Skip to content

← ArchivePaper2020

Stabilized blendshape editing using localized Jacobian transpose descent

Ozan Cetinaslan, Veronica Orvalho

GMODAcademic9 citesFacialRigging

The journal version of the transposition work, adding a hybrid that blends in the pseudo-inverse and a geodesic locality term to keep blendshape weight edits stable.

Abstract

The blendshape approach is the predominant technique for creating high quality facial animation. Facial expressions are generated by altering weight parameters manually for each key-frame by using traditional slider interfaces. However, authoring production quality facial animation with this process requires time-consuming, labor intensive and iterative work for the artists. Direct manipulation interfaces address this problem with a pin-and-drag operation inspired by inverse kinematics approaches. Existing direct manipulation techniques generally make use of the pseudo-inverse of the matrix that contains all the blendshape target shapes' vertex positions. However, pseudo-inverse computation can give unexpected results during facial expression editing, and therefore leaves room for improvement. To this end, we propose an approach to reducing unexpected movements during blendshape weight editing based on using the transpose of the Jacobian matrix rather than the pseudo-inverse. Our approach extracts the deformation directions from the blendshape matrix, and directly maps the sparse constrained point movements to the extracted directions. Furthermore, we offer two additional complementary techniques that build on our proposed transposition algorithm: a hybrid approach that blends the pseudo-inverse algorithm with our transposition algorithm thereby obtaining the benefits of both methods, and a technique that confines the edits to the local region within a geodesic radius on the surface. Our experiments show that, unlike pseudo-inverse based formulations, our transposition algorithm results in smoother and more reliable facial expressions during the weight editing process. The proposed approach improves the fidelity of the generated facial expressions by keeping haphazard movements to a minimum. It is robust, efficient, easy to implement and operates on any blendshape model.

How to read this

Category
Journal extension of the transposition approach to blendshape weight editing.
Contributions
  • Use of the Jacobian transpose instead of the pseudo-inverse to stabilise direct manipulation of blendshape weights.
  • A hybrid that blends the pseudo-inverse and transposition algorithms to get the benefits of both.
  • A locality technique that confines edits to a region inside a geodesic radius on the surface.
  • Experiments reporting smoother and more reliable expressions, with haphazard movement kept to a minimum.
Context
The journal version of cetinaslan-transposition-blendshape-2017, still arguing with the pseudo-inverse formulation in lewis-direct-manipulation-blendshapes-2010. The hybrid is the interesting addition, because it concedes that the pseudo-inverse was solving something the transpose alone does not.Builds on: Transposition Based Blendshape Direct Manipulation
Correctness
The hybrid is both the contribution and the tell. A transpose descends toward a constraint rather than satisfying it, so blending the pseudo-inverse back in is how exactness returns. That means the blend weight is a knob someone has to set, and the paper's guidance on setting it is where to look hard. As with the rest of the line, the evaluation is qualitative on the authors' own models.
Clarity
The clearest paper in the Cetinaslan set. It states the problem, the two candidate solvers, and the compromise, in that order.
How to read it
First pass: abstract and the comparison figures against the pseudo-inverse. Second pass: the transpose derivation and then the hybrid, in that order, since the hybrid only makes sense once you see what the transpose gives up. Third pass: the geodesic locality section, which overlaps with cetinaslan-sketching-manipulators-2020, so read whichever of the two you meet first.

Builds on

Built upon by

Nothing yet.

Related work

Keywords

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →