Active Learning For Relative Attributes

Visual attributes learning serves as an important role in the field of semantics learning,which gets more and more attentions in recent years. Visual attributes, including binaryattributes and relative attributes, are human-nameable and cross-categorical concepts. Theycan be used to describe the objects or scenes by judging whether certain attributes are exhib-ited in an image or a video. Due to their ability of transferring high-level concepts, visualattributes are academic valuable in many applications including Surveillance, Robotics andso on.Even though large amounts of work have been conducted on visual attributes, theyignore an important problem. In fact, it is common sense that modeling these attributesneeds a mass of labels getting which is time-consuming and labor-consuming, and alsotraining a model with vast samples is intractable in real world applications. Therefore, topromote the time efficiency and space efficiency of the attribute learning system, we needto incorporate active learning algorithms which could mine the most informative samples insuch abundant data.In this work, we first investigate the ranking SVM with similarity model (short forRankSVM-with-Sim) which is used for training relative attributes. Then the analysis oftwo active learning algorithms, i.e. version space reduction and expected model change,shows that they have some limitations on applying for relative attributes due to some specificissues. To solve these problems, we propose an active learning algorithm based on expectedgradient length (EGL) and samples’ diversity (SD). Our algorithm includes two steps, thefirst of which is EGL that assists to actively select optimal samples, while the second is SDthat allows the batch mode active selection.We evaluate the proposed algorithm on three databases including OSR, Pubfig andShoes, on each of which two kinds of experiments, i.e. attributes ranking and imageclassification, are conducted. The experimental results indicate that on average, our activelearner could significantly reduce the number of labeled samples needed which is usuallylarge in passive settings. What’s more important, our algorithm could also boost theaccuracy and robustness of attribute learners.