Method and apparatus for generating a displacement map of an input dataset pair

• US 10,380,753 B1
• Filed: 05/30/2018
• Issued: 08/13/2019
• Est. Priority Date: 05/30/2018
• Status: Active Grant

• Current Assignee: AImotive Kft
• Original Assignee: AImotive Kft
• Inventors: Csordas, Robert, Kis-Benedek, Agnes, Szalkai, Balazs
• Primary Examiner(s): Park, Edward
• Assistant Examiner(s): Delaney, Molly
• Application Number:
  US15/992,930
• Time in Patent Office: 440 Days
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06T 2207/10012 Stereo images
• CPC Class Codes: G06T 2207/20016 Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 7/33 using feature-based methods
• CPC Class Codes: G06T 7/593 from stereo images
• CPC Class Codes: G06V 10/52 Scale-space analysis, e.g. wavelet analysis multi-scale boundary representations G06V10/42
• CPC Class Codes: G06V 10/82 using neural networks

Abstract

The invention is a method and an apparatus for generating a displacement map of a first input dataset and a second input dataset of an input dataset pair. The apparatus comprises

• • a neural network based feature extractor (25) for processing the first input dataset and the second input dataset so as to generate a feature map hierarchy (50) comprising a base pair of feature maps (20e, 30e) and a refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d),
  • a displacement unit comprising a first comparator unit for obtaining an initial displacement map using the base pair of feature maps (20e, 30e); and
  • a displacement refinement unit for obtain an updated displacement map for the refinement pair of feature maps (20d, 30d).

Claims

1. A method for generating a displacement map of a first input dataset and a second input dataset of an input dataset pair, each input dataset having at least one spatial dimension and/or time dimension, the method comprising the steps of

processing the first input dataset and the second input dataset by means of a neural network based feature extractor (25) so as to generate a feature map hierarchy (50) comprising a base pair of feature maps (20e, 30e) and a refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d), each pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d, 20e, 30e) constituting one level of the feature map hierarchy (50), the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) having less coarse resolution in all of their at least one spatial dimension and/or time dimension than the base pair of feature maps (20e, 30e);

generating an initial displacement map in a displacement generating operation for the base pair of feature maps (20e, 30e) of the feature map hierarchy (50) based on matching a first feature map of the base pair of feature maps (20e, 30e) with a second feature map of the base pair of feature maps (20e, 30e);

in a displacement refinement operation,

upscaling the initial displacement map in all of its at least one spatial dimension and/or time dimension to the scale of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) of the feature map hierarchy (50) with a respective upscaling factor and multiplying the values of the initial displacement map with the respective upscaling factor to generate an upscaled initial displacement map,

generating a warped version of the first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) by carrying out a warping operation using the upscaled initial displacement map on a first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) of the feature map hierarchy (50),

matching the warped version of the first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) with the second feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) so as to obtain a correction displacement map for the warped version of the first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) and the second feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d), which correction displacement map is added to the upscaled initial displacement map so as to obtain an updated displacement map.

2. The method according to claim 1, characterized in that

at least two refinement pairs of feature maps (20b, 30b, 20c, 30c, 20d, 30d) are comprised in the feature map hierarchy (50), wherein a first refinement pair of feature maps (20d, 30d) being closest to the base pair of feature maps (20e, 30e) in the feature map hierarchy (50) has a less coarse resolution than the base pair of feature maps (20e, 30e), and each successive refinement pair of feature maps (20b, 30b, 20c, 30c) being less close to the base pair of feature maps (20e, 30e) in the feature map hierarchy (50) than the first refinement pair of feature maps (20d, 30d) has a less coarse resolution than a neighbouring refinement pair of feature maps (20d, 30d) being closer to the base pair of feature maps (20e, 30e) in the feature map hierarchy (50) than the respective successive refinement pair of feature maps (20b, 30b, 20c, 30c),

the displacement refinement operation is performed using the first refinement pair of feature maps (20d, 30d) and a respective further displacement refinement operation is performed for each successive refinement pair of feature maps (20b, 30b, 20c, 30c), wherein in each further displacement refinement operation the updated displacement map obtained for the neighbouring refinement pair of feature maps (20d, 30d) being closer to the base pair of feature maps (20e, 30e) in the feature map hierarchy (50) than the respective successive refinement pair of feature maps (20b, 30b, 20c, 30c) is used as the initial displacement map being upscaled, during the upscaling in the respective displacement refinement operation, to the scale of the respective successive refinement pair of feature maps (20b, 30b, 20c, 30c) with a respective upscaling factor and the values of the updated initial displacement map are multiplied with the respective upscaling factor.

