Reconstruction of 3D objects from a single view image is generally an ill-posed problem because of the projection distortion. A 3D approximation method of an object in a single image to a 3D minimum bounding box is proposed in this paper, which works ...
Reconstruction of 3D objects from a single view image is generally an ill-posed problem because of the projection distortion. A 3D approximation method of an object in a single image to a 3D minimum bounding box is proposed in this paper, which works automatically when the object is on the ground plane. However the method does not need any prior information about the object. In this method, an object in an image is first back-projected on the ground plane and then a parallelepiped enough to enclose the corresponding 3D object is estimated based on size of the back-projected shape. Then the parallelepiped is little by little diminished as long as its image does not encroach upon area of the object in the image plane. The minimized parallelepiped called 3D minimum bounding box can be utilized to localize the 3D object on the ground plane and to roughly estimate its volume. However the proposed method tends to output smaller bounding boxes for round-shaped objects than their real minimum bounding boxes. Experimental results on real image objects are shown to demonstrate usefulness of our approach.