Spaces:
Runtime error
Runtime error
kleinhe
commited on
Commit
·
73ba0a5
1
Parent(s):
a995483
speed
Browse files
SMPLX/visualize_joint2smpl/joints2smpl/src/smplify.py
CHANGED
@@ -8,7 +8,7 @@ from customloss import (camera_fitting_loss_3d,
|
|
8 |
)
|
9 |
from prior import MaxMixturePrior
|
10 |
from SMPLX.visualize_joint2smpl.joints2smpl.src import config
|
11 |
-
|
12 |
|
13 |
@torch.no_grad()
|
14 |
def guess_init_3d(model_joints,
|
@@ -41,32 +41,21 @@ class SMPLify3D():
|
|
41 |
|
42 |
def __init__(self,
|
43 |
smplxmodel,
|
44 |
-
step_size=
|
45 |
-
batch_size=1,
|
46 |
num_iters=100,
|
47 |
-
use_collision=False,
|
48 |
-
use_lbfgs=True,
|
49 |
joints_category="orig",
|
50 |
device=torch.device('cuda:0'),
|
51 |
):
|
52 |
|
53 |
# Store options
|
54 |
-
self.batch_size = batch_size
|
55 |
self.device = device
|
56 |
self.step_size = step_size
|
57 |
|
58 |
self.num_iters = num_iters
|
59 |
-
# --- choose optimizer
|
60 |
-
self.use_lbfgs = use_lbfgs
|
61 |
# GMM pose prior
|
62 |
self.pose_prior = MaxMixturePrior(prior_folder=config.GMM_MODEL_DIR,
|
63 |
num_gaussians=8,
|
64 |
dtype=torch.float32).to(device)
|
65 |
-
# collision part
|
66 |
-
self.use_collision = use_collision
|
67 |
-
if self.use_collision:
|
68 |
-
self.part_segm_fn = config.Part_Seg_DIR
|
69 |
-
|
70 |
# reLoad SMPL-X model
|
71 |
self.smpl = smplxmodel
|
72 |
|
@@ -103,35 +92,6 @@ class SMPLify3D():
|
|
103 |
betas: SMPL beta parameters of optimized shape
|
104 |
camera_translation: Camera translation
|
105 |
"""
|
106 |
-
|
107 |
-
# # # add the mesh inter-section to avoid
|
108 |
-
search_tree = None
|
109 |
-
pen_distance = None
|
110 |
-
filter_faces = None
|
111 |
-
|
112 |
-
if self.use_collision:
|
113 |
-
from mesh_intersection.bvh_search_tree import BVH
|
114 |
-
import mesh_intersection.loss as collisions_loss
|
115 |
-
from mesh_intersection.filter_faces import FilterFaces
|
116 |
-
|
117 |
-
search_tree = BVH(max_collisions=8)
|
118 |
-
|
119 |
-
pen_distance = collisions_loss.DistanceFieldPenetrationLoss(
|
120 |
-
sigma=0.5, point2plane=False, vectorized=True, penalize_outside=True)
|
121 |
-
|
122 |
-
if self.part_segm_fn:
|
123 |
-
# Read the part segmentation
|
124 |
-
part_segm_fn = os.path.expandvars(self.part_segm_fn)
|
125 |
-
with open(part_segm_fn, 'rb') as faces_parents_file:
|
126 |
-
face_segm_data = pickle.load(faces_parents_file, encoding='latin1')
|
127 |
-
faces_segm = face_segm_data['segm']
|
128 |
-
faces_parents = face_segm_data['parents']
|
129 |
-
# Create the module used to filter invalid collision pairs
|
130 |
-
filter_faces = FilterFaces(
|
131 |
-
faces_segm=faces_segm, faces_parents=faces_parents,
|
132 |
-
ign_part_pairs=None).to(device=self.device)
|
133 |
-
|
134 |
-
|
135 |
# Split SMPL pose to body pose and global orientation
|
136 |
body_pose = init_pose[:, 3:].detach().clone()
|
137 |
global_orient = init_pose[:, :3].detach().clone()
|
@@ -150,42 +110,29 @@ class SMPLify3D():
|
|
150 |
# -------------Step 1: Optimize camera translation and body orientation--------
|
151 |
# Optimize only camera translation and body orientation
|
152 |
body_pose.requires_grad = False
|
153 |
-
betas.requires_grad =
|
154 |
global_orient.requires_grad = True
|
155 |
camera_translation.requires_grad = True
|
156 |
|
157 |
-
camera_opt_params = [global_orient, camera_translation]
|
158 |
-
|
159 |
-
if self.use_lbfgs:
|
160 |
-
