Object detection in images based on affinity determinations

1. A computer-implemented method comprising:

receiving a collection of images representative of a class;

computing a respective part-based affinity between each pair of images in the collection of images, including:

partitioning a first image of the pair of images into a plurality of partitions;

determining a respective affinity for each partition of the first image, wherein the affinity is based at least in part on a measure of similarity between the partition and a location in a second image of the pair of images having a highest measure of similarity with the partition of the first image;

determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image;

generating a plurality of clusters of the images in the collection of images, wherein the plurality of clusters of images are based on the respective determined part-based affinities between each pair of images in the collection of images;

selecting images in each cluster as being representative of a particular subclass of a plurality of subclasses of the class; and

training a respective deformable parts model using images selected for each of the plurality of subclasses.

2. The method of claim 1 , wherein partitioning a first image of the pair of images into a plurality of partitions comprises:

determining a bounding box within the first image, wherein the bounding box designates an object region within the image; and

partitioning the bounding box in the first image into a plurality of partitions.

3. The method of claim 2 , wherein determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image comprises determining respective affinities of the plurality of partitions within the bounding box of the first image with corresponding locations in the second image.

4. The method of claim 1 , wherein determining a respective affinity for each partition of the first image comprises:

determining a corresponding partition of the second image for the partition of the first image; and

determining the measure of similarity between the partition of the first image and a location in the second image, the location in the second image being located within the corresponding partition of the second image.

5. The method of claim 1 , wherein determining a respective affinity for each partition of the first image comprises determining a measure of similarity between the partition of the first image and a location in the second image using a partition-specific distance function.

6. The method of claim 1 , wherein generating a plurality of clusters comprises:

computing a respective generalization error for each of a number of model orders i, wherein the model order i specifies a maximum number of clusters to generate from the images in the collection of images; and

selecting a model order i having a generalization error that satisfies a threshold.

7. A system comprising:

one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:

receiving a collection of images representative of a class;

computing a respective part-based affinity between each pair of images in the collection of images, including:

partitioning a first image of the pair of images into a plurality of partitions;

determining a respective affinity for each partition of the first image, wherein the affinity is based at least in part on a measure of similarity between the partition and a location in a second image of the pair of images having a highest measure of similarity with the partition of the first image;

determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image;

generating a plurality of clusters of the images in the collection of images, wherein the plurality of clusters of images are based on the respective determined part-based affinities between each pair of images in the collection of images;

selecting images in each cluster as being representative of a particular subclass of a plurality of subclasses of the class; and

training a respective deformable parts model using images selected for each of the plurality of subclasses.

8. The system of claim 7 , wherein partitioning a first image of the pair of images into a plurality of partitions comprises:

determining a bounding box within the first image, wherein the bounding box designates an object region within the image; and

partitioning the bounding box in the first image into a plurality of partitions.

9. The system of claim 8 , wherein determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image comprises determining respective affinities of the plurality of partitions within the bounding box of the first image with corresponding locations in the second image.

10. The system of claim 7 , wherein determining a respective affinity for each partition of the first image comprises:

determining a corresponding partition of the second image for the partition of the first image; and

determining the measure of similarity between the partition of the first image and a location in the second image, the location in the second image being located within the corresponding partition of the second image.

11. The system of claim 7 , wherein determining a respective affinity for each partition of the first image comprises determining a measure of similarity between the partition of the first image and a location in the second image using a partition-specific distance function.

12. The system of claim 7 , wherein generating a plurality of clusters comprises:

computing a respective generalization error for each of a number of model orders i, wherein the model order i specifies a maximum number of clusters to generate from the images in the collection of images; and

selecting a model order i having a generalization error that satisfies a threshold.

13. A computer program product, encoded on one or more non-transitory computer storage media, comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving a collection of images representative of a class;

computing a respective part-based affinity between each pair of images in the collection of images, including:

partitioning a first image of the pair of images into a plurality of partitions;

determining a respective affinity for each partition of the first image, wherein the affinity is based at least in part on a measure of similarity between the partition and a location in a second image of the pair of images having a highest measure of similarity with the partition of the first image;

determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image;

generating a plurality of clusters of the images in the collection of images, wherein the plurality of clusters of images are based on the respective determined part-based affinities between each pair of images in the collection of images;

selecting images in each cluster as being representative of a particular subclass of a plurality of subclasses of the class; and

training a respective deformable parts model using images selected for each of the plurality of subclasses.

14. The computer program product of claim 13 , wherein partitioning a first image of the pair of images into a plurality of partitions comprises:

determining a bounding box within the first image, wherein the bounding box designates an object region within the image; and

partitioning the bounding box in the first image into a plurality of partitions.

15. The computer program product of claim 14 , wherein determining a part-based affinity between the first image and the second image based at least in part on a sum of the respective affinities of the plurality of partitions of the first image with corresponding locations in the second image comprises determining respective affinities of the plurality of partitions within the bounding box of the first image with corresponding locations in the second image.

16. The computer program product of claim 13 , wherein determining a respective affinity for each partition of the first image comprises:

determining a corresponding partition of the second image for the partition of the first image; and

determining the measure of similarity between the partition of the first image and a location in the second image, the location in the second image being located within the corresponding partition of the second image.

17. The computer program product of claim 13 , wherein determining a respective affinity for each partition of the first image comprises determining a measure of similarity between the partition of the first image and a location in the second image using a partition-specific distance function.

18. The computer program product of claim 13 , wherein generating a plurality of clusters comprises:

computing a respective generalization error for each of a number of model orders i, wherein the model order i specifies a maximum number of clusters to generate from the images in the collection of images; and

selecting a model order i having a generalization error that satisfies a threshold.