Root Pose Decomposition Towards Generic Non-rigid Reconstruction

Anonymous
It might take several minutes to load videos. If videos fail to display, please directly download our repo: https://github.com/rpd-share/rpd-share.github.io

More videos are coming. For each video, there are six sub-videos.
Top left: reference image overlayed with input densepose features. Top middle: reconstructed 1st frame shape. Top right: recovered articulations in the canoincal space. Bottom row: reconstruction from front/side/top viewpoints. Correspondences are shown as the same color.



Qualitative results of occluded cats (OVIS dataset)



Learned object of the canonical space
(zoom in by scrolling; change views by dragging the object)


Qualitative results of humans (DAVIS dataset)



Results of sharing the canonical model for an obviously different body size



Learned object of the canonical space
(zoom in by scrolling; change views by dragging the object)

Qualitative results of Hands


Learned object of the canonical space
(zoom in by scrolling; change views by dragging the object)


Qualitative results of chicken (OVIS dataset)

This experiment aims to evaluate RPD when objects quickly changing the distance to camera
Reconstruction by a single 7-second video

Learned object of the canonical space
(zoom in by scrolling; change views by dragging the object)


Qualitative results of duck (OVIS dataset)

Learned object (smoothed) of the canonical space
(zoom in by scrolling; change views by dragging the object)


Failure case

Rapid pose change at the 3rd second leads to an ambiguous pose estimation