Abstract
High-level shape understanding and technique evaluation on large repositoriesof 3D shapes often benefit from additional information known about the shapes.One example of such information is the semantic segmentation of a shape intofunctional or meaningful parts. Generating accurate segmentations withmeaningful segment boundaries is, however, a costly process, typicallyrequiring large amounts of user time to achieve high quality results. In thispaper we present an active learning framework for large dataset segmentation,which iteratively provides the user with new predictions by training new modelsbased on already segmented shapes. Our proposed pipeline consists of threenovel components. First, we a propose a fast and relatively accuratefeature-based deep learning model to provide dataset-wide segmentationpredictions. Second, we propose an information theory measure to estimate theprediction quality and for ordering subsequent fast and meaningful shapeselection. Our experiments show that such suggestive ordering helps reduceusers time and effort, produce high quality predictions, and construct a modelthat generalizes well. Finally, we provide effective segmentation refinementfeatures to help the user quickly correct any incorrect predictions. We showthat our framework is more accurate and in general more efficient thanstate-of-the-art, for massive dataset segmentation with while also providingconsistent segment boundaries.