Structured Output Prediction with Hierarchical Loss Functions for Seafloor Imagery Taxonomic Categorization

In this paper we study the challenging problem of seafloor imagery taxonomic categorization. Our contribution is threefold. First, we demonstrate that this task can be elegantly translated into a Structured SVM learning framework. Second, we introduce a taxonomic loss function in the structured output classification objective during learning that is shown to improve the performance over other loss functions. And third, we show how the Structured SVM can naturally deal with the problem of learning from data imbalance by scaling the cost of misclassification during the optimization. We present a thorough experimental evaluation using the challenging and publicly available Tasmania Coral Point Count dataset, where our models drastically outperform the stateof-the-art-results reported.