irdm-2016

This is the group project for Information Retrieval and Data Mining module. Group Project Option 8 - Mining fine-art paintings for creativity understanding

---

Data

All the data files are downloaded from http://www.wga.hu/index1.html

Team - Members

Name	email
Ilias Antoniou	ilias.antoniou.15@ucl.ac.uk
James Hale	james.hale.15@ucl.ac.uk
Cyrus Parlin	cyrus.parlin.15@ucl.ac.uk

Code

We used python 2.7

Need to install the following modules:

leargist
PIL

Instructions for installing leargist can be found here: https://pypi.python.org/pypi/pyleargist

Procedure

We will briefly describe the procedure needed to be followed in order to run the code.

Procedure

• We run save_images.py script in order to save images on disc. This is needed in order to extract classemes and picodes features later.
• Default setting is to save images under /images directory. A listimages.txt file is generated as well.
• We need to download vlg extractor from vlg. We need to download parameters as well (5GB)
• For linux OS we need opencv2.3.0. This release is quite old so we need to download it from opencv2.3. After downloading it we need to compile it. Build FFMPEG option gives an error so we remove it as we do not need video analysis. We follow similar procedure as described here.
• build configuration: cmake -D CMAKE_BUILD_TYPE=RELEASE -D WITH_FFMPEG=OFF -D CMAKE_INSTALL_PREFIX=/usr/local -D BUILD_ZLIB=ON -D BUILD_PYTHON_SUPPORT=ON ~/opencv/opencv-2.3.0
• Now we are ready to extract classemes and picodes using vlg. We run the following command: ./vlg_extractor --extract_classemes=FLOAT --extract_picodes2048=FLOAT --parameters-dir=parameters/parameters_1.1 ~/PycharmProjects/irdm-2016/listimages.txt ~/PycharmProjects/irdm-2016 ~/PycharmProjects/irdm-2016/features. It saves features under /features/images directory if we stick with default configuration.
• We run main.py script for analysis.

Literature Review

Notes:

• PylearGist: https://pypi.python.org/pypi/pyleargist
• PylearGist an example: http://people.csail.mit.edu/torralba/code/spatialenvelope/
• Picodes: http://vlg.cs.dartmouth.edu/projects/vlg_extractor/vlg_extractor/Home.html
• Large-scale Classification of Fine-Art Paintings: Learning The Right Metric on The Right Feature : http://arxiv.org/pdf/1505.00855v1.pdf
• Quantifying Creativity in Art Networks: http://arxiv.org/pdf/1506.00711v1.pdf

Results

Most similar paintings:

• 
• 
• It seems like the same painting with slightly different colours appear twice in the dataset. Score: 0.8866

Most different paintings:

• 
• 
• The paintings are quite different with similarity score: 0.188156
• Interestingly enough the most similar painting to Ruben's Silenus is the The Drunken Hercules of the same artist. The similarities between the two paintings are obvious:
• 

PageRank analysis on the network:

After performing PageRank analysis on the network proposed by Elgammal and Saleh we end that the top-10 bost important (novel and influential) nodes of the network are the following:

• http://www.wga.hu/html/l/lane/owlshead.html 0.000473825164847
• http://www.wga.hu/html/b/beraud/selfport.html 0.00048041151832
• http://www.wga.hu/html/k/kroyer/zbathing.html 0.000480969828368
• http://www.wga.hu/html/f/friedric/3/307fried.html 0.000489828919249
• http://www.wga.hu/html/c/courbet/4/courb400.html 0.000495143706113
• http://www.wga.hu/html/f/friedric/1/107fried.html 0.00049721862977
• http://www.wga.hu/html/f/friedric/3/306fried.html 0.000511199334791
• http://www.wga.hu/html/b/beruete/manzana2.html 0.000520213748859
• http://www.wga.hu/html/f/friedric/4/409fried.html 0.000660883742561
• http://www.wga.hu/html/f/friedric/4/410fried.html 0.000748484616618