AnimationBlending

Animation

Written with StackEdit.

Code Structure for Animation

struct JointPose {
	Quaternion m_rot; //q
	Vector3 m_trans; //t
	F32 m_scale; //s
}

struct AnimationSample{
	JointPose* m_aJointPose;
}

struct AnimationClip{
	Skeleton* m_pSkeleton; //指向骨骼
	F32 m_framePerSecond;
	U32 m_frameCount; 
	AnimationSample* m_aSamples; //每帧的数据
	bool m_isLooping;
}

len(m_aSamples) is m_frameCount+1 when m_isLooping == False, else m_frameCount;

Animation Retargeting

• Animation is usually compatible with only one skeleton
  • But some closely related skeletons can use these animations
  • Only if engine supports ignoring animation channels that cannot be found in skeleton being animated
  • Advanced research on retargeting animations on totally different skeletons

Continuous Channel Functions

• According SQT data format for each sample,
  • 10 scalar-valued functions per joint
  • the time continuous functions per joint is a channel
  • channel functions should be smooth and continuous across the entire animation clip
  • many engines LERP bettwen the samples, making it piecewise linear approximations to the underlying continuous functions

Metachannels

• channels that do not have to do with posing
  • but used to be synchronized with the animation
• event trigger channel
  • sound, particle effects and so on
• animated locator
  • special joint
  • such as camera position, orientation, fov…
• texture coordinates scrolling
• texture animation
• animated parameter
  • material
  • lighting

Skinning and Matrix Palette Generation

Attaching the vertices of a 3D mesh to a posed skeleton is call skinning

Per-Vertex Skinning Information

• a skinned mesh is attached to a skeleton by means of its vertices.
• each vertex can be attached to one or multiple joints
  • if multiple, the position becomes weighted average of the positions that it would have assumed had it bound to each joint independently
  • so each vertex information should including
    • the index of joints to which it is bound
    • the weighting factor for each joint
  • the maximal no of bound joints is usually 4
• data structure:

  struct Skinned Vertex{
  	float m_position[3];
  	float m_normal[3];
  	float m_u, m_v; //u,v for texture cordinates
  	U8 m_jointIndex[4];
  	float m_jointWeight[3];
  }
  

Mathematics of Skinning

• Skinning matrix: transform vertices from original position in bind pose to current position that correspond to the current pose.
  • always in model space
• One joint example
  • $B_{j\to M}$ : bind pose of the joint j in model space
  • $C_{j\to M}$: current pose of the joint in the model space
  • $V_j$: joint space position of current vertex
  • K_j = (B${}_{j\to M}$)^-1C${}_{j\to M}$
  • V${}_M^C$ = $V_j$ $C_{j\to M}$
    = $V_M^B(B_{j\to M}{)}^{-1}{C}_{j\to M}$
    = V${}_M^B$K${}_j$
  • usually $B_{j\to M}$ is cached in skeleton, $Cj\to M$ calculated using Local Joint Space to Model Space Equation :
    $P_{j\to M}={\Pi }_{i=j}^0{P}_{i\to p(i)}$
• Multiple Joints:
  • for each pose, (K₁, …, K_j) is called matrix palette
  • $K_{j_i}$ is transform matrix for joint i
  • V${}_M^C$ = ${\Sigma }_{i=0}^{N-1}$ w_iV${}_M^B$K${}_{j_i}$

Animation Blending

more than 1 animation clip contribute to the final pose of character

LERP Blending

• (P_LERP)_j = LERP((P_A)_j, (P_B)_j, $\beta$) = (1 - $\beta$)(P_A)_j + $\beta$(P_B)_j
• P${}_{skel}^{LERP}$ = { (P_LERP)_j } |${}_{j=0}^{N-1}$
• $\beta$ is called the blending percentage or blend factor
• SQT format data is easy to blend, not as 4X4 matrix
• T and S: vector LERP, Q: LERP or SLERP
• independently on each joint, parallel calculated

