific representation (Tran teaches generating a new image based on the class-specific identity information from input images (“3.3 Multi-Image DR-GAN” found on p. 1286, right column through p. 1287 left column). Examiner interprets this to be equivalent to generating the one or more images based at least in part, on the class-specific representation.).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the system of Kicanaoglu as modified by Bouhnik to generate a unified pose-independent identity representation from one or more target images and to generate the output image based on that representation, as taught by Tran. Kicanaoglu and Tran both use encoder -decoder architectures to separate identity from pose and generate pose-varied images while preserving identity. Tran further teaches that its learned representation is both discriminative and generative and is used directly for identity preserving image synthesis (see Tran FIG. 1 and p. 1289 left column Conclusions section). Therefore, a person of ordinary skill in the art would have had reason to incorporate Tran’s representation technique into the combined system to improve the representation of the first object identity and preserve that identity in the generated image.
Claim(s) 3 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1), in view of Bouhnik (US 10,540,757 B1), and in further view of Huang (“Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization”; provided by applicant in IDs; examiner provided a copy of the version used by examiner with the Office Action mailed on 04/22/2026, with corresponding citations mentioned in the present office action).
Regarding claim 3, Kicanaoglu and Bouhnik teach the one or more processors of claim 1, including one or more circuits generating one or more images based, at least in part, on one or more first encoders to encode first information.
Bouhnik further teaches using vectors for an “image discriminator” that determines whether or not an image should be displayed through the use of a trained model (Bouhnik, column 12, lines 12-31), and Kicanaoglu further teaches transforming concatenated vectors into a synthetic image (see paragraphs [0088]-[0090], and FIGs 12A-12B).
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more circuits are to generate the one or more images based, at least in part, on a set of mean and variance vectors generated according to the first information.
In a related art, Huang teaches: using adaptive instance normalization layer for the purpose of decoding feature representation in order to reconstruct images, including the use of mean and variance (Abstract; Section 1, paragraph 3, lines 6-10; Section 3.2 lines 1-6; Section 5.1 paragraph 1, lines 4-7).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teaching of Huang in order to improve image reconstruction by achieving “speed comparable to the fastest existing approach, without the restriction to a pre-defined set of style features” (Huang, Abstract). All inventions lie in the field of image analysis with an aim at reconstructing images.
Regarding claim 10, Kicanaoglu and Bouhnik teach the system of claim 8, including one or more processors to generate one or more images of one or more first objects.
Kicanaoglu further teaches: decoding pose-independent information to generate one or more images (see paragraphs [0087]-[0088]).
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more processors are to generate the one or more images based, at least in part, on one or more decoders to decode the first information as a set of mean and variance vectors.
In a related art, Huang teaches: using adaptive instance normalization layer for the purpose of decoding feature representation in order to reconstruct images, including the use of mean and variance (Abstract; Section 1, paragraph 3, lines 6-10; Section 3.2 lines 1-6; Section 5.1 paragraph 1, lines 4-7), while specifically citing the use of a decoder network to generate the final stylized image (Section 1, paragraph 3, third to last sentence).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teaching of Huang in order to improve image reconstruction by achieving “speed comparable to the fastest existing approach, without the restriction to a pre-defined set of style features” (Huang, Abstract). All inventions lie in the field of image analysis with an aim at reconstructing images.
Claim(s) 4-7, 12-14, and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1), in view of Bouhnik (US 10,540,757 B1), and in further view of Kim (US 20180314716 A1).
Regarding claim 4, Kicanaoglu and Bouhnik teach the one or more processors of claim 1, including a first encoder and second encoders used in the generation of images.
Bouhnik further teaches: wherein the one or more images are of a target class (see FIG. 2, “first human/first subject” and “second human/ second person/ second subject” (column 4, lines 10-25). Examiner interprets human to be the target class.), and wherein the one or more circuits are to generate the one or more images based, at least in part, on a class-specific representation, (see Abstract; see column 4, line 4 to column 5, line 5 and FIG. 2 that shows generated image 223 based on a first person’s (from image 203) non-pose dependent clothing (e.g. top) information and a second person’s (from image 211) pose-dependent leg pose information. The first person and second person used to create the generated image of a person are interpreted as class-specific and class-invariant representations.).