3. The method according to claim 1, characterized in that in the displacement generating operation

a plurality of shifted feature maps (82) of the first feature map of the base pair of feature maps (20e, 30e) is generated applying a plurality of different shifts on the first feature map of the base pair of feature maps (20e, 30e), and

the initial displacement map is obtained by generating a resultant shift for each data element position of the second feature map of the base pair of feature maps (20e, 30e) on the basis of investigating the matching between the plurality of shifted feature maps (82) of the first feature map of the base pair of feature maps (20e, 30e) and the second feature map of the base pair of feature maps (20e, 30e),

and/or in the displacement refinement operation

a plurality of shifted feature maps (82) of a first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) is generated applying a plurality of different shifts on the warped version of the first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d),

the correction displacement map is obtained by generating a resultant shift for each data element position of a second feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) on the basis of investigating the matching between the plurality of shifted feature maps (82) of the first feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d) and the second feature map of the refinement pair of feature maps (20b, 30b, 20c, 30c, 20d, 30d).

4. The method according to claim 3, characterised in that in the displacement generating operation for generating an output displacement map for application as the initial displacement map and/or in the displacement refinement operation for generating the output displacement map for application as the correction displacement map, matching of

the plurality of shifted feature maps (82), wherein a shift number thereof is the number of the different shifts, and

the second feature map serving as a reference feature map (80) is performed in the steps of

applying a first comparing convolution unit (84) on the reference feature map (80) to obtain a first intermediate comparison data map,

applying a second comparing convolution unit (94) on each of the plurality of shifted feature maps (82) to obtain a plurality of second intermediate comparison data maps,

adding in an addition operation the first intermediate comparison data map replicated according to the number of the different shifts and the plurality of second intermediate comparison data maps to obtain a resultant comparison data map,

generating a respective resultant shift for each data element of the resultant comparison data map in all of its at least one spatial dimension and/or time dimension, and assigning all of the respective resultant shifts to corresponding data elements in the output displacement map.

5. The method according to claim 4, characterized in that the respective resultant shift for each data element of the resultant comparison data map in all of its at least one spatial dimension and/or time dimension is generated by applying a shift investigating convolution unit (104) on the resultant comparison data map.

6. The method according to claim 5, characterized in that the feature maps have one or more feature channels, and the resultant comparison data map is stacked by the one or more feature channels of the feature maps before applying the shift investigating convolution unit (104).

7. The method according to claim 4, characterized in that the respective resultant shift for each data element of the resultant comparison data map in all of its at least one spatial dimension and/or time dimension is generated by selecting a best-matching shift from the plurality of different shifts.

8. The method according to claim 4, characterized in that the respective resultant shift for each data element of the resultant comparison data map in all of its at least one spatial dimension and/or time dimension is generated by

establishing a plurality of displacement bins for every different shift values,

generating displacement probabilities for each of the displacement bins, the displacement probabilities are calculated based on the resultant comparison data map, and

obtaining the resultant shift by weighting the shift values by the corresponding displacement probabilities.

9. The method according to claim 4, characterized by, before generating a respective resultant shift for each data element of the resultant comparison data map in all of its at least one spatial dimension and/or time dimension, applying a non-linearity layer to the resultant comparison data map.

10. The method according to claim 1, characterized in that the feature maps have one or more feature channels, and in the displacement generating operation and/or in the displacement refinement operation the matching is performed by taking into account the one or more feature channels of the feature maps, and an initial displacement map and a correction displacement map are generated, respectively, with the same or less number of coordinate channels than the number of the at least one spatial dimension and/or time dimension of the input datasets.

11. The method according to claim 1, characterized in that a pair of a first initial displacement map and a second initial displacement map is generated in the displacement generating operation, and, based on the pair of the first initial displacement map and the second initial displacement map and a pair of a first correction displacement map and a second correction displacement map generated in the displacement refinement operation, a pair of a first updated displacement map and a second updated displacement map is generated in the displacement refinement operation.

12. The method according to claim 1, characterized in that the input dataset pair is a pair of images of a stereo image, the displacement map is a disparity map, the displacement generating operation is a disparity generating operation and the displacement refinement operation is a disparity refinement operation.