This package contains implementation of the Spatio-Temporal Hierarchical Matching Pursuit (ST-HMP) descriptor presented in the following paper:

[1] Marianna Madry, Liefeng Bo, Danica Kragic, Dieter Fox, "ST-HMP: Unsupervised Spatio-Temporal Feature Learning for Tactile Data". In IEEE International Conference on Robotics and Automation (ICRA), May 2014 Download: http://www.nada.kth.se/~madry/publications/madry2014ICRA.pdf

The code was developed by Marianna Madry (marianna.madry@gmail.com) at the Royal Institute of Technology (KTH), Sweden and the University of Washington, WA, USA. It is released under the BSD license.

We provide implementation of a complete classification system in which 3D data matrices (spatio-temporal sequences) are represented using the ST-HMP descriptor. The code can be also used for 2D data matrices (spatial signal, such as images) to represente them using the HMP descriptor.

Demo shows how to use the ST-HMP for an object classification task based on sequences of real tactile data. It consists of two parts:

• Learning and extracting of: -- Hierarchical Matching Pursuit (HMP) descriptor -- Spatio-Temporal Hierarchical Matching Pursuit (ST-HMP) descriptor
• Training and testing using SVM classifiers

Running demo:

• demo includes two external dependencies: liblinear and ksvdbox. To recompile the mex codes, please follow instructions in these packages
• in order to run demo, please execute the file: ./code/demo_HMP_STHMP.m
• system parameters are defined and described in a few files in the directory ./code/parameters
• generated results will be saved in the directory ./output
• demo was tested for Linux and Matlab R2011b. If any problems occur, please send an email to Marianna Madry (marianna.madry@gmail.com)

The implementation of the HMP descriptor is based on the Multipath Hierarchical Matching Pursuit software by Liefeng Bo (liefengbo@gmail.com) downloaded from: http://research.cs.washington.edu/istc/lfb/software/hmp/mhmp_cvpr.zip

• Training and testing list:

  • in each line specify class_label path_to_data_file

    For example, as in file ./data/Drimus12RAS_schunk_dexterous/format:mat/_setup/train.list 0 ../data/Drimus12RAS_schunk_dexterous/format:mat/BALL_RUBBER/grasp-1300209406.mat 1 ../data/Drimus12RAS_schunk_dexterous/format:mat/BALSAM/grasp-1300210231.mat

    • 'class_label' should be between 0-9
    • paths to your own train and test list files can be set in: ./code/parameters/set_file_paths.m

• Sensory data format:

• example of data file: ./data/Drimus12RAS_schunk_dexterous/format:mat/BALSAM/grasp-1300210231.mat
• data file contains:
  • input data saved as 3D matrix In the demo: data from six tactile sensors are saved in a structure 'TS' that contains six 3D matricies: TS{1}, TS{2}, .., TS{6} data from joint angles are saved in 'JS' 3D matrix (but these data are not used in the demo)
  • sequence length In the demo: 'seqLength', it is also equal to 'size(TS{1},3)'

Tactile data were collected for five objects using the 3-finger Schunk Dexerous hand with three proximal (14x6 pixels) and three distal (13x6 pixels) tactile sensors. The data in a text format can be found in ./data/Drimus12RAS_schunk_dexterous/format:txt. The same data in the Matlab format can be found in ./data/Drimus12RAS_schunk_dexterous/format:mat

Database was collected by Alin Drimus (drimus@mci.sdu.dk). Please directly contact Alin Drimus to obtain the complete database for 10 object categories. Detailed description of the database can be found in: [2] Alin Drimus, Gert Kootstra, Arne Bilberg, Danica Kragic, "Design of a flexible tactile sensor for classification of rigid and deformable objects", In the Robotics and Autonomous Systems, 2012 .