Hierarchical cluster guided labeling: efficient label collection for visual classification

Abstract

Visual classification is a core component in many visually intelligent systems. For example, recognition of objects and terrains provides perception during path planning and navigation tasks performed by autonomous agents. Supervised visual classifiers are typically trained with large sets of images to yield high classification performance. Although the collection of raw training data is easy, the required human effort to assign labels to this data is time consuming. This is particularly problematic in real-world applications with limited labeling time and resources. Techniques have emerged that are designed to help alleviate the labeling workload but suffer from several shortcomings. First, they do not generalize well to domains with limited a priori knowledge. Second, efficiency is achieved at the cost of collecting significant label noise which inhibits classifier learning or requires additional effort to remove. Finally, they introduce high latency between labeling queries, restricting real-world feasibility. This thesis addresses these shortcomings with unsupervised learning that exploits the hierarchical nature of feature patterns and semantic labels in visual data. Our hierarchical cluster guided labeling (HCGL) framework introduces a novel evaluation of hierarchical groupings to identify the most interesting changes in feature patterns. These changes help localize group selection in the hierarchy to discover and label a spectrum of visual semantics found in the data. We show that employing majority group-based labeling after selection allows HCGL to balance efficiency and label accuracy, yielding higher performing classifiers than other techniques with respect to labeling effort. Finally, we demonstrate the real-world feasibility of our labeling framework by quickly training high performing visual classifiers that aid in successful mobile robot path planning and navigation.