ECCV 2020 Daily - Thursday

3 Marcel Geppert 5 difficult part is initialization when you need to estimate the relative pose of the cameras. The line features give only a single constraint mathematically compared to two from a point feature, so you need more images in order to triangulate the point in the first place. Also, you need additional constraints for pose estimation. Line features cross so quickly that it is not really feasible to do a full relative pose estimation. We need four images and I think it is 17 correspondences on all four images. Then running this in RANSAC is just not feasible.” To resolve this, this method assumes that gravity direction is known – which is a reasonable expectation with modern devices – and with gravity direction known, it is much easier to perform initialization because features can be aligned with the gravity. Part of the problem then becomes the 2D version of the relative pose, but by doing a relative pose estimation for three views in 2D, it can upgrade that to full 3D photos with four views. You can view the team’s supplementary video to see an example of a reconstructed point cloud . They try to reconstruct original images from the data and compare that to traditional SfM methods. You will see that it is mostly dynamic content with people walking around who almost completely disappear because it is not possible to reconstruct that part anymore. DAILY T h u r s d a y "This work aims to remove as much of this information as possible, while maintaining the same results…"