Kicanaoglu and Bouhnik fail to explicitly disclose: using a class-invariant representation different from the target class for generating the one or more images.
In a related art, Kim teaches: a method for learning cross domain relations using a model for direct transfer of attributes between images of different classes/domains (see paragraphs [0011]-[0016], [0060]).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques with more image inputs of different domains/classes. Doing so would make the system more robust by allowing for a variety of objects and their corresponding images to be input and processed. All three references are image analysis based with a focus on reconstructing images.
Regarding claim 5, Kicanaoglu and Bouhnik teach the one or more processors of claim 1, including the use of one or more neural networks.
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more circuits are to train the one or more neural networks based, at least in part, on one or more second images of different objects from the one or more first objects.
In a related art, Kim teaches: wherein the one or more circuits are to train the one or more neural networks based, at least in part, on one or more second images of different objects from the one or more first objects (see paragraphs [0055]-[0057], training the system using a set of images from domain A (i.e. first objects) and a second set of images representing domain B (i.e. second objects that are different objects than the domain A objects).).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques, thereby improving the accuracy of the model for reconstructing image(s).
Regarding claims 6, 13, and 19, Kicanaoglu and Bouhnik teach the one or more processors of claim 1, the system of claim 8, and the method of claim 15. Examiner notes, claims 6 and 19 of the instant application mirror or have broader scopes than claim 13, while all claim techniques of training the one or more neural networks in an unsupervised manner, therefore the rejection for claim 13 applies equally to claims 6 and 19.
Bouhnik further teaches: “Various techniques 20 may be used to train the models including backpropagation, statistical learning, supervised learning, semi-supervised learning, stochastic learning, or other known techniques” (column 19, lines 20-23), without explicitly citing unsupervised training.
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more processors are further to cause the one or more neural networks to be trained in an unsupervised manner based, at least in part, on a training data set comprising images of different classes (as stated in claim 13 of the instant office action).
In a related art, Kim teaches a method for learning cross domain relations using a model for direct transfer of attributes between images of different classes/domains (see paragraphs [0011]-[0016], [0060]). Kim further teaches the method utilizes a GAN to perform unsupervised training of a first generative model for translating images in domain A (i.e. class A) to images in domain B (i.e. class B) (see paragraphs [0015]-[0016]) and training the system using a set of images from domain A and a set of images representing domain B (see paragraphs [0055]-[0057]).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques, thereby improving the accuracy of the model for reconstructing image(s).
Regarding claim 7 and 20, Kicanaoglu and Bouhnik teach: the one or more processors of claim 1 and the method of claim 15.
Bouhnik teaches training one or more neural networks and translating images (Bouhnik column 3, lines 1-11).
However, Kicanaoglu and Bouhnik fail to explicitly disclose: training the one or more neural networks to translate images between two randomly sampled source classes (claim 20: sampled source object classes).
In a related art, Kim teaches: training the one or more neural networks to translate images between two randomly sampled source classes (Paragraphs [0011]-[0016] and [0060] disclose training to translate between two different domains using a GAN. Kim further discloses standard GAN models take random Gaussian noise, while the model taught by Kim uses images instead of noise.).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques, thereby improving the accuracy of the processor.
Regarding claim 12, Kicanaoglu and Bouhnik teach the system of claim 8.
Kicanaoglu and Bouhnik fail to disclose: wherein the one or more processors are to train the one or more neural networks based, at least in part, on one or more second images of the one or more second objects, and wherein the one or more second objects are one or more objects other than the one or more first objects.
In a related art, Kim teaches: wherein the one or more processors are to train the one or more neural networks based, at least in part, on one or more second images of the one or more second objects, and wherein the one or more second objects are one or more objects other than the one or more first objects (see paragraphs [0055]-[0057], training the system using a set of images from domain A (i.e. first objects) and a second set of images representing domain B (i.e. second objects that are different objects than the domain A objects).).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques, thereby improving the accuracy of the model for reconstructing image(s).
