Appearance-flow-based image generation
AssigneeShenzhen Malong Technologies Co., Ltd.
InventorsXintong HAN, Xiaojun HU, Weilin HUANG, Matthew Robert SCOTT
Life of this patent
Estimating the term…
Abstract
Aspects of this disclosure include technologies for appearance-flow-based image generation. In applications for pose-guided person image generation or virtual try-on, the disclosed system can model the appearance flow between source and target clothing regions. Further, a cascaded appearance flow estimation network is used to progressively refine the appearance flow. The resulting appearance flow can properly encode the geometric changes between the source and the target for image generation.
Claims20
A computer-readable storage device encoded with instructions that, when executed, cause one or more processors of a computing system to perform operations comprising:
- identifying a source segmentation map of a source object in a source image;
- synthesizing a target segmentation map based on the source segmentation map and a target pose of a target object;
- determining an appearance flow based on the source segmentation map and the target segmentation map, the appearance flow being configured for correlating respective pixels between a part of the source object and a part of the target object;
- refining the appearance flow based on a combination of a structural regulation, a smooth regulation, and a category-level perceptual loss;
- warping the part of the source object based on the appearance flow;
- and
- generating a target image based on the target pose and the warped part of the source object.
(canceled)
A computer-implemented method for flow-based image generation, comprising:
- determining an appearance flow between a first segment of a source image and a second segment of a target image, the appearance flow being configured for correlating a plurality of pixels between the first segment and the second segment;
- reducing a perceptual loss between the first segment and the second segment, wherein the perceptual loss comprises weighted differences between corresponding features of the first segment and the second segment;
- warping the first segment based on the appearance flow;
- and
- generating a composite image based on the first segment.
(canceled)
A system for flow-based image generation, comprising:
- a processor;
- and
- a memory having instructions stored thereon, wherein the instructions, when executed by the processor, cause the processor to:
- determine an appearance flow between a first segment of a source image and a second segment of a target image, the appearance flow being configured for correlating a plurality of pixels between the first segment and the second segment;
- refine the appearance flow in each of a plurality of cascades based on respective features related to the source image and the target image;
- warp the first segment based on the appearance flow;
- and
- generate a composite image based on the warped first segment.
(canceled)
Dependencies are inferred from the claim text.