GF5 GF5: Animating 3D Characters

Part 2

Part 2: Skinning Weights, One-Hot Binding, and SMPL

Previous: Part 1 Next: Interim

Why This Part Matters

In Part 1, each block was attached rigidly to one joint. That was useful for learning FK, but it is not how real character surfaces behave.

A deformable mesh bends because vertices near a joint are influenced by more than one bone. Part 2 studies that idea directly using SMPL, a standard human body model with a realistic mesh and skinning weights.

The key question for this part is:

How much of the final animation comes from the skeleton pose, and how much comes from the way the mesh is attached to that skeleton?

Learning Goals

By the end of Part 2, you should be able to:

  • explain what a skinning weight means
  • derive one-hot skinning from a full weight matrix
  • explain why one-hot attachment creates artifacts near elbows, shoulders, hips, and knees
  • compare one-hot binding against linear blend skinning (LBS)
  • reuse the same motion on both the toy character and SMPL

Minimal Technical Background

Rest Pose

The rest pose is the default unposed mesh and skeleton.

Skinning starts with:

  • a rest-pose mesh
  • a skeleton
  • a set of per-vertex weights

Skinning Weights

For each vertex v_i, the model stores weights w_ij telling us how strongly joint j influences that vertex.

The weights are usually non-negative and sum to 1.

One-Hot Skinning

One-hot skinning is the simplest baseline:

  • find the joint with the largest weight for each vertex
  • set that joint's weight to 1
  • set all other weights to 0

This creates rigid piecewise motion. It is easy to implement and easy to interpret, which is why it is a good first baseline.

Linear Blend Skinning

In LBS, a posed vertex is a weighted sum of the transforms from several joints:

v'_i = sum_j w_ij T_j v_i

You do not need to derive the full production form of LBS in this part, but you do need to understand the idea: vertices near joints are shared across influences instead of being assigned to a single bone.

SMPL

SMPL is a standard articulated human body model used widely in graphics and vision. In this project it gives us:

  • a human mesh
  • a standard skeleton
  • a realistic set of skinning weights

That makes it a good bridge from the simple Part-1 rig to a more realistic animation pipeline.

Code Map

The main files for this part are:

  • viewer/asset_viewer.py
  • viewer/smpl_support.py
  • viewer/student_submission/part2_skinning.py
  • viewer/skeleton_profiles.py

The viewer already supports:

  • loading SMPL
  • visualising one-hot and LBS results
  • visualising skinning weights
  • exporting videos
  • reusing saved motion clips from Part 1

Part 2 Tasks

Task 1: Load SMPL and Reuse Your Part-1 Motion

Take the motion you created in Part 1 and run it on:

  • one block character
  • SMPL

This is important: the motion should stay the same. What changes now is the character model and the mesh attachment.

Task 2: Implement SMPL One-Hot Skinning

Implement the one-hot version of skinning for SMPL.

Requirements:

  • start from the provided SMPL weight matrix
  • assign each vertex to the single joint with the maximum weight
  • do not hardcode a precomputed one-hot mesh file

This task should produce a deliberately limited baseline that is useful for comparison.

Task 3: Implement SMPL Linear Blend Skinning

Implement linear blend skinning for SMPL using the provided weight matrix and posed joint transforms.

Requirements:

  • use the provided SMPL weights directly
  • compute the posed vertex positions from the weighted combination of joint transforms
  • do not use a hidden pre-written skinning helper in the student version

You are not being asked to rebuild the full SMPL model. The task is the skinning stage only.

Task 4: Compare One-Hot Against LBS

Use the viewer to compare:

  • one-hot SMPL
  • LBS SMPL

Use the same motion on both.

You should examine at least two stress regions, for example:

  • elbow
  • shoulder
  • knee
  • hip

You should also inspect skinning-weight visualisations for at least two joints.

Task 5: Produce Evidence, Not Just Code

By the end of Part 2, you should have concrete results, not just an implementation.

At minimum, collect:

  • one short video of your custom motion on the toy rig
  • one short video showing the same motion on SMPL
  • side-by-side comparisons of one-hot and LBS on at least two challenging poses
  • at least two weight visualisations for selected joints

The interim checkpoint is specified separately in the Interim Report handout.

Self-Check Questions

These questions are for your own understanding and revision. They are not required report questions.

If you can answer them clearly, you are probably on the right track:

  • Why is one-hot attachment a useful baseline even though it looks worse?
  • Why do artifacts appear mainly near joints rather than in the middle of a limb segment?
  • Why can the same motion look acceptable on the toy rig but poor on one-hot SMPL?
  • Why are skinning weights properties of the character model, not of the motion clip?

What To Bring To Help Sessions

Bring one specific artifact or failure case, not just a general complaint.

Good examples:

  • "the elbow collapses in one-hot mode but looks smoother with LBS"
  • "my one-hot assignment seems to use the wrong dominant joint near the shoulder"
  • "the motion transfers correctly, but the deformation near the knee looks too rigid"

Those are much easier to diagnose than "SMPL looks wrong".