Finding semantic parts in images
Embodiments of the present invention relate to finding semantic parts in images. In implementation, a convolutional neural network (CNN) is applied to a set of images to extract features for each image. Each feature is defined by a feature vector that enables a subset of the set of images to be clustered in accordance with a similarity between feature vectors. Normalized cuts may be utilized to help preserve pose within each cluster. The images in the cluster are aligned and part proposals are generated by sampling various regions in various sizes across the aligned images. To determine which part proposal corresponds to a semantic part, a classifier is trained for each part proposal and semantic part to determine which part proposal best fits the correlation pattern given by the true semantic part. In this way, semantic parts in images can be identified without any previous part annotations.
1. One or more non-transitory computer storage media storing computer-useable instructions that, when used by a computing device, cause the computing device to perform a method for finding semantic parts in images, the method comprising: applying a convolutional neural network (CNN) to a set of images, the CNN detecting features for each image, each image being defined by a feature vector; clustering a subset of the set of images in accordance with a similarity between feature vectors; generating a plurality of part proposals, the plurality of part proposals comprising parts at various locations and of various sues for an image of the subset of images; and associating, via information gain matching, a label with at least one of the parts for the image.
2. The one or more computer storage media of claim 1 , wherein the features are detected by the fourth layer of the CNN.
3. The one or more computer storage media of claim 2 , further comprising utilizing Euclidean distance as a distance metric between feature vectors after L 2 normalization.
4. The one or more computer storage media of claim 1 , wherein the clustering is accomplished via normalized cuts.
5. The one or more computer storage media of claim 1 , further comprising aligning each image in the subset of images.
6. The one or more computer storage media of claim 5 , further comprising:
sampling a set of points in a single image of the subset of images, the single image being the closest to the cluster centroid;
propagating the set of points to every other image of the subset of images based on learned alignments; and
wherein the plurality of part proposals are generated by sampling regions around the set of points yielding a set of regions in each image corresponding to exactly one other region per image in the subset of images.
7. The one or more computer storage media of claim 1 , wherein each image is annotated with one or more localized attribute labels.
8. The one or more computer storage media of claim 1 , further comprising:
assuming that a first part proposal is actually a first semantic part;
training a classifier at the first part proposal for the attributes localized at the first semantic part; and
computing an information gain between the classifier and ground truth attributes of other semantic parts of the image.
9. The one or more computer storage media of claim 8 , further comprising:
determining that the information gain at other semantic parts of the image matches an expected pattern; and
determining the first part proposal is the first semantic part.
10. The one or more computer storage media of claim 8 , further comprising:
determining that the information gain at other semantic parts of the image does not match an expected pattern; and
determining the first part proposal is not the first semantic part.
11. A method for associating part proposal with semantic parts, the method comprising:
sampling a set of points in a single image of a subset of images, each image is annotated with one or more localized attribute labels;
propagating the set of points to every other image of the subset of images based on learned alignments;
generating a plurality of part proposals by sampling regions around the set of points yielding a set of regions in each image corresponding to exactly one other region per image in the subset of images;
assuming that a first part proposal is actually a first semantic part;
training a classifier at the first part proposal for the attributes localized at the first semantic part; and
computing an information gain between the classifier and ground truth attributes of other semantic parts of the image.
12. The method of claim 11 , further comprising identifying feature vectors associated with each image in a set of images.
13. The method of claim 12 , further comprising determining the subset of images in the set of images that contains objects of similar poses.
14. The method of claim 11 , determining the information gain at other semantic parts of the image matches an expected pattern.
15. The method of claim 14 , further comprising determining the first part proposal is the first semantic part.
16. The method of claim 11 , determining the information gain at other semantic parts of the image does not match an expected pattern.
17. The method of claim 16 , further comprising determining the first part proposal is the first semantic part.
18. A computerized system comprising:
one or more processors; and
a non-transitory computer storage media storing computer-useable instructions that, when used by the one or more processors, cause the one or more processors to:
apply a convolutional neural network (CNN) to a set of images, the CNN detecting features for each image, each image being defined by a feature vector;
cluster a subset of the set of images in accordance with a similarity between feature vectors;
generate a plurality of part proposals, the plurality of part proposals comprising parts at various locations and of various sizes for an image of the subset of images; and
associate a label with at least one of the parts for the image.
19. The computerized system of claim 18 , wherein computer-useable instructions that, when used by the one or more processors, cause the one or more processors to:
assume that a first part proposal is actually a first semantic part;
train a classifier at the first part proposal for the attributes localized at the first semantic part; and
compute an information gain between the classifier and ground truth attributes of other semantic parts of the image.
20. The computerized system of claim 18 , wherein computer-useable instructions that, when used by the one or more processors, cause the one or more processors to:
determine the first part proposal is the first semantic part if the information gain at other semantic parts of the image matches an expected pattern; and
determine the first part proposal is not the first semantic part if the information gain at other semantic parts of the image does not match an expected pattern.