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 .
Status of Claims
Claims 1, 5-6, 8-11, 18-19, 34-42 and 47-49 are amended. Claims 50 and 51 are new.
Drawings
The drawings were received on 08/12/2026. These drawings are acceptable.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(d):
(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
The following is a quotation of pre-AIA 35 U.S.C. 112, fourth paragraph:
Subject to the following paragraph [i.e., the fifth paragraph of pre-AIA 35 U.S.C. 112], a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
Claims 48-49 are rejected under 35 U.S.C. 112(d) or pre-AIA 35 U.S.C. 112, 4th paragraph, as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends, or for failing to include all the limitations of the claim upon which it depends. Claim 48 is dependent on claim 49 and claim 49 is also dependent on claim 49. Claim 48 recites “The method of claim 49...” thus referring to a subsequently set forth claim rather than a claim previously set forth. Claim 49 recites “The method of claim 49...” thus improperly referring to itself. Neither claim establishes proper dependency from a previously set froth claim.
Applicant may cancel the claim(s), amend the claim(s) to place the claim(s) in proper dependent form, rewrite the claim(s) in independent form, or present a sufficient showing that the dependent claim(s) complies with the statutory requirements.
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, 6-9, 29, 36-39 and 50-51 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR (Dundar, A., Gao, J., Tao, A., & Catanzaro, B. (2022). Fine detailed texture learning for 3d meshes with generative models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12), 14563-14574. (Year: 2022) “Dundar”) in view of ALWALA (Alwala, K. V., Gupta, A., & Tulsiani, S. (2022, June). Pretrain, self-train, distill: A simple recipe for supersizing 3D reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 3763-3772). IEEE. (Year: 2022) “Alwala”) and in further view of BHATTAD (Bhattad, A., Dundar, A., Liu, G., Tao, A., & Catanzaro, B., "View generalization for single image textured 3d models." 2021, In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 6081-6090). (Year: 2021) “Bhattad”).
Regarding claim 1, Dundar teaches “A method, comprising:
at a device:” (a method; Abstract); (an encoder-decoder pipeline; Pg.1, Col 1 Para 2);
“accessing a two-dimensional (2D) dataset comprised of single-view images labeled with texture information;” (For single-view datasets, we experiment on datasets that have camera view annotations. .... the second stage to learn high quality texture maps. Our improvements on learning high quality texture maps also apply to single view image datasets; Pg.3, Col 3);
“accessing a three-dimensional (3D) dataset comprised of multi-view images labeled with geometry information; and” (Multi-view image datasets provide a solution for the limited supervision problem of single-view image datasets; Pg.3 Col 2.3); (We again have the first stage where we focus on learning accurate geometry by using the ground-truth camera parameters; Pg.3, Col 3);
“to reconstruct samples in the 3D dataset, and” (we feed an input image to the encoder and ask the framework to reconstruct the image from a different view; Pg.3, Col 3.1);
“performing a second training step that includes training the machine learning model to reconstruct samples in the 2D dataset and” (In our trainings with single view image datasets, we directly optimize the model to reconstruct the input view; Pg.3, Col 3.1);
However, while Dundar fails to teach “jointly using the 2D dataset and the 3D dataset to train a machine learning model to generate a 3D content given an input 2D image by: performing a first training step that includes training the machine learning model”.
Alwala teaches “jointly using the 2D dataset and the 3D dataset to train a machine learning model to generate a 3D content given an input 2D image by:” (We first pre-train a reconstruction model using multi-view renderings of synthetic data. We then self-train category-specific models on diverse image collections; Pg.2, Fig.2 Desc); (to train a single model to reconstruct objects across hundreds of diverse object categories; Pg.2, Col 1 Para 2);
performing a first training step that includes training the machine learning model” (We first pre-train a reconstruction model using multi-view renderings of synthetic data; Pg.2, Fig.2 Desc);
However, while Dundar and Alwala fail to teach “to generate novel views from the samples in the 2D dataset”.
Bhattad teaches “to generate novel views from the samples in the 2D dataset.” (Given a single 2D image input, .... which can then be used to render novel views; Pg.6081, Fig.1 Desc);
Dundar discloses the framework learning from both multi view and single view images. As well as reconstruction of the datasets. Alwala disclose training steps for the reconstruction framework. Bhattad supplies extending single view reconstruction training to novel views. It would be obvious to an ordinary person skilled in the art to modify Dundar’s reconstruction and view generalization teachings with Alwala’s multi view pretraining and single view self-training reconstruction framework and supplies Bhattad’s teachings such that the model is first trained to reconstruct samples represented by the multi view dataset and then trained on single view data to reconstruct the input view and produce novel view. There would be reason to combine these training techniques to retain geometric consistency learned from multi view data while also improving texture and view generalization from single view data thus improving single image 3D reconstruction.
