20240046610. DETERMINING VISUAL OVERLAP OF IMAGES BY USING BOX EMBEDDINGS simplified abstract (Niantic, Inc.)

From WikiPatents
Jump to navigation Jump to search

DETERMINING VISUAL OVERLAP OF IMAGES BY USING BOX EMBEDDINGS

Organization Name

Niantic, Inc.

Inventor(s)

Anita Rau of London (GB)

Guillermo Garcia-hernando of London (GB)

Gabriel J. Brostow of London (GB)

Daniyar Turmukhambetov of London (GB)

DETERMINING VISUAL OVERLAP OF IMAGES BY USING BOX EMBEDDINGS - A simplified explanation of the abstract

This abstract first appeared for US patent application 20240046610 titled 'DETERMINING VISUAL OVERLAP OF IMAGES BY USING BOX EMBEDDINGS

Simplified Explanation

The abstract describes an image matching system that uses box embeddings to determine visual overlaps between images. The system takes two images of a 3D surface with different camera poses as input. It then uses a machine learning model to generate box encodings for each image, which define a box in an embedding space. The system calculates an asymmetric overlap factor based on the box encodings, which measures the asymmetric surface overlaps between the two images. The overlap factor includes an enclosure factor indicating how much surface from the first image is visible in the second image, and a concentration factor indicating how much surface from the second image is visible in the first image.

  • The system receives two images of a 3D surface with different camera poses.
  • It uses a machine learning model to generate box encodings for each image.
  • The box encodings define a box in an embedding space.
  • An asymmetric overlap factor is calculated based on the box encodings.
  • The overlap factor measures the asymmetric surface overlaps between the two images.
  • The overlap factor includes an enclosure factor and a concentration factor.

Potential applications of this technology:

  • Image recognition and matching in computer vision systems.
  • Object detection and tracking in augmented reality applications.
  • 3D reconstruction and modeling in virtual reality environments.

Problems solved by this technology:

  • Accurately determining visual overlaps between images with different camera poses.
  • Efficiently encoding and comparing image features for matching purposes.
  • Handling asymmetric surface overlaps in a quantitative manner.

Benefits of this technology:

  • Improved image matching accuracy and reliability.
  • Enhanced object detection and tracking capabilities.
  • More efficient and effective 3D reconstruction and modeling.


Original Abstract Submitted

an image matching system for determining visual overlaps between images by using box embeddings is described herein. the system receives two images depicting a 3d surface with different camera poses. the system inputs the images (or a crop of each image) into a machine learning model that outputs a box encoding for the first image and a box encoding for the second image. a box encoding includes parameters defining a box in an embedding space. then the system determines an asymmetric overlap factor that measures asymmetric surface overlaps between the first image and the second image based on the box encodings. the asymmetric overlap factor includes an enclosure factor indicating how much surface from the first image is visible in the second image and a concentration factor indicating how much surface from the second image is visible in the first image.

Retrieved from "http://wikipatents.org/index.php?title=20240046610._DETERMINING_VISUAL_OVERLAP_OF_IMAGES_BY_USING_BOX_EMBEDDINGS_simplified_abstract_(Niantic,_Inc.)&oldid=32774"