Hedging your bets: Optimizing accuracy-specificity trade-offs in large scale visual recognition

摘要: As visual recognition scales up to ever larger numbers of categories, maintaining high accuracy is increasingly difficult. In this work, we study the problem optimizing accuracy-specificity trade-offs in large scale recognition, motivated by observation that object categories form a semantic hierarchy consisting many levels abstraction. A classifier can select appropriate level, trading off specificity for case uncertainty. By trade-off, obtain classifiers try be as specific possible while guaranteeing an arbitrarily accuracy. We formulate maximizing information gain ensuring fixed, small error rate with hierarchy. propose Dual Accuracy Reward Trade-off Search (DARTS) algorithm and prove that, under practical conditions, it converges optimal solution. Experiments demonstrate effectiveness our on datasets ranging from 65 over 10,000 categories.