Dundar, Alwala and Bhattad are analogous art as they are related to training machine learning and 3D reconstruction from input image.
The motivation for the above is to have accurate and efficient training model that produces accurate 3D representation.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar by jointly using the 2D dataset and the 3D dataset to train a machine learning model to generate a 3D content given an input 2D image by: performing a first training step that includes training the machine learning model as taught by Alwala and by to generate novel views from the samples in the 2D dataset as taught by Bhattad.
Regarding claim 6, while Dunbar and Bhattad fail to teach the limitation of claim 6, Alwala further teaches “The method of claim 1, wherein an encoder is shared for the first training step and the second training step.” (All our networks share the same architecture. The encoder is a ResNet-34 [11] with an additional residual linear layer that produces z ∈ R2560; Pg.6, Col 4.2);
Alwala discloses shared architecture, and they disclose the encoder used which is a part of the shared architecture.
The motivation for the above is to have efficient training of machine models.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Bhattad by wherein an encoder is shared for the first training step and the second training step as taught by Alwala.
Regarding claim 7, while Dundar and Bhattad fail to teach the limitation of claim 7, Bhattad teaches “The method of claim 6, wherein:
during the first training step of the machine learning model using the 3D dataset the encoder generates a first 3D representation of an input image from the 3D dataset, and” (Given an input image I depicting a segmented object, our network predicts an implicit 3D representation; Pg.4, Col 3.1);
“during the second training step the encoder generates a second 3D representation of an input image from the 2D dataset.” (Given an input image I depicting a segmented object, our network predicts an implicit 3D representation; Pg.4, Col 3.1);
Alwala discloses 3D representation that is encoded. This occurs regardless of which train source/data is supplied during the input. The pretraining and image is encoded into the 3D representation and during the second training a single view image produces a 3D representation.
The motivation for the above is to have accurate 3D generation from efficient training of machine model.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Bhattad by during the first training step of the machine learning model using the 3D dataset the encoder generates a first 3D representation of an input image from the 3D dataset, and during the second training step the encoder generates a second 3D representation of an input image from the 2D dataset as taught by Alwala.
Regarding claim 8, while Dundar and Bhattad fail to teach the limitation of claim 8, Alwala teaches “The method of claim 7, wherein a renderer is shared for the first training step and the second training step.” (All our networks share the same architecture; Pg.6, Col 4.2);
Alwala discloses shared architecture, and they disclose the encoder used which is a part of the shared architecture.
The motivation for the above is to an efficient renderer used in the 3D generation.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Bhattad by wherein a renderer is shared for the first training step and the second training step as taught by Alwala.
Regarding claim 9, modified with Dundar and Bhattad, Alwala teaches “The method of claim 8, wherein:
during the first training step the renderer converts the first 3D representation of the input image from the 3D dataset into a first set of 2D renderings using viewpoints from the 3D dataset, and” (As we have access to textured 3D meshes, we are able to render multiple views for each object with known camera viewpoints... when volumetrically rendered from the available viewpoints, match the known color and masks at each pixel; Pg.4, Col 3.1);
Alwala discloses input image which gets the 3D representation and rendering multiple 2D views at view points from the multi view data supplied.
Bhattad further teaches “during the second training step the renderer converts the second 3D representation of the input image from the 2D dataset into a second set of 2D renderings that includes a reconstructed view of the input image from the 2D dataset and a novel view of the input image from the 2D dataset.” (Given a single 2D image input, we infer high quality textured 3D models– a triangular mesh and a texture map, which can then be used to render novel views. .... when rendered from that view; Pg.6081, Fig.1 Desc);
Bhattad discloses rendering the predicted 3D object at the source viewpoint to then reconstruct the input and render the same 3D object from another camera top produce novel views.
Claim 36 is directed to a system and its limitations are similar in scope and functions performed by the method of claim 6. Therefore, claim 36 limitations are also rejected with the same rationale as regarding claim 6.
Regarding claim 37, modified with Dundar Bhattad and Chen, Alwala further teaches “The system of claim 36, wherein:
during the first training step the renderer converts the first 3D representation of the input image from the 3D dataset into a first set of 2D renderings using viewpoints from the 3D dataset, and” (As we have access to textured 3D meshes, we are able to render multiple views for each object with known camera viewpoints... when volumetrically rendered from the available viewpoints, match the known color and masks at each pixel; Pg.4, Col 3.1);
Alwala discloses input image which gets the 3D representation and rendering multiple 2D views at view points from the multi view data supplied.
