Compact Description of Local Features

Considering the growing amount of digital images in all kind of databases, the search for specific content remains a problem. Content-based image retrieval based on local features is a promising approach but comes with the problem of being memory and computational intensive. The bag of keypoints approach reduces the feature vectors to one histogramm per image. This paper shows an efficient clustering strategy to obtain the codebook needed for the bag of keypoints computation. With the codebook, compact image descriptors are computed. An object classification with compact descriptors is conducted to test the performance of the obtained compact descriptors. This study shows that multiclass classification with up to 15 classes is possible.