GENERATING MODIFIED TWO-DIMENSIONAL IMAGES BY CUSTOMIZING FOCAL POINTS VIA THREE-DIMENSIONAL REPRESENTATIONS OF THE TWO-DIMENSIONAL IMAGES

Smart overview of the Invention

The patent application focuses on systems and methods for modifying two-dimensional images by using three-dimensional representations. These systems enable scene-based editing, allowing for the adjustment of shadows, human figures, and objects in images. By employing three-dimensional representations, the system can estimate scene scale and visualize 3D planar surfaces, enhancing the editing process. The customization of focal points in images is also a key feature, providing users with more control over their edits.

Technological Context

Advancements in computer vision and image editing technologies have paved the way for sophisticated image-related tasks such as object identification and style transfer. The described system leverages these advancements, utilizing artificial intelligence to facilitate efficient and flexible image editing. By understanding a two-dimensional image as a real-world scene, the system enables realistic modifications that reflect real-world conditions without extensive user input.

System Capabilities

The system employs machine learning models to anticipate potential edits and generate supplementary components for digital images. This pre-processing allows for intuitive user interactions with the image, treating it as a real scene with semantic areas. The system's ability to generate three-dimensional meshes based on two-dimensional images supports flexible editing while reducing the need for manual preparatory steps.

Editing Process

The scene-based image editing system uses segmentation neural networks to create object masks and hole-filling models to generate content fills. These pre-processed components facilitate object-aware modifications like moving or deleting objects in an image. Upon receiving user inputs, the system utilizes pre-generated masks and fills to execute the desired edits efficiently.

Semantic Scene Graphs

Semantic scene graphs are generated to map out characteristics of digital images, including object attributes and relationships. These graphs assist in modifying images by providing detailed information about objects and their interconnections. Users can interact with object attributes displayed via a graphical interface to make modifications, enhancing both individual object edits and relationship-aware adjustments.