Machine-learnt person re-identification
Automated person re-identification may be assisted by consideration of attributes of the person in a joint classification with matching of the person. By both solving for similarities in a plurality of attributes and identities, discriminative interactions may be captured. Automated person re-identification may be assisted by consideration of a semantic color name. Rather than a color histogram, probability distributions are mapped to color terms of the semantic color name. Using other descriptors as well, similarity measures for the various descriptors are weighted and combined into a score. Either or both considerations may be used.
1. A method for person re-identification, the method comprising:
mapping color values separately for different regions from first and second images to first and second probability distributions over a plurality of colors, the plurality of colors in a first color space different than a second color space of the color values;
calculating separately, with a processor, for the different regions a similarity score between the first and second probability distributions;
determining an affinity score as a function of the similarity scores from the different regions, and different weights applied to different similarity scores, the weight being a rank-boosted machine-learnt value; and
identifying a person in the second image as a person in the first image, the identifying being a function of the affinity scores.
2. The method of claim 1 wherein the second color space comprises Red, Green, Blue (RGB) color space, wherein the first color space comprises color terms non subsumable into each other, and wherein mapping comprises mapping from RGB values to probabilities across the color terms.
3. The method of claim 1 wherein mapping comprises mapping to a semantic color name.
4. The method of claim 1 wherein mapping comprises assigning different probabilities to each of the colors for each of the color values, at least two of the probabilities being greater than 0.0.
5. The method of claim 1 wherein calculating comprises calculating between vectors representing the first and second probability distributions.
6. The method of claim 1 wherein calculating comprises calculating Bhattacharyya coefficients.
7. The method of claim 1 wherein determining the affinity score comprises linearly combining the similarity and additional similarities.
8. The method of claim 1 wherein identifying the person comprises comparing the affinity score to a threshold.
9. In a non-transitory computer readable storage medium having stored therein data representing instructions executable by a programmed processor for person re-identification, the storage medium comprising instructions for:
obtaining a first image representing a person and at least a second image;
computing probability distributions of a semantic color name of the first and second images;
determining image descriptors of the first image and second image in addition to the probability distributions;
calculating similarity scores of the probability distributions and the image descriptors;
weighting the similarity scores with rank-boosted machine learnt weights;
summing linearly the weighted similarity scores; and
re-identifying the person in the second image as a function of the combination of the weighted similarity scores.
10. In a non-transitory computer readable storage medium having stored therein data representing instructions executable by a programmed processor for person re-identification, the storage medium comprising instructions for:
obtaining a first image representing a person and at least a second image;
computing probability distributions of the first and second images by mapping from a red, green, blue (RGB) color space to four or more colors in a semantic color name space; the probability distributions including probabilities less than 100%;
determining image descriptors of the first image and second image in addition to the probability distributions;
calculating similarity scores of the probability distributions and the image descriptors;
combining the similarity scores as a function of rank-boosting weights; and
re-identifying the person in the second image as a function of the combination of the similarity scores.
11. The non-transitory computer readable storage medium of claim 10 wherein combining comprises weighting the similarity scores with the rank-boosting weights, the rank-boosting weights comprising machine-learnt weights and linearly summing the weighted similarity scores.
12. The non-transitory computer readable storage medium of claim 10 wherein computing comprises mapping separately for different regions in the first and second images, wherein calculating comprises calculating a regional similarity score separately for the different regions and calculating the similarity score comprises determining as a function of the regional similarity scores with different weights applied to different regional similarity scores.
13. The non-transitory computer readable storage medium of claim 10 wherein calculating similarity scores of the probability distributions comprises calculating between vectors representing the probability distributions.
14. A method for person re-identification, the method comprising:
mapping color values from first and second images to first and second probability distributions over a plurality of colors, the plurality of colors in a first color space different than a second color space of the color values;
calculating, with a processor, similarities between characteristics of the first and second images, the characteristics including a color histogram, a texture histogram, and a covariance matrix;
determining an affinity score as a function of the similarities and weights of the respective similarities, the weights being rank-boosted machine-learnt values; and
identifying a person in the second image as a person in the first image, the identifying being a function of the affinity score.