3d representation of physical environment objects

Abstract

Various implementations provide 3D representations of objects. Such representations may be based on 3D point cloud and/or 2D image inputs that are obtained based on sensor data, e.g., images, depth data, motion data, etc. 3D point cloud input may be used for part segmentation and/or to determine position and/or orientation of object parts, e.g., generating 3D bounding boxes representing the sizes, positions, and orientations, of object parts. 2D image input may be used for part attribute recognition, e.g., to determine whether a chair legs part has a particular type such as star-shaped, straight down, crossed-shaped, etc. Part attributes may be used to produce a relatively simple and relatively accurate representation of the shape of each part within a respective area, e.g., within a bounding box determined for each part using the 3D point cloud input.