Click to Move: Controlling Video Generation with Sparse Motion: Supplementary material
Pierfrancesco Ardino, Marco De Nadai, Bruno Lepri, Elisa Ricci and Stéphane Lathuilière
This page shows the results obtained with different methods on the Cityscapes dataset [1] and KITTI 360 dataset [2] (please refer to the main paper for the description of each method). The position of the moved object at t=0 is highlighted in red in every frame. We observe that in all sequences, the other methods cannot move the objects in the scene. The competitor approaches correctly handle only the ego-motion since the object always remains in its initial position. We also observe that the images generated by the S. Sheng* method suffer from artefacts (see Sequence 1). On the contrary, Click to Move (C2M) successfully moves the object of interest from its initial position. Note that, towards the end of the sequences, the other approaches generate blurry frames while C2M can predict frames with a higher level of details.
References:
[1] Marius Cordts, Mohamed Omran, Sebastian Ramos, TimoRehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke,Stefan Roth, and Bernt Schiele. The cityscapes dataset forsemantic urban scene understanding. In CVPR, 2016.
[2] Jun Xie, Martin Kiefel, Ming-Ting Sun, and Andreas Geiger.Semantic instance annotation of street scenes by 3d to 2d labeltransfer. InCVPR, 2016
[3] Lu Sheng, Junting Pan, Jiaming Guo, Jing Shao, and Chen Change Loy. High-quality video generation from staticstructural annotations.International Journal of Computer Vi-sion, 128:2552–2569, 2020.
Qualitative comparisons
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
