Unsupervised object class discovery via bottom up multiple class learning

1. A computer-implemented method for object class discovery, the method comprising:

receiving multiple images having one or more unknown object classes;

extracting multiple saliency instances from the multiple images, wherein the extracting multiple saliency instances from the multiple images comprises extracting multiple saliency windows as the multiple saliency instances using a window-based saliency detection algorithm;

computing a score for each of the multiple saliency windows;

clustering the multiple saliency windows to generate or update models wherein the clustering comprises:

grouping multiple saliency windows having scores equal to or greater than a predetermined amount into a positive bag;

grouping multiple saliency windows having scores less than the predetermined amount into a negative bag; and

discovering the one or more object classes using the models.

2. The computer-implemented method of claim 1 , wherein the clustering and the discovering are performed automatically and simultaneously.

3. The computer-implemented method of claim 1 , wherein the clustering the multiple saliency windows comprises using a maximum margin multiple class learning algorithm.

4. The computer-implemented method of claim 1 , wherein a predetermined number of saliency windows extracted from individual images of the multiple images are labeled as being associated with the positive bag.

5. The computer-implemented method of claim 4 , wherein the clustering the multiple saliency windows comprises clustering the multiple saliency instances by maximizing margins among positive bags of the multiple images.

6. The computer-implemented method of claim 4 , wherein background saliency instances of the individual images are labeled as being associated with the negative bag, and the clustering the multiple saliency windows comprises clustering the multiple saliency windows by maximizing margins:

among positive bags of the multiple images, and

between positive bags of the multiple images and negative bags of the multiple images.

7. One or more computer storage media storing computer-executable instructions that, when executed on one or more processors, causes the one or more processors to perform acts comprising:

receiving a plurality of images;

identifying multiple saliency instances from the plurality of images;

extracting a predetermined number of saliency instances of an individual image of the plurality of images, the predetermined number of the saliency instances having higher saliency scores than other saliency instances extracted from the individual image;

clustering the multiple saliency instances by optimizing margins among the predetermined number of saliency instances; and

training at least one statistical model using the clustered multiple saliency instances.

8. The one or more computer storage media of claim 7 , wherein the optimizing the margins comprising maximizing the margins.

9. The one or more computer storage media of claim 7 , wherein the acts further comprise:

grouping the predetermined number of saliency instances of individual image into a positive bag; and

determining multiple positive bags for the plurality of images.

10. The one or more computer storage media of claim 9 , wherein the optimizing margins among the predetermined number of saliency instances comprises maximizing margins among the multiple positive bags.

11. The one or more computer storage media of claim 7 , wherein the acts further comprise:

grouping the predetermined number of saliency instances of individual images into a positive bag;

grouping saliency instances associated with backgrounds of the individual images into a negative bag; and

determining multiple negative bags and multiple positive bags for the plurality of images.

12. The one or more computer storage media of claim 11 , wherein the optimizing margins among the predetermined number of saliency instances comprises maximizing margins among the positive bags and margins between the multiple positive bags and the multiple negative bags.

13. The one or more computer storage media of claim 7 , wherein the acts further comprise discovering object classes of the plurality of images using the at least one trained statistical model.

14. The one or more computer storage media of claim 7 , wherein the acts further comprise:

receiving an image containing object that belongs to an object class;

discovering the object class using the at least one trained statistical model; and

retrieving multiple images corresponding to the discovered object classes.

15. A machine learning system for object class discovery, the system comprising:

one or more processors; and

memory to maintain a plurality of components executable by the one or more processors, the plurality of components comprising:

an extractor that:

extracts saliency instances of multiple images, and

determines a first salient group and a second salient group of saliency instances of the multiple images based on saliency scores of the individual ones of the saliency instances, wherein the individual ones of the saliency instances in the first salient group have a score equal to or greater than a predetermined amount and the individual ones of the saliency instances in the second salient group have a score less than the predetermined amount,

a clustering unit that clusters the multiple saliency instances to generate or update machine learning models by optimizing margins:

among two or more of the saliency instances of the first salient group of the multiple saliency instances, and

between the two or more of the saliency instances of the first salient groups and two or more of the saliency instances of the second salient group of the multiple images, and

a detector that discovers an object class using the machine learning models.

16. The system of claim 15 , wherein the second salient group includes randomly sampled saliency instances of the individual ones of the multiple images.

17. The system of claim 15 , wherein the optimizing the margins comprises maximizing the margins.

18. The system of claim 15 , wherein the clustering and the discovering are implemented automatically and simultaneously.