Visual recognition analyses images/videos and provides insights into their visual content. Depends on the available training data, visual recognition can be solved in 4 ways.





set up



(a) Traditional Visual Recognition:

Traditional methods train machine learning models using available training samples which come from the same or similar distribution of the testing samples.

(b) Domain Adaptation for Visual Recognition:

When the distribution of the training samples is different than the testing distribution, traditional methods (a) would fail to address the visual recognition. The typical approach (domain adaptation) is to also use some samples from target together with the source training samples.

(c) Learning Using Privileged Information for Visual Recognition (CVPR 2016):

In some cases, we have access to more information in order to improve the performance compared to (a) and (b). Basically, it may be possible to collect some additional information about the available source dataset such as attributes or depth in training. This additional information is named as Privileged Information and is only available in training.


In paper we proposed a method to exploite Privileged Information by extending the information bottleneck method and combining it with risk minimization.


We provided several experiments with different types of privileged information (Color, Attributes, Skeleton, Bounding Box) as follow:


set up

For each dataset we showed how much the privileged information improves the performance compared to the baseline (not using the privileged information):


set up

For more information read the paper and the supplementary material:

BibTex       PDF        Supplementary Material        

@InProceedings{Motiian_2016_CVPR,
author = {Motiian, Saeid and Piccirilli, Marco and Adjeroh, Donald A. and Doretto, Gianfranco},
title = {Information Bottleneck Learning Using Privileged Information for Visual Recognition},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2016}
}

(d) Domain Adaptation with Privileged Information for Visual Recognition (ECCV 2016):

Some cases may suffer from covariate shift. Therefore domain adaptation methods (b) need to be used to address the covariate shift (by using some target samples in training). Similar to (c) we may also have access to Privileged Information. There are very few works that have exploited both Privileged Information and target samples in training at the same time.

In paper we proposed a method to exploite Privileged Information and some unlabeled target samples by extending the information bottleneck method and combining it with risk minimization.


We provided two experiments with depth as privileged information as follow:


set up

For each experiment we showed how much using privileged information, using targets samples, and using both improve the performance over the baseline (not using the privileged information and target data):


set up

For more information read the paper and the supplementary material:

BibTex       PDF        Supplementary Material        

@inproceedings{motiian2016information,
  title={Information Bottleneck Domain Adaptation with Privileged Information for Visual Recognition},
  author={Motiian, Saeid and Doretto, Gianfranco},
  booktitle={European Conference on Computer Vision},
  pages={630--647},
  year={2016},
  organization={Springer}
}