Abstract: To allow the robot understand the scene while scanning and reconstructing an unknown scene,objects must be segmented and recognized using current scanning data,which requires to solve the partial matching problem based on incomplete point cloud.Existing methods still suffer from low matching accuracy and high computation cost.To this end,we propose a novel multi-level partial matching method.We use Bag of Words framework to reduce the feature dimension and accelerate computation in a coarse level and filter the correspondence between feature points to improve matching precision in a fine level.Specifically,our approach first leverages the multi-scale SVM method to cluster the feature points of models in the dataset.Then,a spatially-sensitive Bag of Words method is used to retrieve candidate models from the dataset.Finally,a partial matching method based on both global and local isometry is applied to filter the feature correspondence between query point cloud and the candidate model.A variety of evaluations and comparisons have shown the feasibility and efficiency of our approach.The proposed method provides a meaningful reference for automatic robotic scene scanning,analysis,and reconstruction and will stimulate other future works.