Method for determining the pose of a target object in an image

1 . A method for determining a pose of a target object in an input image, the method comprising the following steps:

providing a set of reference images that contain the target object and are each annotated with a respective pose of the target object;

inputting the reference images to at least one segmentor that is configured to distinguish parts and/or features of an image belonging to different objects, wherein the at least one segmentor includes: a first, coarse segmentor that is configured to determine sets of pixels of the image belonging to different objects, and a second, dense segmentor that is configured to determine a set of features for each pixel of the image;

determining, from output of the at least one segmentor, a canonical object model that pools information from the reference images in one common coordinate system;

determining, from the output of the at least one segmentor in combination with the annotations of the reference images, a representation of the target object in the output space of at least one segmentor, wherein the determining of the representation of the target object includes:

determining, for each set of pixels belonging to an object according to the coarse segmentor, based on the annotation of the respective reference image, whether this set of pixels belongs to the target object,

when the set of pixels belongs to the target object, computing an aggregate of features outputted by the dense segmentor over all pixels in the set as a positive object-level representation, and

computing an aggregate of the positive object level representations as the representation of the target object;

inputting the input image to the at least one segmentor, and determining, based on a comparison of a resulting output of the at least one segmentor and the representation of the target object, which parts of the resulting output of the at least one segmentor belong to the target object;

combining the parts belonging to the target object into a local reconstruction of the target object; and

evaluating, from a comparison of the local reconstruction and the canonical object model, a sought pose of the target object in the input image.

2 . The method of claim 1 , further comprising: optimizing parameters that characterize a behavior of the dense segmentor towards a goal of maximizing a pairwise similarity between the positive object level representations.

3 . The method of claim 2 , wherein the optimizing starts from a copy of the coarse segmentor.

4 . The method of claim 1 , wherein the comparison of the resulting output of the at least one segmentor obtained for the input image and the representation of the target object includes:

computing, for each set of pixels belonging to an object according to the coarse segmentor, an aggregate of the features outputted by the dense segmentor over all pixels in the set as a candidate representation;

evaluating a similarity between each candidate representation and the representation of the target object; and

determining the candidate representation for which the similarity is highest as a part of the output of the segmentor that belongs to the target object.

5 . The method of claim 1 , wherein the evaluating of the sought pose includes:

determining, based at least in part on the canonical object model and the local reconstruction, a region of interest in the input image that is most salient for the evaluating of the sought pose; and

preferring the determined region of interest in the evaluation of the sought pose.

6 . The method of claim 1 , further comprising:

applying parametrized rotations and/or translations to the canonical object model, and/or to the local reconstruction; and

optimizing parameters of the rotations and/or translations towards a goal of improving a match between the canonical object model and the local reconstruction.

7 . The method of claim 5 , wherein the region of interest is determined based at least in part on:

an estimate of a diameter of the target object derived from the canonical object model;

an optimal translation perpendicular to an image plane; and

a focal length of a camera used to acquire the input image.

8 . The method of claim 5 , wherein the evaluating of the sought pose further comprises:

computing, using a given object detector, a map of detection scores in the determined region of interest; and

updating the determined pose based on a pixel in the map of detection scores with a highest detection score.

9 . The method of claim 8 , further comprising:

applying parametrized rotations and/or translations to the canonical object model, and/or to the local reconstruction; and

optimizing parameters of the rotations and/or translations towards a goal of improving a match between the canonical object model and the local reconstruction;

further optimizing the parameters of the rotations and/or translations towards the goal of improving a match between the object model and the local reconstruction; and

updating the determined pose based on an outcome of the further optimizing.

10 . The method of claim 1 , further comprising:

computing, from the sought pose, an actuation signal; and

actuating, using the actuating signal: a vehicle, and/or a robot, and/or a driving assistance system, and/or a surveillance system, and/or a quality assurance system, and/or a medical imaging system.

11 . A non-transitory machine-readable data carrier on which is stored a computer program for determining a pose of a target object in an input image, the computer program, when executed by one or more computers and/or compute instances, cause the one or more computers and/or compute instances to perform the following steps:

providing a set of reference images that contain the target object and are each annotated with a respective pose of the target object;

inputting the reference images to at least one segmentor that is configured to distinguish parts and/or features of an image belonging to different objects, wherein the at least one segmentor includes: a first, coarse segmentor that is configured to determine sets of pixels of the image belonging to different objects, and a second, dense segmentor that is configured to determine a set of features for each pixel of the image;

determining, from output of the at least one segmentor, a canonical object model that pools information from the reference images in one common coordinate system;

determining, from the output of the at least one segmentor in combination with the annotations of the reference images, a representation of the target object in the output space of at least one segmentor, wherein the determining of the representation of the target object includes:

determining, for each set of pixels belonging to an object according to the coarse segmentor, based on the annotation of the respective reference image, whether this set of pixels belongs to the target object,

when the set of pixels belongs to the target object, computing an aggregate of features outputted by the dense segmentor over all pixels in the set as a positive object-level representation, and

computing an aggregate of the positive object level representations as the representation of the target object;

inputting the input image to the at least one segmentor, and determining, based on a comparison of a resulting output of the at least one segmentor and the representation of the target object, which parts of the resulting output of the at least one segmentor belong to the target object;

combining the parts belonging to the target object into a local reconstruction of the target object; and

evaluating, from a comparison of the local reconstruction and the canonical object model, a sought pose of the target object in the input image.

12 . One or more computers including a non-transitory machine-readable data carrier on which is stored a computer program for determining a pose of a target object in an input image, the computer program, when executed by the one or more computers, cause the one or more computers to perform the following steps:

providing a set of reference images that contain the target object and are each annotated with a respective pose of the target object;

inputting the reference images to at least one segmentor that is configured to distinguish parts and/or features of an image belonging to different objects, where in the at least one segmentor includes: a first, coarse segmentor that is configured to determine sets of pixels of the image belonging to different objects, and a second, dense segmentor that is configured to determine a set of features for each pixel of the image;

determining, from output of the at least one segmentor, a canonical object model that pools information from the reference images in one common coordinate system;

determining, from the output of the at least one segmentor in combination with the annotations of the reference images, a representation of the target object in the output space of at least one segmentor, wherein the determining of the representation of the target object includes:

determining, for each set of pixels belonging to an object according to the coarse segmentor, based on the annotation of the respective reference image, whether this set of pixels belongs to the target object,

when the set of pixels belongs to the target object, computing an aggregate of features outputted by the dense segmentor over all pixels in the set as a positive object-level representation, and

computing an aggregate of the positive object level representations as the representation of the target object;

inputting the input image to the at least one segmentor, and determining, based on a comparison of a resulting output of the at least one segmentor and the representation of the target object, which parts of the resulting output of the at least one segmentor belong to the target object;

combining the parts belonging to the target object into a local reconstruction of the target object; and

evaluating, from a comparison of the local reconstruction and the canonical object model, a sought pose of the target object in the input image.