IBSNet: 用于估计单视角扫描点云交互平分面的神经隐式场

The analysis of spatial relationships between 3D objects is of great significance for scene understanding and interaction. For example, by analyzing the spatial relationship between the robot and the object, the robot can be guided to grasp the object more accurately. By learning the spatial relationship between objects in the real scene, it can guide the generation of virtual scenes that look more natural or better meet the needs of interaction. However, because the single-view scanned point clouds gotten by RGB-D cameras or LiDAR usually have many artifacts and noise, existing methods for analyzing the spatial relationships of objects are often difficult to make accurate predictions when faced with single-view scanned point clouds, which makes these methods impractical for practical applications. For handling the spatial relationship analysis of single-view scanned point clouds, this paper uses the interaction bisector surface(IBS) to express spatial relationships, and proposes a differential unsigned distance field of dual-object to implicitly represent IBS. Inspired by the implicit function learning methods widely used in recent years, this paper designs a neural implicit field to fit the differential unsigned distance field. This neural implicit field takes the single-view scanned point clouds of two objects as input and returns the different unsigned distance field of the two objects. This network uses two multi-layer self-attention point cloud encoders to extract the features of the two input point clouds and combines these features after that. Then these features are inputted into a dual-object unsigned distance decoder to get the unsigned distance value of the query points. Comparative experiments of this method with other methods (Geometry Method, IMNet and Grasping Field) are conducted on the ICON dataset. It simulates single-view scans of each scene from 26 different viewpoints to get the single-view scanned point clouds and split the whole dataset into training set and test set based on a single scene. The robustness of each method is also tested when facing single-view scanning point clouds with different degrees of incompleteness and noise. Experimental results show that the proposed neural implicit field is very robust to the input single-view scanned point clouds with different degrees of incompleteness, and can efficiently predict IBS with accurate shapes.