Joint Image and 3D Shape Part Representation in Large Collections for Object Blending

Abstract

We propose a new approach to perform object shape retrieval from images, it can handle the shape of the part of the object and combine parts from different sources to find a different 3D shape. Our method creates a common representation for images and 3D models that enables mixing elements from both kinds of inputs. Our approach automatically extracts the desired part and its 3D shape from each source without the need of annotations. There are many applications to combining parts from images and 3D models, for example, performing smart online catalogue searches by selecting the parts that we are looking for from images or 3D models and retrieve a 3D shape that has the desired arrangement of parts. Our approach is capable of obtaining the shape of the parts of an object from an image in the wild, independently of the pose of the object and without the need of annotations of any kind.

Keywords

• Shape blending
• joint image and shape embedding
• 3D shape
• computer vision
• computer graphics