Camera invariant depth estimation
AssigneeNVIDIA Corporation
InventorsHang Su, Daniel Lichy, Abhishek Haridas Badki, Jan Kautz, Orazio Gallo
Life of this patent
Estimating the term…
Abstract
Apparatuses, systems, and techniques of identifying one or more distances to one or more objects. In at least one embodiment, one or more distances to one or more objects are identified based, at least in part on one or more first optical rays of a first image depicting an objects and one or more second optical rays a second images depicting the object. In at least one embodiment, one or more machine learning processes (e.g., neural network(s)) are used to identify a correspondence ray to determine distances to one or more objects.
Claims17
A processor, comprising:
- one or more circuits to:
- obtain image data depicting one or more objects from two or more image sources having different optical or lens characteristics;
- sample portions of a virtual spherical surface based, at least in part, on the obtained image data and data representing the different optical or lens characteristics to generate encodings of rays corresponding to one or more pixels of the image data from the two or more image sources, wherein the ray-based encodings are generated to be invariant to the different optical or lens characteristics;
- and
- use one or more neural networks to identify one or more distances to the one or more objects based, at least in part, on the encodings of rays input to the one or more neural networks.
A system, comprising:
- one or more processors to:
- obtain image data depicting one or more objects from two or more image sources having different optical or lens characteristics;
- sample portions of a virtual spherical surface based, at least in part, on the obtained image data and data representing the different optical or lens characteristics to generate encodings of rays corresponding to one or more pixels of the image data from the two or more image sources, wherein the ray-based encodings are generated to be invariant to the different optical or lens characteristics;
- and
- use one or more neural networks to identify one or more distances to the one or more objects based, at least in part, on the encodings of rays input to the one or more neural networks.
A method, comprising:
- obtaining image data depicting one or more objects from two or more image sources having different optical or lens characteristics;
- sampling portions of a virtual spherical surface based, at least in part, on the obtained image data and data representing the different optical or lens characteristics to generate encodings of rays corresponding to one or more pixels of the image data from the two or more image sources, wherein the ray-based encodings are generated to be invariant to the different optical or lens characteristics;
- and
- using one or more neural networks to identify one or more distances to the one or more objects based, at least in part, on the encodings of rays input to the one or more neural networks.
Dependencies are inferred from the claim text.