Visual Object Categorization Based on Hierarchical Shape Motifs Learned From Noisy Point Cloud Decompositions

摘要: Object shape is a key cue that contributes to the semantic understanding of objects. In this work we focus on categorization real-world object point clouds particular types. Therein surface description and representation structure have significant influence accuracy, when dealing with scenes featuring noisy, partial occluded observations. An unsupervised hierarchical learning procedure utilized here symbolically describe characteristics multiple levels. Furthermore, constellation model proposed hierarchically decomposes The decompositions are described as constellations symbols (shape motifs) in gradual order, hence reflecting from local global, i.e., parts over groups entire combination multi-level surfaces decomposition shapes leads which allows conceptualize shapes. discrimination has been observed experiments seven categories instances sensor noise, occlusions well inter-category intra-category similarities. Experiments include evaluation approach, comparisons Fast Point Feature Histograms, Vocabulary Tree neural network-based Deep Learning method. conducted alternative datasets analyze generalization capability approach.