RALF - Reinforced Active Learning Formulation

RALF is the framework used in [1] and part of the project Semi-supervised learning in image collections. This framework combines active learning and reinforcement learning to enable a time-varying trade-off among different exploration and exploitation sampling criteria that is learned online during the sampling process.

In the following framework, we provide different sampling criteria for exploration as well as exploitation. We propose a novel exploration criteria graph density ([1], Sec. 3.2) that consistently outperforms previous exploration criteria for label propagation as well as other algorithms such as SVM or KNN. More recently, we show in [2] that this criteria also helps to find more representative labels for metric learning in comparison to the other one. Additionally, we make available our implementation of the previous method with our improvements ([1] Sec. 7.1). Finally, we provide the implementation of our RALF algorithm.

Poster

Download

Code

RALF is an implementation in Matlab including HOG descriptors for ETH-80 and C-PASCAL that are used in [1]

Datasets (images only)

• ETH-80 introduced in [3] contains 3,280 images with 8 object classes photographed from 41 viewpoints in front of a uniform background.
• Cropped PASCAL 2008 introduced in [4] with currently 4,450 images of aligned objects from 20 classes. Bounding box annotations of PASCAL VOC 2008 are used to extract the objects to evaluate these images in a multi-class setting.
   

Documentation

Just unzip the downloaded file. You will find demo files for each experimental part in the directory demo. 

Citation

@inproceedings{ebert12cvpr,
   Author = {Sandra Ebert and Mario Fritz and Bernt Schiele},
   title = {RALF: A Reinforced Active Learning Formulation for Object Class    
            Recognition},
   booktitle = {IEEE Conference on Computer Vision and Pattern Recognition
                (CVPR)},
   year = {2012}
}

Contact

For further informations please contact Sandra Ebert (ebert at mpi-inf.mpg.de)