On the sampling of web images for learning visual concept classifiers

Visual concept learning often requires a large set of training images. In practice, nevertheless, acquiring noise-free training labels with sufficient positive examples is always expensive. A plausible solution for training data collection is by sampling the largely available user-tagged images from social media websites. With the general belief that the probability of correct tagging is higher than that of incorrect tagging, such a solution often sounds feasible, though is not without challenges. First, user-tags can be subjective and, to certain extent, are ambiguous. For instance, an image tagged with "whales" may be simply a picture about ocean museum. Learning concept "whales" with such training samples will not be effective. Second, user-tags can be overly abbreviated. For instance, an image about concept "wedding" may be tagged with "love" or simply the couple's names. As a result, crawling sufficient positive training examples is difficult. This paper empirically studies the impact of exploiting the tagged images towards concept learning, investigating the issue of how the quality of pseudo training images affects concept detection performance. In addition, we propose a simple approach, named semantic field, for predicting the relevance between a target concept and the tag list associated with the images. Specifically, the relevance is determined through concept-tag co-occurrence by exploring external sources such as WordNet and Wikipedia. The proposed approach is shown to be effective in selecting pseudo training examples, exhibiting better performance in concept learning than other approaches such as those based on keyword sampling and tag voting.