Animation Part 3: Compression, Animation Pipeline

Animation Part 3

11.8 Compression Techniques

4 bytes * 10 channels * 30 samples/second = 1200 bytes per second per joint , quite large!

channel omisson

• scale channel: 3 to 1, even 0
• quaternions are always normalized, so w in (x, y, z ,w) can be omitted and reconstructed at runtime
• for poses that do not change, the entire animation for that frame can be stored as a single bit, indicating not changing

quantization

• converting a 32-bit IEEE float into an n-bit integer representation
  • for example, 32-bit float is too precise for a quat that lies in [-1, 1]
  • often 16 bit is enough
• encoding & decoding in quantization
  • encoding : float to n-bit integer presentation
  • decoding: opposite
  • decoding only recovers approximation, because quantization is lossy

Sampling frequency and key omission

• three problems causing animation data large
  • up to 10 channels(channel omission helps)
  • large number of joints(hard to solve for high-res character)
  • high sample rate
    • reduce sample rate overall(when 15 fps satisfies, don’t do 30)
    • omit linear changing frames, use LERP to restore

Curved-based compression

• use non-uniform B-splines to approximate each channel curve
  

selective loading and streaming

• load only necessary clips
• keep core clips

11.9 Animation System Architecture

• three distinct layers for a Animation System
• Animation Pipeline
  • generate a single local skeleton pose, and post-process
• Action State Machine (ASM)
  • upper-level animation transitions
• Animation Controller
  • high-level control of animation mode
  • such as drive mode, run and gun mode…

11.10 The Animation Pipeline

• input: animation clips and blend specifications
• output: local and global poses, matrix palette
• stages:
  1. Clip decompression and pose extraction
  2. Pose Blending
  3. Global pose generation
  4. Post-Processing
  5. Recalculating of global poses
  6. Matrix palette generation
• review: local pose and global pose of a joint
  • local joint pose: the joint transformation in parent joint space
  • global joint pose: the joint transformation in model space
  • $P_{j\to M}={\Pi }_{i=j}^0{P}_{i\to p(i)}$, where $p(0)\equiv M$
  • M is model space, $P_{i\to p(i)}$ is transform from local joint i to parent p(i)

Data structure

• shared resources
  • Skeleton: joint hierarchy and its bind pose
  • Skinned meshes
  • Animation Clips
    
• A set of animation clips only apply to one skeleton
  • better use different meshes for the same skeleton to create new characters, so that they can all share a set of animation clips
  • animation retargeting system makes it less strict
• per-instance data
  • each character instance has its own state data
  • including:
    • Clip state
      • Local clock: determines which frame to be extracted
      • Playback rate: speed of the clip
    • Blend specification
      • how to blend multiple clips
    • Partial Skeleton joint weight
    • Local Pose
    • Global Pose
    • Matrix palette

Flat-weighted Average Blend v.s. Blend Tree

pass

Cross Fading Architectures

• cross-fades with a flat weighed average
  • Ramping down A, Ramping up B
  • Complex blends cross fades, from ABC to DE
    • keep relative weights of A,B,C( or D,E) constant
    • ramping up or down ABC, DE as a whole
• Cross Fades with Expression Trees
  • simply introduce a new LERP node at root
    

Animation Pipeline Optimization

• different on most hardware platforms
• main RAM <—> Local cache of processing units
• hUMA: heterogeneous unified memory architecture
• however, optimization techniques can apply to multiple platforms at most times, more or less