Claims:1. A method for obtaining a target view of a scene from a target viewpoint, comprising:
encoding a plurality of source images of the scene with a convolutional network to extract feature maps, wherein each of the source images is from a source viewpoint different from the target viewpoint;
processing each of the feature maps to obtain image data for the target view, the processing comprising a transformation based on the target viewpoint and a proxy three-dimensional (3D) geometry;
aggregating the image data over the plurality of source images to obtain the target view of the scene.
, Description:RELATED APPLICATIONS
[0001] The present application claims the benefit of priority to U.S. Provisional Patent Application No. 63/058,100 filed on July 29, 2020.

[0002] This application claims priority to U.S. Non-Provisional Application number 17/028,434, filed September 22, 2020, titled “DEEP NOVEL VIEW SYNTHESIS FROM UNSTRUCTURED INPUT” the entire disclosure of which is hereby incorporated by reference.

TECHNICAL FIELD
[0003] Embodiments generally relate to view synthesis. More particularly, embodiments relate to deep novel view synthesis from unstructured input.

BACKGROUND
[0004] Previously, a variety of methods have been proposed to tackle the problem of novel view synthesis from a set of input images. The proposed methods may be categorized by the restrictions on the image viewpoints and the possible deviations from the input viewpoints.

BRIEF DESCRIPTION OF THE DRAWINGS
[0005] The various advantages of the embodiments will become apparent to one skilled in the art by reading the following specification and appended claims, and by referencing the following drawings, in which:
[0006] FIG. 1 is an illustration of an example of offline and online process sequences according to an embodiment;
[0007] FIG. 2 is a block diagram of an example of a recurrent mapping and blending network according to an embodiment;
[0008] FIGs. 3-5 are comparative illustrations of examples of traditional results and enhanced results according to embodiments;
[0009] FIGs. 6A and 6B are flowcharts of examples of methods of operating performance-enhanced computing systems according to embodiments;
[0010] FIG. 7 is a block diagram of an example of a performance-enhanced computing system according to an embodiment;
[0001] FIG. 8 is an illustration of an example of a semiconductor package apparatus according to an embodiment;
[0002] FIG. 9 is a block diagram of an example of a processor according to an embodiment; and
[0011] FIG. 10 is a block diagram of an example of a multi-processor based computing system according to an embodiment.

DESCRIPTION OF EMBODIMENTS
[0012] Previous approaches to conducting novel view synthesis may have involved light field, three-dimensional geometry based and/or mapping based methods. Light field methods do not require information about the scene geometry, but assume a dense camera grid, or restrict the target view to be a linear interpolation of the input viewpoints. Light field methods have the problem of a restricted input set-up and/or a restricted deviation from the input viewpoints. For example, a typical light field set-up is a number of images arranged on a 2D plane.
[0013] Three-dimensional (3D) geometry based methods gather information in the 3D geometry of the scene, or object. In the simplest case, the color information of the viewpoints observing the given point in 3D can be aggregated in the novel target view. Recently, neural features are learned on the 3D geometry that can be rendered with another neural network. Current 3D geometry-based methods rely on a rather precise 3D geometry that is difficult to obtain with current structure-from-motion and multiple view stereo methods. Due to this reason, the current renderings of these methods are not as sharp as real images of the scene.