Method for pose invariant vessel fingerprinting

1. A computer-implemented method for matching objects, the method being executed by at least one processor, comprising the steps of:

receiving at least two images wherein one of the at least two images has a first target object and a second of the at least two images has a second target object;

extracting at least one first patch from the first target object and at least one second patch from the second target object;

constructing a distance-based part encoding between each of the at least one first patch and the at least one second patch based upon a corresponding codebook of image parts including at least one of part type and pose;

warping a viewpoint of one of the at least one first patch to a viewpoint of the at least one second patch; and

applying a parts level similarity measure based on the view-invariant distance measure for each of the at least one first patch and the at least one second patch to determine whether the first target object and the second target object are the same or different objects.

2. The method of claim 1 , wherein the codebook of image parts comprises a set of neighborhood local patches to a distinctive anchor shape, wherein a distance to each codebook entry forms an embedded distance vector utilized as a part encoding.

3. The method of claim 2 , wherein the codebook of image parts is a subset of a codebook based object representation, comprising a codebook vocabulary of instances of distinctive part types of an object that are characteristic of objects of the same type.

4. The method of claim 3 , wherein the codebook vocabulary is constructed by:

normalizing each image of an object of a plurality of images of objects in terms of height and tesselating a region of interest (ROI) of the object into coarse grid cell regions;

extracting commonly seen distinctive parts instances from the plurality of images to form a representative local patch set;

designating patches in the representative local patch as seed patches to collect a more extensive set of instances of matching image patches from a larger dataset based on an image-based modified normalized cross-correlation (MNCC) similarity measure; and

clustering the extensive set of instances of matching image patches into a codebook vocabulary of at the least one plurality of codebook instances according to object type and pose.

5. The method of claim 4 , wherein the step of constructing a distance-based part encoding comprises the step of computing a distance vector for each of the at least one first patch and the at least one second patch based on the Euclidan distance between the HOG representation of the at least one first patch and the at least one second patch and each of the vocabulary elements of a corresponding codebook in the codebook vocabulary.

6. The method of claim 1 , wherein the first target object and the second target object are vessels.

7. The method of claim 1 , wherein the step of warping a viewpoint is based upon a mapping between two jointly constructed low-dimensional manifolds to obtain a view-invariant distance measure in a lower dimensional space.

8. The method of claim 7 , wherein the mapping between two jointly constructed low-dimensional manifolds is learned through local linear embedding (LLE).

9. The method of claim 8 , further comprising the steps of:

detecting at least one matching codebook element for each of the at least one first patch from the first target object and at least one second patch;

for each matching codebook element, obtaining a matching codebook element response is obtained and forming a feature vector; and

for each formed feature vector,

obtaining a K nearest neighborhood from a stored sample dataset;

determining a coordinate in a lower-dimensional space using LLE;

determining a corresponding warped LLE coordinate;

computing a linear reconstruction weight using a non-parametric mapping; and

reconstructing a view rectified feature vector in a higher dimensional space based on a weighting of the warped LLE coordinate with the reconstructed weight.

10. The method of claim 1 , wherein the parts-level similarity measure is based on a same-different classifier and on view-rectified distances.

11. A system for matching objects; comprising:

at least one image capturing device for receiving at least two images wherein one of the at least two images has a first target object and a second of the at least two images has a second target object; and

a processor for:

extracting at least one first patch from the first target object and at least one second patch from the second target object;

constructing a distance-based part encoding between each of the at least one first patch and the at least one second patch based upon a corresponding codebook of image parts including at least one of part type and pose;

warping a viewpoint of one of the at least one first patch to a viewpoint of the at least one second patch; and

applying a parts level similarity measure based on the view-invariant distance measure for each of the at least one first patch and the at least one second patch to determine whether the first target object and the second target object are the same or different objects.

12. The system of claim 11 , wherein the codebook of image parts comprises a set of neighborhood local patches to a distinctive anchor shape, wherein a distance to each codebook entry forms an embedded distance vector utilized as a part encoding.