Bhattad further teaches “during the second training step the renderer converts the second 3D representation of the input image from the 2D dataset into a second set of 2D renderings that includes a reconstructed view of the input image from the 2D dataset and a novel view of the input image from the 2D dataset” (Given a single 2D image input, we infer high quality textured 3D models– a triangular mesh and a texture map, which can then be used to render novel views. .... when rendered from that view; Pg.6081, Fig.1 Desc);
Bhattad discloses rendering the predicted 3D object at the source viewpoint to then reconstruct the input and render the same 3D object from another camera top produce novel views.
Claim 38 is directed to a system and its limitations are similar in scope and functions performed by the method of claim 8. Therefore, claim 38 limitations are also rejected with the same rationale as regarding claim 8.
Claim 39 is directed to a system and its limitations are similar in scope and functions performed by the method of claim 9. Therefore, claim 39 limitations are also rejected with the same rationale as regarding claim 9.
Regarding claim 50, while Dundar and Bhattad fail to teach the limitations, Alwala teaches “The method of claim 1, wherein the first training step includes:
generating a 3D representation of a sampled input from the 3D dataset,” (We pretrain our implicit reconstruction network using a dataset of synthetic 3D shapes... Given an input image I depicting a segmented object, our network predicts an implicit 3D representation; Pg.4, Col 3.1);
“generating a set of 2D renderings from the first 3D representation,” (we are able to render multiple views for each object with known camera viewpoints... the single-view (implicit) 3D predictions, when volumetrically rendered from the available viewpoints; Pg.4, Col 3.1);
“computing a loss between the set of 2D renderings and the sampled input, and” (we supervise the reconstruction network via a novel-view rendering loss. ... we volume-render the predicted mask and color using the implicit representation and penalize deviations from the ground-truth; Pg.4, Fig.4 Desc);
“updating the machine learning model based on the loss.” (train our reconstruction network by simply enforcing that the single-view (implicit) 3D predictions, when volumetrically rendered... match the known color and masks at each pixel; Pg.4, Col 3.1);
Alwala discloses that its first training step comprises pretraining an image implicit reconstruction network using texture 3D models. Also, given an input image, the reconstruction network predicts an implicit 3D representation. Alwala also discloses volumetrically rendering the predicted 3D representation from viewpoints and supervising the reconstruction network using novel view rendering loss that penalizes deviations fort eh render mask and color. Thus, teaching 3D representation from an input of the 3D training dataset, generating a 2D renderings from the 3D representation, computing a loss between the rend predictions and updating the reconstruction network by training the network based on that loss.
The motivation for the above is to have accurate generation of 3D representation based on training data and losses.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Bhattad by generating a 3D representation of a sampled input from the 3D dataset, generating a set of 2D renderings from the first 3D representation, computing a loss between the set of 2D renderings and the sampled input, and updating the machine learning model based on the loss as taught by Alwala.
Regarding claim 51, while Dundar and Alwala fail to teach the limitations, Bhattad teaches “The method of claim 1, wherein the second training step includes:
generating a 3D representation of a sampled input from the 2D dataset,” (Given an image X1 and its corresponding camera pose C1, we first infer this image into its geometric and texture representations; Pg.6085, Col 3.2.1 Para 2);
“generating a set of 2D renderings from the 3D representation, the set of 2D renderings including a reconstructed view of the sampled input and a novel view of the sampled input,” (R is a differentiable renderer function that generates the image... we first infer this image into its geometric and texture representations, and render an intermediate image from a novel camera view.... a cycle consistency check between X1 and X′ 1 using the reconstruction loss (Eq 1) and perceptual loss; Pg.6085, Col 3.2.1 Para 2);
“computing a first loss between the reconstructed view and the sampled input,” (a cycle consistency check between X1 and X′ 1 using the reconstruction loss (Eq 1) and perceptual loss; Pg.6085, Col 3.2.1 Para 2);
“computing a second loss between the novel view and the sampled input, and” (We use GAN losses on rendered images from novels views; Pg.6085, Col 3.2 Para 2); (For fake images, we use the novel view images, I1; Pg.6085, Col 3.2.1 Para 2);
“updating the machine learning model based on the first loss and the second loss.” (We apply a cycle consistency check between X1 and X′ 1 using the reconstruction loss (Eq 1) and perceptual loss (Eq 2). ... significantly improves our overall 3D synthesis; Pg.6085, Col 3.2.1 Para 2);
Bhattad discloses single image training where an input image is inferred into geometric and texture representations, and a differentiable render generates images form the inferred representation. The renders an indeterminate image from a novel viewpoint and corresponding to the original viewpoint. The computes the reconstruction and perceptual losses between the original view and original input image. It would be obvious to a person skilled in the art to use the sample input image as the reference associated with both the reconstructed view and generated novel view output during the single vie second training step, thus computing losses associated with the reconstructed and novel vies and updating the model according to both losses.
