Three-dimensional modeling of an object
The present approach relates to the use of a point cloud of an object to initialize or seed a space carving technique used to generate a 3D model of the object. In one implementation, feature matching is performed on 2D images, with matched features constituting the points of a point cloud model. The point cloud generated in this manner, is one input of a foreground/background segmentation algorithm, which generates a set of segmented 2D images used by a space carving routine to generate a 3D model of the object.
1. A method of generating a 3D model of an object, comprising:
performing feature matching on a set of two-dimensional (2D) images to generate a three dimensional (3D) point cloud model of an object, wherein performing the feature matching on the set of 2D images comprises matching features of the object present in two or more 2D images of the set of 2D images, wherein the features constitute points of the 3D point cloud model;
providing the 3D point cloud model as an input to a segmentation routine;
segmenting foreground and background of the set of 2D images or subsamples of the set of 2D images based on the 3D point cloud model to generate a plurality of 2D segmented images defining silhouettes of the object;
projecting the silhouettes on an initial volume to generate a plurality of projection regions;
performing a space carving operation on the initial volume by excluding volumes outside of the plurality of projection regions from the initial volume to generate the 3D model of the object; and
comparing the 3D model to a prior 3D model of the object or to landmark distances observed in the object to determine a degree of error in the 3D model.
2. The method of claim 1 , wherein segmenting the foreground and the background of the set of 2D images or the subsamples of the set of 2D images outputs respective camera pose estimations for each of the plurality of 2D segmented images.
3. The method of claim 2 , wherein the space carving operation uses the respective camera pose estimations associated with each of the plurality of 2D segmented images to generate the 3D model.
4. The method of claim 1 , wherein the set of 2D images is acquired by a drone executing a flight plan about the object.
5. The method of claim 1 , wherein the object is an industrial asset, a commercial asset, or infrastructure.
6. The method of claim 1 , wherein the 3D model comprises a water-tight convex hull representation of the object.
7. The method of claim 1 , further comprising:
comparing the 3D model to the prior 3D model to identify surface or structural irregularities.
8. The method of claim 1 , wherein the space carving operation is provided with a respective camera pose estimation for each of the plurality of 2D segmented images.
9. The method of claim 1 , wherein determining the degree of error in the 3D model comprises comparing the landmark distances observed in the object to corresponding landmark distances in the 3D model to determine a scale factor between the 3D model and the object.
10. The method of claim 9 , wherein the degree of error is a root mean square error, and wherein the root mean square error is determined based on a root mean square error calculation between the landmark distances observed in the object and the corresponding landmark distances in the 3D model.
11. The method of claim 6 , wherein a surface of the water-tight convex hull representation is colored mapped.