Systems and methods for querying three-dimensional (3D) models using descriptors

A computer-program product tangibly embodied in a non-transitory machine-readable storage medium, comprising instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• generating or obtaining one or more depth images of a physical structure or of a three-dimensional (3D) model of the physical structure, wherein a depth image corresponds to a two-dimensional (2D) projection of the physical structure or the 3D model from a corresponding viewpoint and includes pixel information representing distances of surfaces of the physical structure from the corresponding viewpoint;
• encoding, into a first encoded representation of binary values, the one or more depth images at a first resolution generated according to a first spatial scale by a first convolutional layer of a machine-learning model, wherein the first convolutional layer encodes all of the information contained within the one or more depth images at the first spatial scale;
• encoding, into a second encoded representation of binary values, the one or more depth images at a second resolution according to a second spatial scale by a second convolutional layer of the machine-learning model, wherein the second convolutional layer encodes all of the information contained within the one or more depth images at the second spatial scale, wherein the first spatial scale and the second spatial scale are different;
• combining the first encoded representation from the output of the first convolutional layer with the second encoded representation from the output of the second convolutional layer into a single descriptor of binary values characterizing the physical structure or the 3D model;
• generating, from a data structure, a plurality of existing descriptors each characterizing a corresponding existing 3D model;
• and
• querying the data structure of the plurality of existing descriptors to identify one or more output descriptors using the single descriptor as input, wherein the one or more output descriptors comprise existing descriptors from the plurality of existing descriptors that exhibit a similarity to the single descriptor that is greater than a threshold similarity value.

A computer-implemented method comprising:

• generating or obtaining a plurality of depth images of a physical structure or of a three-dimensional (3D) model of the physical structure from different viewpoints, wherein each depth image corresponds to a two-dimensional (2D) projection of the physical structure or the 3D model from a corresponding viewpoint of the different viewpoints and includes pixel information representing distances to the surfaces of the physical structure or 3D model that are visible from the corresponding viewpoint;
• encoding, into a first encoded representation, the plurality of depth images at a first resolution generated according to a first spatial scale by a machine-learning model, wherein the first encoded representation encodes all of the information contained within the plurality of depth images at the first spatial scale;
• encoding, into a second encoded representation, the first encoded representation of the plurality of depth images at a second resolution according to a second spatial scale by the machine-learning model, wherein the second encoded representation encodes all of the information contained within the first encoded representation at the second spatial scale, wherein the first spatial scale and the second spatial scale are different;
• and
• concatenating the first encoded representation of the plurality of depth images at the first resolution and the second encoded representation of the plurality of depth images at the second resolution into a single descriptor characterizing the physical structure or the 3D model.