The motivation for the above is to have accurate generation of 3D representation based on training data and losses.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Alwala by generating a 3D representation of a sampled input from the 2D dataset, generating a set of 2D renderings from the 3D representation, the set of 2D renderings including a reconstructed view of the sampled input and a novel view of the sampled input, computing a first loss between the reconstructed view and the sampled input, computing a second loss between the novel view and the sampled input, and updating the machine learning model based on the first loss and the second loss as taught by Bhattad.
Claim(s) 2, 3, 29, 33, 34, 42, 46 and 47 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD in view of CHEN (No. WO-2021178936-A1 “Chen”).
Regarding claim 2, Dundar, Alwala and Bhattad modified with Chen teaches “The method of claim 1, wherein the single-view images in the 2D dataset are not labeled with geometry information.” (Fig. 5 Showcase a 2D image not labeled with any geometry information);
Regarding claim 3, Dundar, Alwala and Bhattad modified with Chen teaches “The method of claim 1, wherein the multi-view images in the 3D dataset are not labeled with texture information.” (Fig. 5 Showcase a 3D image not labeled with any texture information);
Regarding claim 29, while Dundar, Alwala and Bhattad do not teach the limitation of claim 29, Chen teaches “The method of claim 1, further comprising, at the device:
deploying the trained machine learning model.” (training machine learning models, deploying machine learning models; Para 0136);
The motivation for the above is to have accurate and efficient machine training model that produce accurate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Qian by deploying the trained machine learning model as taught by Chen.
Regarding claim 33, Dundar, Alwala and Bhattad modified with Chen teaches “The method of claim 29, further comprising, at the device:
processing a single input 2D image, by the deployed machine learning model, to generate 3D content; and” (Magic123, that generates 3D content from a single reference image; Pg. 3, Col 2);
“outputting the 3D content.” (Magic123, a novel image-to-3D pipeline that yields high-quality 3D outputs; Pg. 3 Para 3);
Regarding claim 34, Dundar, Alwala and Bhattad modified with Chen teaches “A system, comprising:” (a system; Para 0057);
“a non-transitory memory storage comprising instructions; and one or more processors in communication with the memory, wherein the one or more processors execute the instructions to:” (non-transitory computer-readable storage media having stored thereon executable instructions (or other memory to store executable instructions) that, when executed (i.e., as a result of being executed) by one or more processors; Para 0167);
Claim 34 is directed to a system and its limitations are similar in scope and functions performed by the method of claim 1. Therefore, claim 34 limitations are also rejected with the same rationale as regarding claim 1.
Regarding claim 42, Dundar, Alwala and Bhattad modified with Chen teaches “A non-transitory computer-readable media storing computer instructions which when executed by one or more processors of a device cause the device to:” (non-transitory computer-readable storage media having stored thereon executable instructions (or other memory to store executable instructions) that, when executed (i.e., as a result of being executed) by one or more processors; Para 0167)
Claim 42 is directed to a non-transitory computer-readable media and its limitations are similar in scope and functions performed by the method of claim 1. Therefore, claim 42 limitations are also rejected with the same rationale as regarding claim 1.
Claim 46 is directed to a non-transitory computer-readable media and its limitations are similar in scope and functions performed by the method of claim 33. Therefore, claim 46 limitations are also rejected with the same rationale as regarding claim 33.
Claim 47 is directed to a method and its limitations are similar in scope and functions performed by the method of claim 1. Therefore, claim 47 limitations are also rejected with the same rationale as regarding claim 1.
Claim(s) 10-17 and 40 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD in view of CHEN and further in view of QIAN ( “Qian”).
Regarding claim 10, modified with Dundar, Alwala and Bhattad, Chen further teaches “The method of claim 9, wherein during the first training step, a first loss is computed between the first set of 2D renderings and a ground truth from the 3D dataset.” (this 3D information can be fed as input, along with the camera information from the input image, to a differentiable graphics Tenderer... These renderings can then be compared against the relevant ground truth data, using an appropriate loss function, to determine the loss values; Para 0026);
Chen discloses an input image which is a 2D image that gets rendered by the differentiable Tenderer. That rendering is then compared to the ground truth data to determine the loss values from.
