Estimating depth from a single image
During a training phase, a machine accesses reference images with corresponding depth information. The machine calculates visual descriptors and corresponding depth descriptors from this information. The machine then generates a mapping that correlates these visual descriptors with their corresponding depth descriptors. After the training phase, the machine may perform depth estimation based on a single query image devoid of depth information. The machine may calculate one or more visual descriptors from the single query image and obtain a corresponding depth descriptor for each visual descriptor from the generated mapping. Based on obtained depth descriptors, the machine creates depth information that corresponds to the submitted single query image.
1. A method comprising:
accessing, by one or more processors of a machine, a query image;
calculating, by one or more processors of the machine, a query visual descriptor from the query image;
obtaining, by one or more processors of the machine, a reference depth descriptor that corresponds to the query visual descriptor from a database that correlates reference visual descriptors of reference images with corresponding reference depth descriptors of the reference images, each of the reference images corresponding to a different reference depth map;
generating, by one or more processors of the machine, a query depth map that corresponds to the query image based on the obtained reference depth descriptor;
subdividing, by one or more processors of the machine, the query image into superpixels; and
modifying, by one or more processors of the machine, the created depth map by modifying an orientation of a plane represented by a superpixel among the superpixels.
2. The method of claim 1 , wherein:
the modifying of the orientation of the plane represented by the superpixel is based on a summation of depth values of color pixels in the superpixel.
3. The method of claim 1 , wherein:
the modifying of the orientation of the plane represented by the superpixel is based on a cardinality of depth values of color pixels in the superpixel.
4. The method of claim 1 , wherein:
the modifying of the orientation of the plane represented by the superpixel is based on a ratio of a summation of depth values of color pixels in the superpixel to a cardinality of the depth values of the color pixels in the superpixel.
5. The method of claim 1 , wherein:
the modifying of the orientation of the plane represented by the superpixel includes assigning a constant depth value to each pixel within the superpixel.
6. The method of claim 1 , further comprising:
deforming the plane represented by the superpixel based on a random sample consensus algorithm.
7. The method of claim 1 , wherein:
the receiving of the query image receives the query image without any corresponding depth map.
8. The method of claim 1 , wherein:
the reference images and the query image are red-green-blue images devoid of depth values.
9. The method of claim 1 , wherein:
the accessing of the query image includes receiving the query image within a request to estimate depth information from the query image; and
the generating of the query depth map is in response to the request to estimate the depth information.
10. The method of claim 1 , further comprising:
prior to the obtaining of the reference depth descriptor, causing the database to correlate the reference visual descriptors of the reference images with their corresponding reference depth descriptors of the reference images.
11. The method of claim 10 , further comprising:
accessing the reference images and corresponding reference depth maps;
generating the reference visual descriptors and their corresponding reference depth descriptors based on the accessed reference images and corresponding reference depth maps; and
generating a matrix that correlates the generated reference visual descriptors with their generated corresponding depth descriptors to cause the database to correlate the reference visual descriptors with their corresponding reference depth descriptors.
12. The method of claim 10 , wherein:
each of the reference depth maps corresponds to a different reference image among the reference images and includes a depth pixel that corresponds to a color pixel in the corresponding reference image.
13. The method of claim 1 , wherein:
the query image depicts a surface of a physical item and includes a header that specifies a camera parameter; and
the generated query depth map includes a three-dimensional representation of the surface of the physical item whose surface is depicted in the query image.
14. The method of claim 13 , further comprising:
providing the three-dimensional representation of the surface of the physical item to a rendering engine configured to create a three-dimensional visualization of the surface of the physical item.
15. The method of claim 13 , wherein:
the three-dimensional representation of the surface of the physical item includes a three-dimensional point cloud; and
the method further comprises:
calculating a length of the surface of the physical item based on the three-dimensional point cloud.
16. The method of claim 15 , further comprising:
providing the calculated length of the surface of the physical item to a shipping application.
17. A non-transitory machine-readable storage medium comprising instructions that, when executed by one or more processors of a machine, cause the machine to perform operations comprising:
accessing a query image;
calculating a query visual descriptor from the query image;
obtaining a reference depth descriptor that corresponds to the query visual descriptor from a database that correlates reference visual descriptors of reference images with corresponding reference depth descriptors of the reference images, each of the reference images corresponding to a different reference depth map;
generating a query depth map that corresponds to the query image based on the obtained reference depth descriptor;
subdividing the query image into superpixels; and
modifying the created depth map by modifying an orientation of a plane represented by a superpixel among the superpixels.
18. The non-transitory machine-readable storage medium of claim 17 , wherein:
the modifying of the orientation of the plane represented by the superpixel is based on a cardinality of depth values of color pixels in the superpixel.
19. A system comprising:
one or more processors; and
a memory storing instructions that, when executed by at least one processor among the one or more processors, cause the system to perform operations comprising:
accessing a query image;
calculating a query visual descriptor from the query image;
obtaining a reference depth descriptor that corresponds to the query visual descriptor from a database that correlates reference visual descriptors of reference images with corresponding reference depth descriptors of the reference images, each of the reference images corresponding to a different reference depth map;
generating a query depth map that corresponds to the query image based on the obtained reference depth descriptor;
subdividing the query image into superpixels; and
modifying the created depth map by modifying an orientation of a plane represented by a superpixel among the superpixels.
20. The system of claim 19 , wherein the operations further comprise:
prior to the obtaining of the reference depth descriptor, causing the database to correlate the reference visual descriptors of the reference images with their corresponding reference depth descriptors of the reference images.