20-Minute Presentation Outline for "RigNet: Neural Rigging for Articulated Characters"

1. Title Slide (1 slide, 1 minute)

• Content:
  • Title, authors, affiliations.
  • Key visual: Figure 1 from the paper (input mesh, predicted skeleton, deformed examples).
  • Mention supplementary video: https://youtu.be/J30VETgWlDg.

---

2. Overview & Thesis (1 slide, 1.5 minutes)

• Content:
  • Principal thesis: End-to-end neural framework for automated rigging (skeleton + skinning) of diverse 3D characters.
  • Problem: Manual rigging is time-consuming; prior work relies on templates or limited shape classes.
  • Key contributions:
    1. Modular deep architecture (joint prediction, connectivity, skinning).
    2. User control over skeleton granularity.
    3. State-of-the-art results on animator-quality rigs.

---

3. Background & Motivation (2 slides, 3 minutes)

• Slide 1:
  • Traditional rigging: Animator-created skeletons (Figure 2) require anatomical intuition.
  • Prior work limitations:
    • Template-based (Pinocchio) fails on novel structures.
    • Volumetric methods (Xu et al.) lose surface details.
    • Geometric skinning (BBW) lacks anatomical awareness.
• Slide 2:
  • Why neural networks?
    • Learn from large, diverse datasets.
    • Capture animator intuition (joint placement, skinning).
  • Key challenges:
    • Variable joint count/topology.
    • Skeleton-skinning interdependence.

---

4. Method Overview (2 slides, 4 minutes)

• Slide 1: Pipeline diagram (Figure 4).
  • Three stages:
    1. Joint prediction: GMEdgeNet + clustering.
    2. Bone connectivity: BoneNet + MST.
    3. Skinning: Geodesic-aware GNN.
  • Highlight user control (bandwidth slider in Figure 5).
• Slide 2: Key technical innovations.
  • GMEdgeConv: Combines mesh and geodesic neighborhoods.
  • Differentiable clustering: Adaptive joint count.
  • Volumetric geodesic distances for skinning.

---

5. Results & Comparisons (3 slides, 5 minutes)

• Slide 1: Qualitative results (Figure 8, 9).
  • Show skeleton predictions vs. animators/Pinocchio.
  • Skinning error maps (L1 norm).
• Slide 2: Quantitative metrics (Tables 1, 2).
  • Skeleton: CD-J2J (3.9% vs. 7.2% for Pinocchio).
  • Skinning: 82.3% precision vs. 76.3% for NeuroSkinning.
• Slide 3: Generalization (Figure 10).
  • Test cases: Quadrupeds, robots, fictional characters.

---

6. Questions for Authors (1 slide, 2 minutes)

• Content:
  • Why not end-to-end training (joints → skinning)?
  • Handling multi-resolution meshes?
  • Dataset bias: Missing small parts (fingers, clothes) in training.
  • Failure cases (Figure 13): Extra/missing joints.

---

7. Evaluation & Critiques (1 slide, 2 minutes)

• Strengths:
  • Complete solution (skeleton + skinning).
  • Outperforms geometric and learning baselines.
• Weaknesses:
  • Limited user control beyond granularity.
  • Dependency on consistent mesh orientation.
• Significance: Democratizes rigging for non-experts.

---

8. Conclusion & Verdict (1 slide, 1.5 minutes)

• Summary:
  • First learning-based end-to-end rigging framework.
  • Practical for games, films, and crowdsourcing.
• Verdict:
  • Thumbs up for novelty, technical depth, and impact.
  • Future work: Multi-resolution skeletons, physics integration.

---

Total Slides: 12 slides

Time Allocation:

• Technical sections (methods/results): 9 minutes.
• Background/evaluation: 5 minutes.
• Q&A/verdict: 3 minutes.

Visuals to Include:

• Pipeline diagram (Figure 4).
• Comparisons (Figures 8, 9).
• Interactive demo video snippets.
• Failure cases (Figure 13).

Guidelines Met:

• Follows CIS660 structure.
• Balances high-level concepts (SigGraph Recommendations).
• Uses visuals to ground explanations.