Regarding claim 11, modified with Dundar, Alwala and Bhattad, Qian teaches “The method of claim 10, wherein during the first training step, the first set of 2D renderings are supervised using the first loss.” (Fig. 3 Showcases 2D priors that is rendered and transition to a loss term); (the optimization of NeRF requires at least two loss functions: (i) the reference view reconstruction supervision; and (ii) the novel view guidance; Pg. 4, Col 2.1.1);
Regarding claim 12, modified with Dundar, Alwala and Bhattad, Chen further teaches “The method of claim 11, wherein the first loss compares one or more features between the first set of 2D renderings and a ground truth from the 3D dataset.” (The rendered representation(s) can then be compared 514 against the corresponding ground truth data to determine one or more loss values; Para 0053); (high quality 3D ground-truth; Para 0030);
Regarding claim 13, modified with Dundar, Alwala and Bhattad, Chen further teaches “The method of claim 12, wherein the one or more features include texture.” (3D properties such as geometry, texture; Para 0030);
Regarding claim 14, modified with Dundar, Alwala and Bhattad, Qian teaches “The method of claim 12, wherein the one or more features include silhouette.” (generating high-resolution 3D shapes; Pg. 5, Col 2.1.2);
Regarding claim 15, modified with Dundar, Alwala and Bhattad, Chen further teaches “The method of claim 12, wherein the one or more features include geometry information.” (3D properties such as geometry, texture; Para 0030);
Regarding claim 16, modified with Dundar, Alwala and Bhattad, Qian teaches “The method of claim 15, wherein the geometry information includes depth.” (the actual geometric details, we further extract the depth map; Pg. 4, Col 2.1);
Regarding claim 17, modified with Dundar, Alwala and Bhattad, Qian teaches “The method of claim 15, wherein the geometry information includes surface normal.” (the normal maps of geometry for the generated 3D model following [51]. We use the finite differences of the depth to estimate the normal vector of each point; Pg. 5);
Qian discloses a normal vector which relates to the surface normal, both used to determine orientation.
Regarding claim 40, modified with Dundar, Alwala and Bhattad, Qian teaches “The system of claim 39, wherein during the first training step, the first set of 2D renderings are supervised using a first loss computed between” (Fig. 3 Showcases 2D priors that is rendered and transition to a loss term); (the optimization of NeRF requires at least two loss functions: (i) the reference view reconstruction supervision; and (ii) the novel view guidance; Pg. 4, Col 2.1.1);
While modified with Dundar, Alwala, Bhattad and Qian does not teach “the first set of 2D renderings and a ground truth from the 3D dataset”.
Chen further teaches “the first set of 2D renderings and a ground truth from the 3D dataset.” (this 3D information can be fed as input, along with the camera information from the input image, to a differentiable graphics Tenderer... These renderings can then be compared against the relevant ground truth data, using an appropriate loss function, to determine the loss values; Para 0026);
Modifying modified with Dundar, Alwala and Bhattad with Qian and Chen, it would be obvious to combine the 2D renderings of Chen and the supervised loss of Qian.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein during the first training step, the first set of 2D renderings are supervised using a first loss computed between as taught by Qian and by the first set of 2D renderings and a ground truth from the 3D dataset as taught by Chen.
Claim(s) 18-27 and 41 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of ABBELOOS (No. EP-4235588-A1 “Abbeloos”) and TANG (Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., & Chen, D., "Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior.", 03 April 2023, In Proceedings of the IEEE/CVF international conference on computer vision (pp. 22819-22829). (Year: 2023) “Tang”).
Regarding claim 18, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 18, Abbeloos teaches “The method of claim 9, wherein during the second training step:
a second loss is computed between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset, and” (Loss values are then calculated for the respective inputted initial views and the corresponding outputted reconstructed views; Para 0041);
While Abbeloos does not teach “a third loss is computed between the novel view of the input image from the 2D dataset and the input image from the 2D dataset.”
Tang teaches “a third loss is computed between the novel view of the input image from the 2D dataset and the input image from the 2D dataset.” (incorporating constraints from the reference image at the frontal view and diffusion prior at novel views; Abstract);
Abbeloos discloses loss values that relate to the third and second loss being computed. Include the loss values with the reconstructed view of an input image and novel view of an input image taught by Tang.
The motivation for the above is to accurate calculations of loss with reconstructed and novel views.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by a second loss is computed between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset and a third loss is computed between the novel view of the input image from the 2D dataset and the input image from the 2D dataset as taught by Abbeloos and Tang.
Regarding claim 19, while Dundar, Alwala, Bhattad and Abbeloos fails to teach the limitations of claim 19, Tang teaches “The method of claim 18, wherein during the second training step:
the reconstructed view of the input image from the 2D dataset is supervised using the second loss, and” (we also modify it with image reconstruction loss at the reference view... reconstructing complex objects due to the lack of multi-view supervision; Pg. 22825-22826);
“the novel view of the input image from the 2D dataset is supervised using the third loss.” (To enable novel view from sparse input views, .... neural representation conditioned on the input images rather than only leveraging input views for supervision; Pg. 22820, Col 2);
Tang discloses the supervision of novel and reconstructed views which relate to reconstructed and novel views being supervised by loss. Combine with Abbeloos teaching of loss values, the second and third loss are computed are disclosed in claim 18.