13. The system of claim 12 , wherein the codebook of image parts is a subset of a codebook based object representation, comprising a codebook vocabulary of instances of distinctive part types of an object that are characteristic of objects of the same type.

14. The system of claim 11 , wherein warping a viewpoint is based upon a mapping between two jointly constructed low-dimensional manifolds to obtain a view-invariant distance measure in a lower dimensional space.

15. The system of claim 14 , wherein the mapping between two jointly constructed low-dimensional manifolds is learned through local linear embedding (LLE).

16. The system of claim 11 , wherein the at least one image capturing device is a plurality of cameras located at one of the same and different locations.

17. The system of claim 16 , wherein the plurality of cameras comprises pairs of wide angle and pan-tilt-zoom (PTZ) camera.

18. The system of claim 16 , wherein control is transferred from one camera to another camera when the first target object matches the second target object.

19. A computer-readable medium storing computer code for matching objects, the code being executed by at least one processor, wherein the computer code comprises code for:

receiving at least two images wherein one of the at least two images has a first target object and a second of the at least two images has a second target object;

extracting at least one first patch from the first target object and at least one second patch from the second target object;

constructing a distance-based part encoding between each of the at least one first patch and the at least one second patch based upon a corresponding codebook of image parts including at least one of part type and pose;

warping a viewpoint of one of the at least one first patch to a viewpoint of the at least one second patch; and

applying a parts level similarity measure based on the view-invariant distance measure for each of the at least one first patch and the at least one second patch to determine whether the first target object and the second target object are the same or different objects.

20. The computer readable medium of claim 19 , wherein the codebook of image parts comprises a set of neighborhood local patches to a distinctive anchor shape, wherein a distance to each codebook entry forms an embedded distance vector utilized as a part encoding.

21. The computer readable medium of claim 20 , wherein the codebook of image parts is a subset of a codebook based object representation, comprising a codebook vocabulary of instances of distinctive part types of an object that are characteristic of objects of the same type.

22. The computer readable medium of claim 21 , wherein the codebook vocabulary is constructed by:

normalizing each image of an object of a plurality of images of objects in terms of height and tesselating a region of interest (ROI) of the object into coarse grid cell regions;

extracting commonly seen distinctive parts instances from the plurality of images to form a representative local patch set;

designating patches in the representative local patch as seed patches to collect a more extensive set of instances of matching image patches from a larger dataset based on an image-based modified normalized cross-correlation (MNCC) similarity measure; and

clustering the extensive set of instances of matching image patches into a codebook vocabulary of at the least one plurality of codebook instances according to object type and pose.

23. The computer readable medium of claim 22 , wherein code for constructing a distance-based part encoding comprises code for computing a distance vector for each of the at least one first patch and the at least one second patch based on the Euclidan distance between the HOG representation of the at least one first patch and the at least one second patch and each of the vocabulary elements of a corresponding codebook in the codebook vocabulary.

24. The computer readable medium of claim 19 , wherein code for warping a viewpoint is based upon a mapping between two jointly constructed low-dimensional manifolds to obtain a view-invariant distance measure in a lower dimensional space.

25. The computer readable medium of claim 24 , wherein the mapping between two jointly constructed low-dimensional manifolds is learned through local linear embedding (LLE).

26. The computer readable medium of claim 25 , further comprising code for:

detecting at least one matching codebook element for each of the at least one first patch from the first target object and at least one second patch;

for each matching codebook element, obtaining a matching codebook element response is obtained and forming a feature vector; and

for each formed feature vector,

obtaining a K nearest neighborhood from a stored sample dataset;

determining a coordinate in a lower-dimensional space using LLE;

determining a corresponding warped LLE coordinate;

computing a linear reconstruction weight using a non-parametric mapping; and

reconstructing a view rectified feature vector in a higher dimensional space based on a weighting of the warped LLE coordinate with the reconstructed weight.

27. The computer readable medium of claim 19 , wherein the parts-level similarity measure is based on a same-different classifier and on view-rectified distances.