Spectral Clustering of ROIs for Object Discovery

Manual labeling of large image sets is expensive Unsupervised methods are important, because ground truth information is in some applications not available Goals Group images according to the pictured objects without additional information (only the number of categories is known) Our idea: combine a generic object detector with spectral clustering techniques for object discovery Clustering based on similarities of objects, not on similar backgrounds Difficulty: find one appropriate bounding box out of many hypotheses Generic object detector (from [1]) Examples of spectral clustering (from [5]) Spectral Clustering Techniques Use the eigen-decomposition of a modified similarity matrix to project the data into a low-dimensional subspace prior to clustering [8] Kernel matrix represents pairwise similarities between feature vectors Cluster projected data with standard methods: usually k-means, w