IP Library Granted Patent US 12694480
Granted Patent B2
US 12694480 · App. 18/477,219 · Granted Jul 28, 2026

High-resolution multiview-consistent rendering and alpha matting from sparse views

Inventors: Sergio Orts Escolano (Zurich, CH); Zhiwen Fan (Austin, TX); Di Qiu (Kitchener, CA); Yinda Zhang (Palo Alto, CA); Daoye Wang (Zurich, CH); Erroll Wood (Cambridge, GB); Abhimitra Meka (San Francisco, CA); Hossam Isack (Oakland, CA); Paulo Fabiano Urnau Gotardo (Zurich, CH); Kripasindhu Sarkar (Zurich, CH); Thabo Beeler (Egg, CH); Zhengyang Shen (Sunnyvale, CA); Alexander Sahba Koumis (San Francisco, CA)
Assignee: GOOGLE LLC
G06T5/50G06T7/194H04N5/272G06T2207/20084G06T2207/20221
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12694480
App. No.
18/477,219
Granted
Jul 28, 2026
Kind
B2
Abstract

A method including capturing a first plurality of images that include a foreground object and a background, capturing a second plurality of images that include the background, generating an alpha matte based on the first plurality of images and the second plurality of images using a trained machine learned model trained using a loss function configured to cause the trained machine learned model to learn high-frequency details of the foreground object, generating a foreground object image based on the first plurality of images and the second plurality of images using the trained machine learned model, and synthesizing an image including the foreground object image and a second background scene using the alpha matte.

Claims (37)

1 . A non-transitory computer-readable storage medium comprising instructions stored thereon that, when executed by at least one processor, cause a computing system to:

generate an alpha matte based on a first plurality of images that include a foreground object on a first background scene and a second plurality of images including the first background scene using a model trained using a loss function configured to cause the model to determine a geometry of the foreground object that has three dimensions and first high-frequency details of the foreground object in a first stage and to determine second high-frequency details of the foreground object in a second stage;

generate a foreground object image based on the first plurality of images and the second plurality of images using the model; and

generate an image including the foreground object image and a second background scene using the alpha matte.

2 . The non-transitory computer-readable storage medium of claim 1 , wherein the loss function configured to cause the model to determine at least one of the first high-frequency details or the second high-frequency details of the foreground object includes a Sobel loss.

3 . The non-transitory computer-readable storage medium of claim 1 , wherein the loss function includes an annealing patch-wise smoothness loss.

4 . The non-transitory computer-readable storage medium of claim 1 , wherein the instructions further cause the computing system to:

capture the first plurality of images that include the foreground object and the first background scene using a sparse-view camera rig; and

capture the second plurality of images that include the second background scene using the sparse-view camera rig.

5 . The non-transitory computer-readable storage medium of claim 1 , wherein the model is a neural network including a Sobel field multilayer perceptron (MLP).

6 . The non-transitory computer-readable storage medium of claim 5 , wherein a Sobel map is used to supervise the training of the model.

7 . A method comprising:

selecting an alpha matte generated using a model;

selecting a first image that includes a foreground object generated using the model;

selecting a background image; and

generating a second image that includes the foreground object and the background image using the alpha matte, the first image and the alpha matte being generated using the model trained using a loss function configured to cause the model to determine a geometry of the foreground object that is three dimensional and first high-frequency details of the foreground object in a first stage and to determine second high-frequency details of the foreground object in a second stage.

8 . The method of claim 7 , wherein the model is trained using a first plurality of images including the foreground object and a background scene, a second plurality of images including the background scene, and the loss function includes a Sobel loss.

9 . The method of claim 8 , wherein the first plurality of images and the second plurality of images are captured using a camera rig having greater than two and less than twenty cameras.

10 . The method of claim 7 , wherein the loss function includes an annealing patch-wise smoothness loss.

11 . The method of claim 7 , wherein the model is a neural network including a Sobel field multilayer perceptron (MLP).

12 . The method of claim 11 , wherein a Sobel map is used to supervise training of the model.

13 . The method of claim 7 , wherein

the foreground object is generated with a new viewpoint,

the alpha matte is modified based on the foreground object with the new viewpoint, and

the generating of the second image includes removing a portion of the background image using the modified alpha matte and replacing the removed portion with the foreground object with the new viewpoint.

14 . A method comprising:

generating an alpha matte based on a first plurality of images that include a foreground object on a first background scene and a second plurality of images including the first background scene using a model trained using a loss function configured to cause the model to determine a geometry of the foreground object that has three dimensions and first high-frequency details of the foreground object in a first stage and to determine second high-frequency details of the foreground object in a second stage;

generating a foreground object image based on the first plurality of images and the second plurality of images using the model; and

generating an image including the foreground object image and a second background scene using the alpha matte.

15 . The method of claim 14 , wherein the loss function configured to cause model to determine at least one of the first high-frequency details or the second high-frequency details of the foreground object includes a Sobel loss.

16 . The method of claim 14 , wherein the loss function includes an annealing patch-wise smoothness loss.

17 . The method of claim 14 , wherein the method further comprises:

capturing the first plurality of images that include the foreground object and the first background scene using a sparse-view camera rig; and

capturing the second plurality of images that include the second background scene using the sparse-view camera rig.

18 . The method of claim 14 , wherein the model is a neural network including a Sobel field multilayer perceptron (MLP).

19 . The method of claim 18 , wherein a Sobel map is used to supervise training of the model.

20 . The method of claim 14 , wherein the generating of the image includes removing a portion of the second background scene using the alpha matte and replacing the removed portion with the foreground object image.