Computer Vision News - June 2020

2 Summary Research 16 - Baseline : simple model with no occlusion mask, no first order Taylor expansions, supervised with at the highest resolution only; - Pyr .: the pyramid loss is added; - Pyr. + : the occlusion aware network is added; - Jac.w/oequivarianceconstraints : local affine transformations are added; - Full : with equivariance constraints; In the image on the side, we can observe that results obtained by employing the full pipeline proposed are substantially better. are reported below, showing clearly how the first motion model method referred to as Ours, in the figure, outperforms the other two. The quantitative results (not shown) also prove superiority of this method with respect to the others according to all four metrics. ℒ ← This is further confirmed by the quantitative evaluation in the table. The second experiment was a comparison with the state-of-the-art for video reconstruction. Two qualitative examples

RkJQdWJsaXNoZXIy NTc3NzU=