Applications of LERP Blending

• Temporal Blending
  • finding intermediate poses between sampled poses
• Motion continuity: Cross Fading
  • clip transitions may pop, not smooth
  • C0 continuity: path trace smooth
  • C1 continuity: velocity smooth
  • C2 continuity: second derivatives of the motion path smooth
  • C3 and higher …
  • LERP blending used in transitions to achieve C0 or C1 continuity is called cross fading
  • overlap time, $\beta$ from 0 to 1
  • types of cross fading
    • smooth transition: Clip A and B play simultaneously, $\beta$ changes over time
    • Frozen transition: A stops and B gradually takes over
      • usually used in unrelated clips
    • blend factor $\beta$ functions
      • linear, cubic
      • ease-in, ease-out
    • Core poses
      • every clip starts and ends in these poses
      • C0 continuity: ensure two core poses match
      • C1 continuity: authoring a single smooth animation into A end and B start

Directional Locomotion

pivotal movement: always face to where it turns
targeted movement: moving sideways, backward, forward, keeping its face to a target

• targeted movement
  • forward, strafe left, strafe right
  • LERP blend between two adjacent clips, $\beta$ according to angle
  • backward not included because of legs crossing each other
    • one feasible approach: use to hemispherical blends, one for forward and another for backward; explicit transition animation is played between hemisphere switch
• pivotal movement
  • always play forward clips
  • slightly leaning sideways when turning
  • author three clips: move straight forward, extremely left turn, extremely right turn
  • LERP blending between them to create desired turn pose

Complex LERP Blends

prepackage certain commonly used blends for ease of use

• generalized one-dimensional LERP blending
  

• Simple two-dimensional LERP blending

  • (b_x, b_y), 4 corner clips A, B, C, D
  • two steps
    • use b_x to blend between A and B to B_AB, C and D to B_CD
    • use b_y to blend B_AB and B_CD
      

• Triangular two-dimensional LERP blending

  • blend between arbitrary number of clips positioned at arbitrary positions
  • (P_LERP)_j = $\alpha$(P₀)_j + $\beta$(P₁)_j + $\gamma$(P₂)_j
  • $\alpha$ + $\beta$ + $\gamma$ = 1
  • barycentric coordinates b
    • b = $\alpha$b₀ + $\beta$b₁ + $\gamma$b₂

• generalized two-dimensional LERP blending

  • Delaunay triangul

Partial-Skeleton Blending

wave hands when walking, a partial blocking is required

• the set of blending percentages: {$\beta$_j}|${}_{j=0}^{N-1}$
• in LERP, $\beta$_j=0 can be used to mask out certain joints
• for the walking-wave-hands example
  • mask out joints on except right arm
• problems:
  • abrupt change: gradually change $\beta$ instead of abrupt change
  • looking not natural because human body movements are never independent, but partial-skeleton blending makes the two partial movements independent of each other
    • in the walking-wave-hands example, human should look slightly different when run/walk/stand waving hands
    • so we turn to additive blending

Additive Blending

• difference clip: difference between two clips
  • also called additive animation clip
• D = S - R
  • D: difference clip, S: source, R: reference
  • R + $\beta$D: partial add
    *$\grave{S}=\grave{R}+D$ can create new clip, with S appearing naturally blended on $\grave{R}$ (for example ,run is R, walk is $\grave{R}$, S is run tiredly, then $\grave{S}$ is walking tiredly, D is tired)
• Mathematical Formulation
  • $D_j=S_j{R}_j^{-1}$, for joint j
    • $D_j$ is difference pose
    • $R_j$ is reference pose
    • $R_j$ is source pose
    • matrix subtraction is multiplying inverse
  • $A_j=D_j{T}_j=(S_j{R}_j^{-1})T_j$
    • $A_j$ is new additive pose
    • $T_j$ is the target pose we want to add on
• Diffrence clip can alse be temporal LERP
• Additive blend percentage
  • for example, half-tired
  • LERP as usual
  • $A_j=LERP(T_j,D_j{T}_j,\beta )=(1-\beta )(T_j)+\beta (D_j{T}_j)$
• Additive v.s. Partial Blending
  • additive blending looks more natural
  • limitations of additive blending
    • over rotate of joint when applying multiple additive clips
    • carefully choose reference clip and target clip
      • keep hip rotations to a minimum in the reference clip
      • shoulder and elbow joints should be in neutral poses in the reference clip
      • new difference clip should be created for each core pose
• Applications for additive blending
  • stance variation: difference clip is each stance
  • locomotion noise: add randomness to locomotion movements
  • Aim and Look-At: add partial difference to aim partially
  • overloading the time axis: use LERP to blend between difference clips