mygit007hub / Image-Segmentation-by-Correlation-Clustering

This is the project that we did for the Computer Vision course at Stony Brook University. It is a learning based method where the number of clusters do not have to be specified from the beginning. Image Segmentation by Correlation Clustering gave us extremely good results with Boundary displacement Error as low as 4.85 compared to 10.81 in the work "Higher-Order Correlation Clustering for Image Segmentation"

Image-Segmentation-by-Correlation-Clustering

1. Libraries used VLFeat's SLIC superpixels, SVM training

2. Structure of the code

2.1 Files for Training superPixel.m ­ Initial Superpixel generation code construct_superPixelGraph.m ­ Constructing pairwise superpixel graphs from superpixels. featureExtraction.m ­ Extracting pairwise features between adjacent superpixels in the graphs. ground_truth_by_maximum_consesus.m ­ As BSD (Berkeley data) gives the ground truth segmentation by multiple subjects for each image, we form a single ground truth image by taking the maximum consensus over each segmentation of the image. assign_label_to_edges.m ­ Using the ground truth, the edges of the constructed graph are labeled +1 or ­1 train_classifier.m ­ For the features extracted and the labels assigned to the edges, the SVM is trained retrain_classifier.m ­ Retrained with Hard Negative samples color_segments.m ­ used to color the segments in the image for displaying

2.2 Files for Multiple iterations Approach The folders: order1_files, order2_files, order3_files, order4_files, order5_files, order6_files, order7_files, order8_files, contain the files for each iteration respectively. For each iteration, we have, construct_superPixelGraph.m featureExtraction.m assign_label_to_edges.m superpixels_order.m ­ Generates the superpixels for the next iteration 2.2 Files for Correlation Clustering: correlation_clustering.py ­ (Source ­ https://github.com/filkry/py­correlation­clustering) It takes in the input of the edge labeled graph in the form of a text file. The file is parsed to construct the graph using network library. Correlation clustering is performed on the graph and the output of clusters formed is given in a text file.

2.3 Files for Evaluation compare_segmentations.m ­ To calculate performance measures like Probabilistic Rand Index and Variation of Information. (Source ­ https://code.google.com/p/phd­workspace/source/browse/trunk/Berkeley_Segmentation/segb ench/SegmentationBenchmark/?r=5) compare_image_boundary_error.m ­ To calculate performance measures like Boundary Displacement Error (Source ­ https://code.google.com/p/phd­workspace/source/browse/trunk/Berkeley_Segmentation/segb ench/SegmentationBenchmark/?r=5)