← ArchivePaper2016
Position Manipulation Techniques for Facial Animation
The dissertation behind Cetinaslan's blendshape interface papers: sketch driven direct manipulation localized with geodesic circles, a Jacobian transpose solver that stops the pseudo-inverse jumps, and a localized Verlet framework, validated on production facial models.
Abstract
Computer generated 3D facial animation, where humanoid or cartoon character faces are animated to express emotions, plays a tremendous role to enrich entertainment, perception, education and communication. Despite notable advances in the field of computer graphics addressing the 3D character animation, the evolution of the current tools and workflows of professional studios are still cumbersome to create the content for games and movies. As a result, producing high quality facial animation is a laborious, difficult and time-consuming process which requires both technical expertise and artistic talent. Animators widely focus on a set of control structure, which is called rig, to create the appealing expressions by manipulating the 3D virtual face. However, constructing facial animation using the rig requires deep understanding of the mechanics beneath that control structure and it involves long training periods and hard manual effort to carefully locate and manipulate the rig elements to create believable results. This dissertation describes novel methods to ease the facial animation workflow by proposing a set of rig-focused algorithms that are designed to create desired facial expressions practically. Instead of considering the rig as a set of control structure, we approach the facial rig as an artistically designed deformation space to create the required facial deformations for each pose. Therefore, three topics, which lie in the rig deformation space, are investigated. The first method explored in this research addresses the direct manipulation of the facial rig structure using a sketch-based interface. The method allows to manipulate a large number of blendshapes, which is a widely adapted technique for high-quality rigging, simultaneously through a simple stroke in the form of freehand drawing. Although there have been recent efforts on direct manipulation of blendshape models, the final results do not fully satisfy the artists. This is due to the fact that the mathematical frameworks of direct manipulation cause unintuitive and unexpected results with a global deformation impact. Our solution reduces the deformation impact to the local facial geometry by employing the geodesic circles, which allows to gain intuitive and satisfactory results. To this end, we have analyzed the blendshape input data, which includes the vertex coordinates of the target shapes, and reconfigured the content of the input data according to the desired manipulation area. Our method is generic and responds automatically without any manual editing. Inspired by artists' brush painting on canvas, the method provides drawing strokes directly on the 3D model. The second method proposed in the scope of this research is a novel transposition based approach for blendshape direct manipulation. The existing direct manipulation methods employ the pseudo-inverses for the underlying mathematical frameworks. Nevertheless, pseudo-inverse based approaches have a well known instability which causes unexpected facial movements during the artistic editing. Therefore, we present a novel and stable method, inspired by the Jacobian Transpose Method, to overcome the general instability problem of the blendshape direct manipulation with a pin-and-drag interface. The third method deployed in this dissertation provides a novel Verlet integration framework which allows a local physically-based deformation on the arbitrary selected area of the facial model by using position-based constraints. Traditional frameworks deform the models as a whole. The proposed framework allows an interactive selection of the deformation influence area over the model, which improves computational performance for the position manipulation during the simulation. To this end, we take advantage of a geodesic distance computation technique based on heat kernel to determine the selected area. This approach allows producing life-like deformation effects by enhancing the animation especially during the collision handling phase. The applications of the proposed methods to actual data are showcased by executing them to the production quality facial models. We validate our results by comparing them with the results of the existing methods. The comparisons are performed qualitatively and quantitatively that are demonstrated on graphical charts and side by side visual figures. Consequently, the proposed methods in this dissertation aim to reduce time and human effort during facial animation process by preserving the high quality results with stable and practical solutions.
How to read this
- Category
- PhD dissertation (University of Porto, 2016) that gathers Cetinaslan's three blendshape interface methods into one argument.
- Contributions
- The framing that a facial rig is an artistically designed deformation space rather than a pile of controls, which is what licenses editing it by position instead of by slider.
- Sketch driven direct manipulation of blendshapes, localized with geodesic circles so a stroke edits the nearby geometry instead of the whole face, with the blendshape input reconfigured automatically for the chosen area.
- A transposition based direct manipulation solver, inspired by the Jacobian transpose method, that removes the instability of pseudo-inverse formulations under pin-and-drag editing.
- A localized Verlet integration framework with position based constraints, where the deformation influence area is selected interactively through heat kernel geodesic distances, aimed at collision handling on cartoon faces.
- Qualitative and quantitative comparisons against the existing methods on production quality facial models, presented as charts and side by side figures.
- Context
- This is the long form of the papers the archive already lists separately: cetinaslan-sketch-controllers-2015 opened the line, cetinaslan-direct-manipulation-2018 and cetinaslan-sketching-manipulators-2020 carry the sketch thread, cetinaslan-transposition-blendshape-2017 and cetinaslan-jacobian-transpose-2020 the solver thread, and cetinaslan-verlet-facial-2016 the physics one. All of it argues with lewis-direct-manipulation-blendshapes-2010, whose pseudo-inverse is the thing being replaced. Read the thesis when you want the three threads with their shared motivation and full experiments in one place; the individual papers are the citable units.Builds on: Sketch-Based Controllers for Blendshape Facial Animation · Direct Manipulation Blendshapes
- Correctness
- A thesis is examined rather than peer reviewed in the conference sense, so weigh the experiments yourself. The strengths are the ones the papers share: the transposition argument is standard inverse kinematics reasoning and holds, and the geodesic locality is a clean, reproducible construction. The limits are also shared: geodesic radius is a proxy for anatomical influence, the transpose descends toward a constraint rather than satisfying it, and the evaluations are on the authors' own models with no external user study. What the thesis adds over the papers is the quantitative comparison chapter, which is the part to read if you doubt the qualitative figures.
- Clarity
- Well organised and easy to navigate: motivation, three method chapters, evaluation. The English is workmanlike and occasionally awkward, but the mathematics is laid out more fully than the page limited papers allow, which is the reason to open it.
- How to read it
- First pass: the abstract and the introduction's problem statement, then skim the three method chapter openings to see how they relate. Second pass: pick the thread you care about (sketching, solver, or Verlet) and read that chapter with its evaluation section, since each is self contained. Third pass: the comparison chapter as a whole, which is where the thesis earns its keep over the papers. The PDF is open access from the Porto repository.
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 →