Qualitative Comparisons

We perform qualitative comparisons under different settings.

(a) Joint training of the canonical model and deformation network from a dynamic sequence;
(b) First training the canonical model with a static sequence, then the deform network with a dynamic sequence;
(c) Estimating global transformations between the two stages in (b);
(d) Progressively learning both global and local deformations for each frame in an online manner;
(e) Incorporating additional regularizations into (d).

Quantitative Comparisons

Ablation Study

Here, we show an ablation on the decomposition of global and local deformation.

We also conclude that learning in a progressive manner improves the learning of global transformation as in accompanying figure.

Our experiments show the improvement in visual results upon applying regularization.