The motivation for the above is to have accurate and efficient supervision of the second and third losses.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala, Bhattad and Abbeloos by the reconstructed view of the input image from the 2D dataset is supervised using the second loss, and the novel view of the input image from the 2D dataset is supervised using the third loss as taught by Tang.
Regarding claim 20, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 20, Abbeloos teaches “The method of claim 19, wherein the second loss compares one or more features between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset.” (perceptual loss term, which is a value representative of differences between features extracted respectively from the initial view and the reconstructed view; Para 0043);
The motivation for the above is to have accurate comparison and determination of features in the 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the second loss compares one or more features between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset as taught by Abbeloos.
Regarding claim 21, modified with Dundar and Alwala , Bhattad further teaches “The method of claim 20, wherein the one or more features include texture. (We extract geometry and texture from randomly paired images; Pg. 6086); (Fig. 5 showcases texture);
Regarding claim 22, modified with Dundar and Alwala, Bhattad further teaches “The method of claim 20, wherein the one or more features include geometry information.” (We extract geometry and texture from randomly paired images; Pg. 6086); (Fig. 5 showcases geometry mesh);
Regarding claim 23 while Dundar, Alwala and Bhattad fail to teach the limitation of claim 23, Abbeloos teaches “The method of claim 19, wherein the third loss compares one or more features between the novel view of the input image from the 2D dataset and the input image from the 2D dataset.” (perceptual loss term, which is a value representative of differences between features extracted respectively from the initial view and the reconstructed view; Para 0043);
Abbeloos discloses a loss term that compares the features between the initial view and reconstructed view. It would be obvious to substitute the reconstructed view with a novel view that Abbeloos discloses.
The motivation for the above is to have accurate comparison and determination of features in the 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the third loss compares one or more features between the novel view of the input image from the 2D dataset and the input image from the 2D dataset as taught by Abbeloos.
Regarding claim 24, modified with Dundar and Alwala, Bhattad further teaches “The method of claim 23, wherein the one or more features include texture.” (We extract geometry and texture from randomly paired images; Pg. 6086); (Fig. 5 showcases texture);
Regarding claim 25, modified with Dundar and Alwala, Bhattad further teaches “The method of claim 23, wherein the one or more features include geometry information.” (We extract geometry and texture from randomly paired images; Pg. 6086); (Fig. 5 showcases geometry mesh);
Regarding claim 26, modified with Dundar and Alwala, Bhattad further teaches “The method of claim 9, wherein during the second training step:
using the encoder to generate a third 3D representation of the novel view of the input image from the 2D dataset, and” (geometry and texture predictions from input images, and render them from novel views; Pg. 6085); (Using this intermediate image, we infer back the geometry and texture representations; Pg.6085 Col 3.2.1 Para 4);
“enforcing cycle consistency by minimizing a difference between the third 3D representation of the novel view of the input image from the 2D dataset and the second 3D representation of the input image from the 2D dataset.” (a cycle consistency loss that improves view generalization; Abstract); (we regularize deformation to be small and hence minimize overall deformation to be as close to zero; Pg. 6085); (impose multi-view consistency; GAN loss penalizes if rendered novel views are not realistic for both the shape and appearance, and the cycle consistency loss regularizes the network to infer predictions for occluded regions consistent with the original view; Pg. 6085);
Regarding claim 27, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 27, Abbeloos teaches “The method of claim 9, wherein during the second training step, using at least one discriminator to ensure that the novel view of the input image from the 2D dataset is realistic compared to the input image from the 2D dataset.” (the discriminator has been defined as being configured to calculate a realism score representative of a probability of a novel view outputted by the generator to be a real view; Para 0034);
The motivation for the above is to have a more realistic 3D representation for accurate machine learning training.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein during the second training step, using at least one discriminator to ensure that the novel view of the input image from the 2D dataset is realistic compared to the input image from the 2D dataset as taught by Abbeloos.
Regarding claim 41, “The system of claim 39, wherein during the second training step:
the reconstructed view of the input image from the 2D dataset is supervised using a second loss” (we also modify it with image reconstruction loss at the reference view... reconstructing complex objects due to the lack of multi-view supervision; Pg. 22825-22826);
“computed between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset, and” (Loss values are then calculated for the respective inputted initial views and the corresponding outputted reconstructed views; Para 0041);
“the novel view of the input image from the 2D dataset is supervised using a third loss computed between the novel view of the input image from the 2D dataset and the input image from the 2D dataset.” (To enable novel view from sparse input views, .... neural representation conditioned on the input images rather than only leveraging input views for supervision; Pg. 22820, Col 2); (incorporating constraints from the reference image at the frontal view and diffusion prior at novel views; Abstract);
It would be obvious to combine the loss values of Abbeloos and Tang’s reconstructed view. Similar in rationale to claims 18 and 19 which have a similar scope when combined.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala, and Bhattad the reconstructed view of the input image from the 2D dataset is supervised using a second loss computed between the reconstructed view of the input image from the 2D dataset and the input image from the 2D dataset, and the novel view of the input image from the 2D dataset is supervised using a third loss computed between the novel view of the input image from the 2D dataset and the input image from the 2D dataset as taught by Abbeloos and Tang.
