A Scalable Active Framework for Region Annotation in 3D Shape Collections

摘要

Large repositories of 3D shapes provide valuable input for datadrivenrnanalysis and modeling tools. They are especially powerfulrnonce annotated with semantic information such as salient regionsrnand functional parts. We propose a novel active learning methodrncapable of enriching massive geometric datasets with accurate semanticrnregion annotations. Given a shape collection and a userspecifiedrnregion label our goal is to correctly demarcate the correspondingrnregions with minimal manual work. Our active frameworkrnachieves this goal by cycling between manually annotatingrnthe regions, automatically propagating these annotations across thernrest of the shapes, manually verifying both human and automaticrnannotations, and learning from the verification results to improvernthe automatic propagation algorithm. We use a unified utility functionrnthat explicitly models the time cost of human input acrossrnall steps of our method. This allows us to jointly optimize forrnthe set of models to annotate and for the set of models to verifyrnbased on the predicted impact of these actions on the human efficiency.rnWe demonstrate that incorporating verification of all producedrnlabelings within this unified objective improves both accuracyrnand efficiency of the active learning procedure. We automaticallyrnpropagate human labels across a dynamic shape network usingrna conditional random field (CRF) framework, taking advantagernof global shape-to-shape similarities, local feature similarities, andrnpoint-to-point correspondences. By combining these diverse cuesrnwe achieve higher accuracy than existing alternatives. We validaternour framework on existing benchmarks demonstrating it to be significantlyrnmore efficient at using human input compared to previousrntechniques. We further validate its efficiency and robustness by annotatingrna massive shape dataset, labeling over 93,000 shape parts,rnacross multiple model classes, and providing a labeled part collectionrnmore than one order of magnitude larger than existing ones.