Skip to content

Latest commit

 

History

History
115 lines (86 loc) · 4.72 KB

README.md

File metadata and controls

115 lines (86 loc) · 4.72 KB
Raw

Hadoop implementation of the Pooled Time Series (PoT) algorithm

PoT java implementation using Apache Hadoop.

Dependencies

  • Maven (Version shouldn't matter much. Tested with 2.x and 3.x.)
  • OpenCV 2.4.x (Tested with 2.4.9 and 2.4.11)

Pre-requisites

If you get any errors running brew install opencv related to numpy, please run:

  1. pip install numpy

Now move on to OpenCV (More detailed instructions in wiki/Installing-opencv)

  1. brew install opencv --with-java

The above should leave you with a:

/usr/local/Cellar/opencv/<VERSION>/share/OpenCV/java

Directory which contains the associated dylib OpenCV dynamic library along with the OpenCV jar file.

Getting started

  1. cd hadoop-pot-assembly
  2. mvn install assembly:assembly
  3. Set OPENCV_JAVA_HOME, e.g., to export OPENCV_JAVA_HOME=/usr/local/Cellar/opencv/2.4.9/share/OpenCV/java
  4. Set POOLED_TIME_SERIES_HOME, e.g., to export POOLED_TIME_SERIES_HOME=$HOME/hadoop-pot/src/main
  5. Run pooled-time-series, e.g., by creating an alias, alias pooled-time-series="$POOLED_TIME_SERIES_HOME/bin/pooled-time-series"

The above should produce:

usage: pooled_time_series
 -d,--dir <directory>            A directory with image files in it
 -f,--file <file>                Path to a single file
 -h,--help                       Print this message.
 -j,--json                       Set similarity output format to JSON.
                                 Defaults to .txt
 -o,--outputfile <output file>   File containing similarity results.
                                 Defaults to ./similarity.txt
 -p,--pathfile <path file>       A file containing full absolute paths to
                                 videos. Previous default was
                                 memex-index_temp.txt

So, to call the code e.g., on a directory of files called data, you would run (e.g., with OpenCV 2.4.9):

pooled-times-series -d data

Alternatively you can create (independently of this tool) a file with absolute file paths to video files, 1 per line, and then pass it with the -p file to the above program.

Running Hadoop Jobs

Config and Getting Started

Add the following to your .bashrc

export HADOOP_OPTS="-Djava.library.path=<path to OpenCV jar> -Dmapred.map.child.java.opts=-Djava.library.path=<path to OpenCV jar>"
alias pooled-time-series-hadoop="$POOLED_TIME_SERIES_HOME/bin/pooled-time-series-hadoop"

Build and clean up the jar for running

# Compile everything
mvn install assembly:assembly

# Drop the LICENSE file from our jar that will give us headaches otherwise
zip -d target/pooled-time-series-1.0-SNAPSHOT-jar-with-dependencies.jar META-INF/LICENSE

Documentation moving to the wiki

We are moving our documentation to the wiki. Please bear with us and report issues as you find them.

Research Background and Detail

This is a source code used in the following conference paper [1]. It includes the pooled time series (PoT) representation framework as well as basic per-frame descriptor extractions including histogram of optical flows (HOF) and histogram of oriented gradients (HOG). For more detailed information on the approach, please check the paper.

If you take advantage of this code for any academic purpose, please do cite:

[1] Mattmann, Chris A., and Madhav Sharan. "Scalable Hadoop-Based Pooled Time Series of Big Video Data from the Deep Web." Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval. ACM, 2017.
[2] M. S. Ryoo, B. Rothrock, and L. Matthies, "Pooled Motion Features for First-Person Videos", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015.

https://arxiv.org/abs/1610.06669
http://arxiv.org/pdf/1412.6505v2.pdf

@inproceedings{mattmann2017scalable,
  title={Scalable Hadoop-Based Pooled Time Series of Big Video Data from the Deep Web},
  author={Mattmann, Chris A and Sharan, Madhav},
  booktitle={Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval},
  pages={117--120},
  year={2017},
  organization={ACM}
}

@inproceedings{ryoo2015pot,
 title={Pooled Motion Features for First-Person Videos},
 author={M. S. Ryoo and B. Rothrock and L. Matthies},
 booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
 year={2015},
 month={June},
 address={Boston, MA},
}

Evaluation

HMDB Dataset - http://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/