SOUMAJYOTI / Multiclass-Object-classification

Performs mutiple class object classification using an SVM based on bag-of-words model and spatial pyramid matching.

Multiclass-Object-classification

Written by Soumajyoti Sarkar as part of the Tagme Machine Learning Contest 2014 held by Indian Institute of Science Bangalore.

A brief description of the algorithms used is explained here https://docs.google.com/file/d/0B_i42sRcBimDLTQ1VXN2SWdyemM/

1. Installation =============== Please add the vlfeat software and libsvm software (links given in the writeup)
   into the Matlab PATH variable or set the path of thes softwares from the File->Set Path GUI option in the Matlab window.

2. Preparing the training dataset. ============

   1. Obtain the concatenated HOG histograms of all the training images by running HistogramListGlobal_HOG.m from the command window as follows: > histograms_HOG = HistogramListGlobal_HOG; The output gets stored in HOG_model folder.
   2. Obtain the concatenated SIFT histograms of all the training images by running encoder_classify.m from the command window as follows: > encoder_classify; The output gets stored in SIFT_model folder.
   3. Obtain the concatenated pyramid histograms of all the training images by running spatial_pyramid.m from the command window as follows: > spatial_pyramid; The output gets stored in Pyramid folder.

3. Testing the models on test images ===========================

   1. Run the script demo_tag.m that computes the classes to which each image in the category belongs as follows: > demo_tag;
   2. Create the file labels.txt by running the script write_labels_file.m as follows; > write_labels_file(dir, label); Here dir is the path of the folder containing test images. This step would create a file labels.txt that contain the results.

   The steps 1 and 2 in this section should be run in succession.

4. Few notes about the folders/files in general =========================

   hog - Contains the codes for extracting HOG features. kernel - Contains the kernel codes- rbf, chi-square kernel and histogram-intersection kernel. spatial_pyramid - Contains the codes for pyramid building.

   c_enhance.m - performs the preprocessing operations as described in writeup. computeH_HOG.m - computes the histograms of words of a particular class from images. computeHistogram.m - computes histogram of words of an image. computeVlist_HOG.m - computes the visual vocabulary of a particulatr class from images. encodeimage.m - computes the Fisher encoding of an image from SIFT descriptors. getImageSet.m - Gets all images from a folder. quantizeDescriptors.m - Quantizes the descriptors to form visual words from visual vocabulary.

5. Contact ========================= Soumajyoti Sarkar Email: [e-mail]