Learning robust color name models from web images

We use images that have been collected using an Internet search engine to train color name models for color naming and recognition tasks. Considering color histogram bands as being words of an image and the color names as classes, we use the supervised latent Dirichlet allocation to train our model. To pre-process the training data, we use state-of the art salient object detection and a Kullback-Leibler divergence based outlier detection. In summary, we achieve state-of-the-art performance on the eBay data set and improve the similarity between labels assigned by our model and human observers by approximately 14%.