Since the preparation of labeled datafor training semantic segmentation networks of pointclouds is a time-consuming process, weakly supervisedapproaches have been introduced to learn fromonly a small fraction of data....Since the preparation of labeled datafor training semantic segmentation networks of pointclouds is a time-consuming process, weakly supervisedapproaches have been introduced to learn fromonly a small fraction of data. These methods aretypically based on learning with contrastive losses whileautomatically deriving per-point pseudo-labels from asparse set of user-annotated labels. In this paper, ourkey observation is that the selection of which samplesto annotate is as important as how these samplesare used for training. Thus, we introduce a methodfor weakly supervised segmentation of 3D scenes thatcombines self-training with active learning. Activelearning selects points for annotation that are likelyto result in improvements to the trained model, whileself-training makes efficient use of the user-providedlabels for learning the model. We demonstrate thatour approach leads to an effective method that providesimprovements in scene segmentation over previouswork and baselines, while requiring only a few userannotations.展开更多
基金supported by Guangdong Natural Science Foundation(2021B1515020085)Shenzhen Science and Technology Program(RCYX20210609103121030)+4 种基金National Natural Science Foundation of China(62322207,61872250,U2001206,U21B2023)Department of Education of Guangdong Province Innovation Team(2022KCXTD025)Shenzhen Science and Technology Innovation Program(JCYJ20210324120213036)the Natural Sciences and Engineering Research Council of Canada(NSERC)Guangdong Laboratory of Artificial Intelligence and Digital Economy(ShenZhen).
文摘Since the preparation of labeled datafor training semantic segmentation networks of pointclouds is a time-consuming process, weakly supervisedapproaches have been introduced to learn fromonly a small fraction of data. These methods aretypically based on learning with contrastive losses whileautomatically deriving per-point pseudo-labels from asparse set of user-annotated labels. In this paper, ourkey observation is that the selection of which samplesto annotate is as important as how these samplesare used for training. Thus, we introduce a methodfor weakly supervised segmentation of 3D scenes thatcombines self-training with active learning. Activelearning selects points for annotation that are likelyto result in improvements to the trained model, whileself-training makes efficient use of the user-providedlabels for learning the model. We demonstrate thatour approach leads to an effective method that providesimprovements in scene segmentation over previouswork and baselines, while requiring only a few userannotations.