Hybrid and self-aware long-term object tracking

1. A method of tracking an object, comprising:

performing a hybrid search over a sequence of frames, the hybrid search comprising:

intermittently performing a global search on a first set of frames of the sequence of frames; and

performing a local search on a second set of frames of the sequence of frames, the second set of frames being different from the first set of frames;

updating a similarity function based on a result of the hybrid search; and

tracking the object based on the hybrid search.

2. The method of claim 1 , in which a subset of frames of the second set of frames is located between a first frame and a second frame of the first set of frames.

3. The method of claim 2 , in which the global search comprises:

searching at a single scale over a selected frame of the first set of frames to identify potential locations of the object;

searching at different scales around the identified potential locations to select a likely box for the object; and

searching at finer scales around the likely box to select a most likely box for the object, a fit of the most likely box being closer to the object in comparison to a fit of the likely box.

4. The method of claim 3 , in which updating the similarity function comprises:

analyzing a current similarity map of the most likely box of the selected frame and previous similarity maps of previous frames;

determining whether the most likely box is correct based on the analyzing; and

updating the similarity function when the most likely box is correct.

5. The method of claim 4 , in which the current similarity map of the most likely box of the selected frame and the previous similarity maps of previous frames are analyzed with a recurrent neural network.

6. The method of claim 4 , in which determining whether the most likely box is correct comprises determining a classification score based on the analyzing.

7. The method of claim 4 , in which updating the similarity function comprises forming training pairs comprising a positive pair and a plurality of negative pairs.

8. The method of claim 7 , in which:

the positive pair comprises an initial target and the most likely box from a current the selected frame; and

the plurality of negative pairs comprises the initial target and at least one candidate box that is different from the most likely box.

9. An apparatus for tracking an object, the apparatus comprising:

a memory; and

at least one processor coupled to the memory, the at least one processor configured:

to perform a hybrid search over a sequence of frames by:

intermittently performing a global search on a first set of frames of the sequence of frames; and

performing a local search on a second set of frames of the sequence of frames, the second set of frames being different from the first set of frames;

to update a similarity function based on a result of the hybrid search; and

to track the object based on the hybrid search.

10. The apparatus of claim 9 , in which a subset of frames of the second set of frames is located between a first frame and a second frame of the first set of frames.

11. The apparatus of claim 10 , in which the at least one processor is configured to perform the global search by:

searching at a single scale over a selected frame from the first set of frames to identify potential locations of the object;

searching at different scales around the identified potential locations to select a likely box for the object; and

searching at finer scales around the likely box to select a most likely box for the object, a fit of the most likely box being closer to the object in comparison to a fit of the likely box.

12. The apparatus of claim 11 , in which the at least one processor is configured to update the similarity function by:

analyzing a current similarity map of the most likely box of the selected frame and previous similarity maps of previous frames;

determining whether the most likely box is correct based on the analyzing; and

updating the similarity function when the most likely box is correct.

13. The apparatus of claim 12 , in which the current similarity map of the most likely box of the selected frame and the previous similarity maps of previous frames are analyzed with a recurrent neural network.

14. The apparatus of claim 12 , in which the at least one processor is configured to determine whether the most likely box is correct by determining a classification score based on the analyzing.

15. The apparatus of claim 12 , in which the at least one processor is configured to update the similarity function by forming training pairs comprising a positive pair and a plurality of negative pairs.

16. The apparatus of claim 15 , in which:

the positive pair comprises an initial target and the most likely box from the selected frame; and

the plurality of negative pairs comprises the initial target and at least one candidate box that is different from the most likely box.

17. A non-transitory computer-readable medium having program code recorded thereon for tracking an object, the program code executed by a processor and comprising:

program code to perform a hybrid search over a sequence of frames, the program code to perform the hybrid search comprising:

program code to intermittently perform a global search on a first set of frames of the sequence of frames; and

program code to perform a local search on a second set of frames of the sequence of frames, the second set of frames being different from the first set of frames;

program code to update a similarity function based on a result of the hybrid search; and

program code to track the object based on the hybrid search.

18. The non-transitory computer-readable medium of claim 17 , in which a subset of frames of the second set of frames is located between a first frame and a second frame of the first set of frames.

19. The non-transitory computer-readable medium of claim 18 , in which the program code to perform the global search comprises:

program code to search at a single scale over a selected frame of the first set of frames to identify potential locations of the object;

program code to search at different scales around the identified potential locations to select a likely box for the object; and

program code to search at finer scales around the likely box to select a most likely box for the object, a fit of the most likely box being closer to the object in comparison to a fit of the likely box.

20. The non-transitory computer-readable medium of claim 19 , in which the program code to update the similarity function comprises:

program code to analyze a current similarity map of the most likely box of the selected frame and previous similarity maps of previous frames;

program code to determine whether the most likely box is correct based on the analyzing; and

program code to update the similarity function when the most likely box is correct.

21. The non-transitory computer-readable medium of claim 20 , in which the current similarity map of the most likely box of the selected frame and the previous similarity maps of previous frames are analyzed with a recurrent neural network.

22. The non-transitory computer-readable medium of claim 20 , in which the program code to determine whether the most likely box is correct comprises program code to determine a classification score based on a comparison of the current similarity map and the previous similarity maps.

23. The non-transitory computer-readable medium of claim 20 , in which the program code to update the similarity function comprises program code to form training pairs comprising a positive pair and a plurality of negative pairs.

24. The non-transitory computer-readable medium of claim 23 , in which:

the positive pair comprises an initial target and the most likely box from the selected frame; and

the plurality of negative pairs comprises the initial target and at least one candidate box that is different from the most likely box.

25. An apparatus for tracking an object, the apparatus comprising:

means for performing a hybrid search over a sequence of frames, the hybrid search comprising:

means for intermittently performing a global search on a first set of frames of the sequence of frames; and

means for performing a local search on a second set of frames of the sequence of frames, the second set of frames being different from the first set of frames;

means for updating a similarity function based on a result of the hybrid search; and

means for tracking the object based on the hybrid search.

26. The apparatus of claim 25 , in which a subset of frames of the second set of frames is located between a first frame and a second frame of the first set of frames.

27. The apparatus of claim 26 , in which the means for periodically performing the global search comprises:

means for searching at a single scale over a selected frame of the first set of frames to identify potential locations of the object;

means for searching at different scales around the identified potential locations to select a likely box for the object; and

means for searching at finer scales around the likely box to select a most likely box for the object, a fit of the most likely box being closer to the object in comparison to a fit of the likely box.

28. The apparatus of claim 27 , in which the means for updating the similarity function comprises:

means for analyzing a current similarity map of the most likely box of the selected frame and previous similarity maps of previous frames;

means for determining whether the most likely box is correct based on the analyzing; and

means for updating the similarity function when the most likely box is correct.

29. The apparatus of claim 28 , in which the current similarity map of the most likely box of the selected frame and the previous similarity maps of previous frames are analyzed with a recurrent neural network.

30. The apparatus of claim 28 , in which the means for determining whether the most likely box is correct comprises means for determining a classification score based on a comparison of the current similarity map and the previous similarity maps.