FineGrainClassification

A Matlab reimplementation of the ICCV 2013 paper Fine-Grained Categorization by Alignments. The key idea of this paper is to get features from the foreground segments which are aligned by divide the foreground into sub-regions. For more details, please check the original paper.

Explanation of the code structure

• Folder align contains all the Matlab codes to run the ICCV 2013 method. Run the following scripts orderly:

  1. step1_trainEncoder.m;
  2. step2_encoding.m;
  3. step3_libsvm_kernel.m;
  4. step4_libsvm_aggre.m;
  5. step5_libsvm_traintest.m. (The reason for this complexity is that I need to parallelly encode all images)

• Folder bow is the baseline method using Bag-of-Words features, which is based on the VLFeat bow examples.

• The three used datasets are: CUB Bird 2010, CUB Bird 2011 and Stanford Dog.

• Other folders are some personal experiments, e.g. latent SVM in the folder latent. You can just ignore them.