myshi commited on
Commit
faea09f
·
verified ·
1 Parent(s): 90aeffe

README: retargeting = measurement + optimisation

Browse files
Files changed (2) hide show
  1. README.md +47 -18
  2. assets/lean_cadence.gif +3 -0
README.md CHANGED
@@ -63,6 +63,50 @@ width. All rows are data from this dataset.</em></p>
63
 
64
  ## 2. Retargeting
65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
  **Body model.** Each target body is represented by its SMPL-X skeleton and its rest-pose mesh, partitioned by the
67
  skinning weights into fifteen rigid blocks (pelvis, upper torso, head, thighs, shanks, feet, upper arms, forearms and
68
  **hands**), each with a signed-distance field, a mass and a centre of mass. Foot vertices define the sole height,
@@ -108,24 +152,9 @@ clip and every body:
108
  | lean | trunk lean not following the body | (θ_f − θ_f*)² + (θ_s − θ_s*)² on the mean forward / side lean (deg) | 0.3 |
109
 
110
  The penetration allowance a is the largest of 1 cm, the 95th-percentile penetration of the same block pair on the
111
- source body, and the rest-pose overlap on the target body. The lean targets shift the copy's mean forward and side
112
- trunk lean by k_f·Δw and k_s·Δw for the hip-width difference Δw between target and source,
113
- with k_f = 150° and k_s = 65° per metre, the two body associations measured in MotionPersona.
114
-
115
- <p align="center">
116
- <img src="assets/lean_associations.png" width="100%" alt="Adult performers' effects on forward and side trunk lean against hip width, with quartile means, population fit and pair slopes">
117
- </p>
118
- <p align="center"><em>The two associations assigned to body shape, measured on the captured data. Grey: each adult
119
- participant's effect on trunk lean (the participant term of a participant × context decomposition) against hip
120
- width; blue: hip-width quartile means ± SE; dashed: fit over all 39 adults with bootstrap 95 % band. Wider hips go
121
- with more forward lean (+4.5° per 2 cm over all adults, +2.6° within the 12 metadata-matched pairs that share all
122
- three persona attributes but differ in body) and more side lean (+1.2° and +1.5° per 2 cm).</em></p>
123
-
124
- The retargeting targets of 150° and 65° per metre lie between the within-pair and the population slopes and inside
125
- the 95 % intervals of the pair slopes (41–249° and 19–135° per metre). These two are the only body associations that
126
- survived the tests in the accompanying analysis (an age control, a consistent sign in all nine contexts, and a
127
- directional check on the matched pairs); all other gait descriptors, such as cadence and leg-normalised step length,
128
- are preserved from the source rather than adapted.
129
 
130
  <p align="center">
131
  <img src="assets/terms.gif" width="100%" alt="Weighted residual of each objective term over 400 iterations, with quality checks">
 
63
 
64
  ## 2. Retargeting
65
 
66
+ Retargeting a clip onto another body requires a definition of *the same motion on that body*: which properties of
67
+ the source are kept, and which change with the body. MotionPersonaX takes this definition from measurements on the
68
+ captured data (Section 2.1) and realises it with an explicit optimisation (Section 2.2).
69
+
70
+ ### 2.1 Measurement: what changes with the body
71
+
72
+ Because every participant performs the same grid of contexts and locomotion types, each of 27 gait descriptors
73
+ (timing, space, limbs and trunk; lengths normalised by leg length, hip width or shoulder width) can be decomposed into a
74
+ participant effect, a context effect and their interaction. The participant effect is large and stable: it exceeds the
75
+ context effect by an order of magnitude for speed, jerk and sway, with split-half reliability of 0.57–0.96. Body
76
+ measures, however, explain only a small part of it. Each participant effect is regressed on nine body measures with
77
+ leave-one-participant-out prediction, and an association is attributed to the body only if it survives an age
78
+ control, keeps its sign after partialling out age, has the same sign in all nine contexts, and is confirmed in
79
+ direction by twelve metadata-matched adult pairs that share all three persona attributes but differ in body.
80
+
81
+ Two associations pass, both on trunk lean: wider hips go with more forward lean and more side lean.
82
+
83
+ <p align="center">
84
+ <img src="assets/lean_associations.png" width="100%" alt="Adult performers' effects on forward and side trunk lean against hip width, with quartile means, population fit and pair slopes">
85
+ </p>
86
+ <p align="center"><em>The two associations assigned to body shape, measured on the captured data. Grey: each adult
87
+ participant's effect on trunk lean (the participant term of a participant × context decomposition) against hip
88
+ width; blue: hip-width quartile means ± SE; dashed: fit over all 39 adults with bootstrap 95 % band. Wider hips go
89
+ with more forward lean (+4.5° per 2 cm over all adults, +2.6° within the 12 metadata-matched pairs that share all
90
+ three persona attributes but differ in body) and more side lean (+1.2° and +1.5° per 2 cm).</em></p>
91
+
92
+ The retargeter implements these two associations with slopes of k_f = 150° and k_s = 65° per metre of hip width,
93
+ between the within-pair and the population estimates and inside the 95 % intervals of the pair slopes (41–249° and
94
+ 19–135° per metre). No other descriptor passes all tests: no reliable body association is detected for cadence; the
95
+ apparent ones for leg-normalised step length, knee range, foot clearance and joint speed vanish under the age control;
96
+ and hand distance scales with shoulder width. These descriptors are therefore preserved from the source rather than
97
+ adapted. Within the descriptors and the statistical resolution of this study, a variance-weighted shape predictor
98
+ explains at most 7.5 % of the differences between participants; the rest stays with the participant.
99
+
100
+ <p align="center">
101
+ <img src="assets/lean_cadence.gif" width="100%" alt="One clip retargeted onto a child body, the source body and a wide grid body: forward lean follows hip width while cadence and normalised step stay equal">
102
+ </p>
103
+ <p align="center"><em>An angry forward walk of p09 on a 113 cm child body (p13), on the source body and on a 195 cm
104
+ grid body.
105
+ Forward lean follows hip width (−6.74°, −3.36°, +3.45°; slope 149.7° per metre against the 150° target), while
106
+ cadence (2.00 steps/s) and leg-normalised step length (0.69) are identical on all three bodies.</em></p>
107
+
108
+ ### 2.2 Optimisation
109
+
110
  **Body model.** Each target body is represented by its SMPL-X skeleton and its rest-pose mesh, partitioned by the
111
  skinning weights into fifteen rigid blocks (pelvis, upper torso, head, thighs, shanks, feet, upper arms, forearms and
112
  **hands**), each with a signed-distance field, a mass and a centre of mass. Foot vertices define the sole height,
 
152
  | lean | trunk lean not following the body | (θ_f − θ_f*)² + (θ_s − θ_s*)² on the mean forward / side lean (deg) | 0.3 |
153
 
154
  The penetration allowance a is the largest of 1 cm, the 95th-percentile penetration of the same block pair on the
155
+ source body, and the rest-pose overlap on the target body. The lean targets θ_f* and θ_s* shift the copy's mean
156
+ forward and side trunk lean by k_f·Δw and k_s·Δw for the hip-width difference Δw between target and source
157
+ (Section 2.1).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
158
 
159
  <p align="center">
160
  <img src="assets/terms.gif" width="100%" alt="Weighted residual of each objective term over 400 iterations, with quality checks">
assets/lean_cadence.gif ADDED

Git LFS Details

  • SHA256: 27ff76b16612fd811233914ad7cdb2dcc90636f2f28664dca83a4628941ccd24
  • Pointer size: 132 Bytes
  • Size of remote file: 2.4 MB