Systems and methods for reconstructing a three-dimensional object from an image

A system for three-dimensional (3D) object reconstruction using a single input image of a target 3D object, the system comprising:

• a memory storing a 3D neural network model and a plurality of processor-executed instructions;
• and
• one or more processors executing the plurality of processor-executed instructions to perform operations comprising:
• generating, from the single input image, a reference view of the target 3D object corresponding to a first camera pose;
• generating, for a 3D neural network model, a 3D representation of the 3D object comprising a plurality of colors and a plurality of densities corresponding to a plurality of points, based on the reference view;
• generating, for at least one point in the reference view, a first estimated visibility depth indicating a range along a first plurality of rays that restrain a color contribution from the first point, based on a first density corresponding to the first point;
• constraining, based at least in part on estimated visibility depths, the plurality of colors and the plurality of densities for the plurality of points to generate a constrained 3D representation of the target 3D object prior to performing any training of the 3D neural network model;
• training, after the constraining, the 3D neural network model based on the constrained 3D representation;
• and
• outputting, via a user interface, a visualization of the target 3D object based on constrained 3D representation comprising the plurality of constrained colors and the plurality of constrained densities.

A non-transitory processor-readable storage medium storing a plurality of processor executed instructions for three-dimensional (3D) object reconstruction using a single input image of a target 3D object, the plurality of processor-executed instructions being executed by one or more processors to perform operations comprising:

• generating, from the single input image, a reference view of the target 3D object corresponding to a first camera pose;
• generating, for a 3D neural network model, a 3D representation of the 3D object comprising a plurality of colors and a plurality of densities corresponding to a plurality of points, based on the reference view;
• generating, for at least one point in the reference view, a first estimated visibility depth indicating a range along a first plurality of rays that restrain a color contribution from the first point, based on a first density corresponding to the first point;
• constraining, based at least in part on estimated visibility depths, the plurality of colors and the plurality of densities for the plurality of points to generate a constrained 3D representation of the target 3D object prior to performing any training of the 3D neural network model;
• training, after the constraining, the 3D neural network model based on the constrained 3D representation;
• and
• outputting, via a user interface, a visualization of the target 3D object based on constrained 3D representation comprising the plurality of constrained colors and the plurality of constrained densities.