Regarding claim 14, Kicanaoglu and Bouhnik teach the system of claim 8. The limitations found in claim 14 equally mirror the limitations found in claim 7, with the exception of translating pose images. Therefore, claim 14 is rejected based on the prior art of Kicanaoglu, Bouhnik, and Kim taught in claim 7 of the present office action. The additional use of pose images fails to overcome the rejection because Bouhnik teaches using pose dependent information, as exemplified in claim 8’s 103 rejection.
Regarding claim 18, Kicanaoglu and Bouhnik teach the method of claim 15.
Kicanaoglu and Bouhnik fail to explicitly disclose: further comprising training the one or more neural networks based, at least in part, on one or more second images of a different object class from an object class of the one or more generated images.
In a related art, Kim teaches: further comprising training the one or more neural networks based, at least in part, on one or more second images of a different object class from an object class of the one or more generated images (see paragraphs [0055]-[0059], training the system using a set of images from domain A (i.e. first objects) and a second set of images representing domain B (i.e. second objects that are different objects than the domain A objects) and FIG. 8A-8C representing a learning model trained with various handbag and shoe images (FIG. 8A) and generated output of shoe images (FIG. 8B) based on shoe images and second images of different objects, specifically exemplified by handbags.).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik to incorporate the teachings of Kim in order to improve relation extraction techniques, thereby improving the accuracy of the model for reconstructing image(s).
Claim(s) 11 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1), in view of Bouhnik (US 10,540,757 B1), in further view of Tran (“Disentangled Representation Learning GAN for Pose-Invariant Face Recognition”), in further view of Kim (US 20180314716 A1).
Regarding claim 11, Kicanaoglu and Bouhnik teach the system of claim 8.
The limitations found in claim 11 equally mirror the limitations found in claim 4 and taught by Kicanaoglu, Bouhnik, and Kim, with the exception of the use of latent representation. For the sake of brevity, please refer to claim 4’s 103 rejection for Kicanaoglu, Bouhnik, and Kim’s teachings of: wherein the one or more images are of a target class, and
wherein the one or more processors are to generate the one or more images based, at least in part, on a class-specific
In a related art, Tran teaches: a relevant GAN variant, known as an “Adversarial Autoencoder (AAE)” that uses an autoencoder to reconstruct the input image and a latent vector is generated by the encoder to be used in subsequent tasks like training the discriminator model (p. 1287, seciont 3.4, subsection “AAdversarial Autoencoder (AAE)”). Tran further specifies the autoencoder is trained to learn a latent representation (p. 1287, seciont 3.4, subsection “AAdversarial Autoencoder (AAE)”).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the teachings of Kicanaoglu and Bouhnik, previously modified by Kim (as seen in claim 4), to incorporate the teachings of a variant GAN, taught by Tran, that uses latent representation information in order to help train the model to distinguish between real/fake distributions (Tran (p. 1287, seciont 3.4, subsection “AAdversarial Autoencoder (AAE)”), thereby making the system more accurate.
Regarding claim 17, Kicanaoglu and Bouhnik teach the method of claim 15, including generating one or more images and using one or more first and second encoders, but fail to explicitly disclose generating the images based on a class-specific latent representation of a target object class and a class-invariant latent representation of a different source object class.
In a related art, Tran teaches: a pose-invariant representation identity representation that is explicitly disentangled from pose and is used for image synthesis. Tran explains that its encoder-decoder architecture learns an identity representation that is “explicitly disentangled” from pose and can synthesize identity-preserving faces at target poses (see Tran Abstract, Fig. 1, p. 1283-1284 section 1). Thus, Tran’s identity representation reasonably corresponds to the claimed class-specific latent representation, while the separately represented pose information reasonably corresponds to class-invariant information.
