**4. Conclusions**

A 3D abstraction method receives as input the camera intrinsic parameters and several pictures of the scene. There are two different approaches: The first one does not require the camera extrinsics estimated from an external SfM pipeline, nor the Ground Truth camera poses [2]. It sources the line correspondences from a line matching method, and is able to generate 3D sketches from sets of pictures. This

kind of approaches get an edge against datasets with low number of images, or when these present corrupted texture, blurring, and low definition images where the feature point descriptor fails to detect a fair number of keypoints. The reduced number of correspondences limit the thickness of the point cloud generated by the SfM pipelines, and therefore the accuracy of the estimated camera extrinsics. With inaccurate estimations for the cameras, exploiting homography constraints is not adequate to source line correspondences. Oppositely, [2] is able to reconstruct simple line-based sketches with fair precision and number of lines. It required lower number of images to obtain more complete abstractions than method [1]. The range of scenarios where it is advantageous to use method [1] for 3D abstraction includes sets of pictures of simple objects, with low texture, poor illumination, low resolution, blurring or under other conditions that make difficult the success of a point based algorithm. In these scenarios it outperforms the competition in terms of quantity of lines, precision and completeness of the abstraction. Another conclusion is that camera extrinsics are unavoidably required for 3D abstractions featuring many lines, because the estimation for the camera poses will not be accurate if the line matching method returns matching outliers or line fragmentation.

On the other hand, for datasets with moderate number of images, which clear textures, the second approach can be profitable. In this case, the geometric relations from the related points among the images will permit the feature point based pipeline to generate a moderately dense 3D point cloud. In this case, the poses of the cameras obtained by the point based pipeline can be trusted, and used as basis for line matching and linear projection to generate the 3D sketch. The results obtained with method [1] with datasets of hundreds of images are very good. An abstraction using this method will team perfectly with a dense reconstruction.

Both approaches are valid for their range of applications. The first one is valid for difficult datasets with noise and low number of images. The second approach will shine with datasets with high texture and many pictures, because it will will profit of the high precision obtained from feature point based 3D reconstruction pipelines for locating the cameras.
