Granted Patent B2

US 12,340,556 · App. 18/117,672 · Granted Jun 24, 2025

Inventors: Jason Devitt (Redwood City, CA); Haoyang Wang (Redwood City, CA); Konstantin Azarov (Redwood City, CA)

Assignee: Compound Eye, Inc.

G06V10/751G06N3/08G06V10/7715G06V10/82

Abstract

Claims (13)

1. A method comprising:

determining a pixel hash for each pixel in a first image and each pixel in a second image;

determining a first correspondence map relating matching pixels in the first and second images by, for each of a set of pixels in the first image:

determining a set of correspondence vectors for the respective pixel based on correspondence vectors of neighboring pixels, wherein each correspondence vector identifies a corresponding pixel in the second image that is paired with the respective pixel; and

selecting a correspondence vector from the set of correspondence vectors based on a cost value determined between the pixel hashes of the paired pixels;

using a neural network to generate a second correspondence map with a greater number of correspondence vectors than the first correspondence map, wherein the neural network ingests the first correspondence map and at least one of the first or second image; and

determining a dense depth map based on the second correspondence map.

2. The method of claim 1 , wherein the set of correspondence vectors for an analysis pixel comprises each of:

an average correspondence vector between the correspondence vector of a right and left neighboring pixel of the analysis pixel; and

an average correspondence vector between the correspondence vector of a neighboring pixel above and a neighboring pixel below the analysis pixel.

3. The method of claim 1 , wherein the neural network is trained to fill gaps in the first correspondence map.

4. The method of claim 1 , wherein the first and second image are associated with a first and second timestamp, respectively, wherein the first and second timestamps are distinct.

5. The method of claim 1 , wherein at least one of the first or second correspondence map comprises an integer-accurate correspondence map, the method further comprising refining the integer-accurate correspondence map using a new pixel hash for each pixel in the first image and in the second image, wherein the new pixel hashes are distinct from the pixel hashes used to determine the integer-accurate correspondence map.

Assignments (2)

Continuity (5)

References Cited (140)