In a related art, Kim teaches: learning relations between “two different domains” (paragraph [0028]) and translating images from domain A to domain B; for example, Kim explains generating a shoe image from a handbag image and teaches that its cross-domain translation maintains key attributes such “orientations” and “facial identities” (see paragraphs [0011]-[0016], [0054]-[0058], [0060], and FIG. 8A-8C representing a learning model trained with various handbag and shoe images (FIG. 8A) and generated output of shoe images (FIG. 8B) based on shoe images and second images of different objects, specifically exemplified by handbags.).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the method of Kicanaoglu and Bouhnik to use Tran’s disentangled identity and pose representations because Kicanaoglu and Tran both separate identity from pose for identity preserving image generation, and Tran teaches that such disentanglement supports pose-invariant representation and image synthesis.
It would have further been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the method of Kicanaoglu as modified by Bouhnik and Tran to apply those representations across different source and target object classes, as taught by Kim. Kim teaches cross-domain image translation between different image domains while maintaining attributes such as orientation and identity. A person of ordinary skill would have had reason to use Kim’s technique to preserve the target-object identity while transferring pose-related information from the source image.
Claim(s) 16 is rejected under 35 U.S.C. 103 as being unpatentable over Kicanaoglu et al. (US 20180336434 A1), in view of Bouhnik (US 10,540,757 B1), in further view of Tran (“Disentangled Representation Learning GAN for Pose-Invariant Face Recognition”), and in further view of Huang (“Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization”).
Regarding claim 16, Kicanaoglu and Bouhnik teach the method of claim 15.
Kicanaoglu and Bouhnik fail to explicitly disclose: wherein the one or more first encoders are to encode the first information at least by generating a class specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects, and wherein generating the one or more images is further based, at least in part, on one or more decoders to decode the class-specific representation as a set of spatially invariant means and variances.
In a related art, Tran further teaches: wherein the one or more first encoders are to encode pose-independent information at least by generating a class-specific representation of the one or more first objects based, at least in part, on one or more target images of the one or more first objects (Tran p. 1285, right column, last paragraph through p. 1286, left column, first and second paragraphs, discloses an encoder and decoder aimed to learn an identity representation from a face image (i.e. target image of one or more objects), including through the use of identity classification. Specifically, Tran teaches learning “pose-invariant identity representation for PIFR” (p. 1285, right column, third paragraph), a well-known technique based on pose-independent information.). Tran also teaches using a decoder throughout the process of synthesizing images (Abstract; p. 1283, right column second paragraph, lines 7-12; p. 1285 section 3.2 lines 5-7).
However, Tran fails to explicitly disclose: wherein generating the one or more images is further based, at least in part, on one or more decoders to decode the class-specific representation as a set of spatially invariant means and variances.
In another related art, Huang teaches: using adaptive instance normalization layer for the purpose of decoding feature representation in order to reconstruct images, including the use of mean and variance computed across spatial dimensions independently for each channel, (Abstract; Section 1, paragraph 3, lines 6-10; Section 3.2 lines 1-10; Section 5.1 paragraph 1, lines 4-7), while specifically citing the use of a decoder network to generate the final stylized image (Section 1, paragraph 3, third to last sentence).
It would have been obvious to a person of ordinary skill in the art before the effective filing date to have modified the method of Kicanaoglu and Bouhnik to generate the first pose independent information as a class-specific representation based on target images, as taught Tran, in order to provide a more robust representation of the first-object identity.
It would have further been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the method of Kicanaoglu modified by Bouhnik and Tran to use Huang’s adaptive instance-normalization technique to represent the class-specific information using spatially invariant mean and variance statistics for image generation in order to efficiently control the appearance of the generated image using the class-specific representation. Huang teaches its method achieves “speed comparable to the fastest existing approach, without the restriction to a pre-defined set of style features” (Huang, Abstract). All inventions lie in the field of image analysis with an aim at reconstructing images.
Conclusion
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 SAMUEL DAVID BAYNES whose telephone number is (571)272-0607. The examiner can normally be reached Monday - Friday 8:00 am - 5:00 pm.
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, Stephen R Koziol can be reached at (408)918-7630. 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.
/SDB/
Samuel D. Baynes
Examiner, Art Unit 2665
/Stephen R Koziol/Supervisory Patent Examiner, Art Unit 2665