Claim(s) 4, and 28 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of BOBU (No. US-20230145208-A1 “Bobu”).
Regarding claim 4, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 4, Bobu teaches “The method of claim 1, wherein the training is semi-supervised.” (training may be performed in either a supervised, partially supervised, or unsupervised manner; Para 0126);
Dundar, Alwala, Bhattad and Bobu are analogous art they are related to training machine learning model.
The motivation for the above is to an efficient supervised machine learning model.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the training is semi-supervised as taught by Bobu.
Regarding claim 28, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 28, Bobu teaches “The method of claim 1, further comprising, at the device:
using a geometry prior to train, without supervision, the machine learning model to learn geometry.” (perform 3D geometry processing and media processing via graphics core(s) 2180A-2180N. In at least one embodiment, for 3D geometry processing commands, command streamer 2103 supplies commands to geometry pipeline; Para 0401); (geometry data for 3D pipeline; Para 0459); (training may be performed in either a supervised, partially supervised, or unsupervised manner; Para 0126);
The motivation for the above is to an efficient supervised machine learning model.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by using a geometry prior to train, without supervision, the machine learning model to learn geometry as taught by Bobu.
Claim(s) 5 and 35 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of OVED (No. US-10867436-B2 “Oved”).
Regarding claim 5, modified with Dundar and Bhattad, Alwala teaches “The method of claim 1, wherein the second training step is performed after the first training step.” (We first pre-train a reconstruction model using multi-view renderings of synthetic data. We then self-train category-specific models on diverse image collections; Pg.2, Fig.2 Desc);
The motivation for the above is to efficient training of machine learning models.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar and Bhattad by wherein the second training step is performed after the first training step as taught by Alwala.
Claim 35 is directed to a system and its limitations are similar in scope and functions performed by the method of claim 5. Therefore, claim 35 limitations are also rejected with the same rationale as regarding claim 5.
Claim(s) 30-32 and 43-45 are rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of KIM (No. US-10937237-B1 “Kim”).
Regarding claim 30, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 30, Kim teaches “The method of claim 29, wherein the trained machine learning model is deployed for use by at least one downstream application.” (the client device 106 can include a client application 112; Col 7, Line 1-2);
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the trained machine learning model is deployed for use by at least one downstream application as taught by Kim.
Regarding claim 31, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 31, Kim teaches “The method of claim 30, wherein the at least one downstream application includes a virtual reality application.” (applications such as databases of three-dimensional assets, virtual or augmented reality environments, or other environments; Col 7, Line 12-13);
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the at least one downstream application includes a virtual reality application as taught by Kim.
Regarding claim 32, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 32, Kim teaches “The method of claim 30, wherein the at least one downstream application includes an augmented reality application.” (applications such as databases of three-dimensional assets, virtual or augmented reality environments, or other environments; Col 7, Line 12-13);
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the at least one downstream application includes an augmented reality application as taught by Kim.
Regarding claim 43, while Dundar, Alwala and Bhattad fail to teach the limitations, Chen teaches “The non-transitory computer-readable media of claim 42, wherein the device is further caused to:
deploy the trained machine learning model,” (training machine learning models, deploying machine learning models; Para 0136);
While Chen does not teach “wherein the trained machine learning model is deployed for use by at least one downstream application”.
Kim teaches “wherein the trained machine learning model is deployed for use by at least one downstream application.” (the client device 106 can include a client application 112; Col 7, Line 1-2);
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by deploy the trained machine learning model, wherein the trained machine learning model is deployed for use by at least one downstream application as taught by Chen and Kim.
Claim 44 is directed to a non-transitory computer-readable media and its limitations are similar in scope and functions performed by the method of claim 31. Therefore, claim 44 limitations are also rejected with the same rationale as regarding claim 31.
Claim 45 is directed to a non-transitory computer-readable media and its limitations are similar in scope and functions performed by the method of claim 32. Therefore, claim 45 limitations are also rejected with the same rationale as regarding claim 32.
