Towards extensible automatic image annotation with the bag-of-words approach

• Authors: Robert Nagy;Klaus Meyer-Wegener
• Venue: Proceedings of the international workshop on Very-large-scale multimedia corpus, mining and retrieval
• Year: 2010

Visual-word-based image categorization has proven to be very effective in several publications and contests. Recently, various approaches have been proposed to address the need for scalability and computational performance of classification based on Bag of Words. Despite these efforts, extensibility still remains an issue. Classifiers and histograms of visual words are both heavily dependent on an immutable general visual vocabulary created during the training step based on training images. Adding a new category that is insufficiently represented by the visual words in the vocabulary requires recreation of the visual vocabulary, complete recomputation of histograms and retraining of classifiers. When adding a new category, current approaches need to fully rebuild the whole recognition system. We address the problem of extensibility by combining class-specific vocabularies with outlier visual words. Classification is achieved by computing a scoring function for each class-specific vocabulary and selecting the highest score value. We show first results of our highly parallelizable and distributable approach on the Caltech 256 dataset.