Claim(s) 48 is rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of VINOD (Vinod, V., Shah, T., and Lagun, D., “TEGLO: High Fidelity Canonical Texture Mapping from Single-View Images”, <i>arXiv e-prints</i>, Art. no. arXiv:2303.13743, 2023. doi:10.48550/arXiv.2303.13743. “Vinod”).
Regarding claim 48, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 48, Vinod teaches ““The method of claim 49, wherein the single-view images are labeled with texture information.” (by mapping the input image pixels onto the texture space we can achieve near perfect reconstruction; Abstract); (takes a single-view image and its approximate camera pose to map the pixels onto a texture; Pg.2, Col 1 Para 2);
Vinod discloses using single view images collections for texture representation and mapping pixel of single vie image into a texture space. In combination with Dundar disclosing reconstruction textured 3D models using single vie image datasets and training for learning texture. Thus, it would have been obvious to a person skilled in the art to label the single view image with texture information because the texture information is the information extracted from and used to supervise the single view images for texture learning.
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the single-view images are labeled with texture information as taught by Vinod.
Claim(s) 49 is rejected under 35 U.S.C. 103 as being unpatentable over DUNDAR in view of ALWALA and in further view of BHATTAD and further in view of YAMASHITA (Yamashita, K., Enyo, Y., Nobuhara, S., & Nishino, K. (2023, January). nlmvs-net: Deep non-lambertian multi-view stereo. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (pp. 3036-3045). IEEE. (Year: 2023) “Yamashita”).
Regarding claim 49, while Dundar, Alwala and Bhattad fail to teach the limitation of claim 49, Yamashita teaches ““The method of claim 49, wherein the multi-view images are labeled with geometry information.” (Fig.1 showcase view images of an object and resulting with depths and normal that showcase the ground truth depth maps and normal maps.);
Yamashita discloses a training dataset with multi view images of an object that is associated with ground truth geometry information. Thus, teaching multi view images labeled or associated with corresponding geometry information.
The motivation for the above is to have a user friendly an application to generate 3D representations.
Therefore, it would have been obvious for an ordinary skilled person in the art before the effective filing date of claimed invention to have modified Dundar, Alwala and Bhattad by wherein the multi-view images are labeled with geometry information as taught by Yamashita.
Response to Arguments
Applicant’s arguments, see Pg.17, filed 08/12/2026, with respect to Drawings have been fully considered and are persuasive. The objection of 04/10/2026 has been withdrawn.
Applicants’ replacement sheets for drawings have overcome the previous objection.
Applicant’s arguments, see Pg.17, filed 08/12/2026, with respect to Specifications have been fully considered and are persuasive. The objection of 04/10/2026 has been withdrawn.
Applicants’ amendments to specifications have overcome the previous objection.
Applicant’s arguments, see Pg.17, filed 08/12/2026, with respect to Claim 47 have been fully considered and are persuasive. The objection of 04/10/2026 has been withdrawn.
Applicant’s arguments, see Pg.18, filed 08/12/2026, with respect to the rejection(s) of claim(s) 1 under 35 USC § 103 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Dundar, Alwala . Applicant argues that the Qian reference relates to simultaneously using 2D and 3D priors, which refer to prior trained diffusion models, to generate novel views from an image and does not teach or suggest jointly using 2D and 3D datasets to train a learning model to generate 3D content given an input 2D image. Further, the Chen reference does not cure this deficiency in the Qian reference. Neither the Qian nor the Chen reference disclose at least the above highlighted limitation recited in applicant's independent claims, the combination of such references cannot teach or suggest the same.
Examiner replies that under the combination of Dundar, Alwala and Bhattad. Dundar discloses both multi view and single view training and reconstruction of a different view in the multi view setting and reconstruction of the input view when training with single vie datasets. (SEE Pg.3, Col 3-3.1). Alwala discloses the missing training steps concept by a first pre training a reconstruction model using multi view data and after, self-training using the image collections. (SEE Pg.2 Fig 2). Bhattad discloses using novel view rendering during single view training to improve the predicated 3D object’s behavior at unseen viewpoints.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Ho, L. N., Tran, A. T., Phung, Q., & Hoai, M. (2021, October). Toward realistic single-view 3d object reconstruction with unsupervised learning from multiple images. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 12580-12590). IEEE. (Year: 2021) (Ho) – Discloses eliminating the symmetry requirement with a novel unsupervised algorithm that can learn a 3D reconstruction network from a multi-image dataset. Employing a novel albedo loss that improves the reconstructed details and realisticity.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 BRIGITER D PROTAZI whose telephone number is (571)272-7995. The examiner can normally be reached Monday - Friday 7:30-5.
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, Said A Broome can be reached at 5712722931. 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.
/B.D.P./Examiner, Art Unit 2612
/Said Broome/Supervisory Patent Examiner, Art Unit 2612