camera_optimizer = torch.optim.LBFGS(camera_opt_params, max_iter=self.num_iters,
|
161 |
-
lr=self.step_size, line_search_fn='strong_wolfe')
|
162 |
-
for i in range(10):
|
163 |
-
def closure():
|
164 |
-
camera_optimizer.zero_grad()
|
165 |
-
smpl_output = self.smpl(global_orient=global_orient,
|
166 |
-
body_pose=body_pose,
|
167 |
-
betas=betas)
|
168 |
-
model_joints = smpl_output.joints
|
169 |
-
loss = camera_fitting_loss_3d(model_joints, camera_translation,
|
170 |
-
init_cam_t, j3d, self.joints_category)
|
171 |
-
loss.backward()
|
172 |
-
return loss
|
173 |
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
|
|
179 |
smpl_output = self.smpl(global_orient=global_orient,
|
180 |
body_pose=body_pose,
|
181 |
betas=betas)
|
182 |
model_joints = smpl_output.joints
|
|
|
|
|
183 |
|
184 |
-
loss = camera_fitting_loss_3d(model_joints[:, self.smpl_index], camera_translation,
|
185 |
-
init_cam_t, j3d[:, self.corr_index], self.joints_category)
|
186 |
-
camera_optimizer.zero_grad()
|
187 |
loss.backward()
|
188 |
-
|
|
|
|
|
189 |
|
190 |
# Fix camera translation after optimizing camera
|
191 |
# --------Step 2: Optimize body joints --------------------------
|
@@ -193,43 +140,15 @@ class SMPLify3D():
|
|
193 |
body_pose.requires_grad = True
|
194 |
global_orient.requires_grad = True
|
195 |
camera_translation.requires_grad = True
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
if self.use_lbfgs:
|
206 |
-
body_optimizer = torch.optim.LBFGS(body_opt_params, max_iter=self.num_iters,
|
207 |
-
lr=self.step_size, line_search_fn='strong_wolfe')
|
208 |
-
for i in range(self.num_iters):
|
209 |
-
def closure():
|
210 |
-
body_optimizer.zero_grad()
|
211 |
-
smpl_output = self.smpl(global_orient=global_orient,
|
212 |
-
body_pose=body_pose,
|
213 |
-
betas=betas)
|
214 |
-
model_joints = smpl_output.joints
|
215 |
-
model_vertices = smpl_output.vertices
|
216 |
-
|
217 |
-
loss = body_fitting_loss_3d(body_pose, preserve_pose, betas, model_joints[:, self.smpl_index], camera_translation,
|
218 |
-
j3d[:, self.corr_index], self.pose_prior,
|
219 |
-
joints3d_conf=conf_3d,
|
220 |
-
joint_loss_weight=600.0,
|
221 |
-
pose_preserve_weight=5.0,
|
222 |
-
use_collision=self.use_collision,
|
223 |
-
model_vertices=model_vertices, model_faces=self.model_faces,
|
224 |
-
search_tree=search_tree, pen_distance=pen_distance, filter_faces=filter_faces)
|
225 |
-
loss.backward()
|
226 |
-
return loss
|
227 |
-
|
228 |
-
body_optimizer.step(closure)
|
229 |
-
else:
|
230 |
-
body_optimizer = torch.optim.Adam(body_opt_params, lr=self.step_size, betas=(0.9, 0.999))
|
231 |
-
|
232 |
-
for i in range(self.num_iters):
|
233 |
smpl_output = self.smpl(global_orient=global_orient,
|
234 |
body_pose=body_pose,
|
235 |
betas=betas)
|
@@ -240,31 +159,15 @@ class SMPLify3D():
|
|
240 |
j3d[:, self.corr_index], self.pose_prior,
|
241 |
joints3d_conf=conf_3d,
|
242 |
joint_loss_weight=600.0,
|
243 |
-
|
|
|
244 |
model_vertices=model_vertices, model_faces=self.model_faces,
|
245 |
-
search_tree=
|
246 |
-
body_optimizer.zero_grad()
|
247 |
loss.backward()
|
248 |
-
|
249 |
|
250 |
-
|
251 |
-
with torch.no_grad():
|
252 |
-
smpl_output = self.smpl(global_orient=global_orient,
|
253 |
-
body_pose=body_pose,
|
254 |
-
betas=betas, return_full_pose=True)
|
255 |
-
model_joints = smpl_output.joints
|
256 |
-
model_vertices = smpl_output.vertices
|
257 |
|
258 |
-
final_loss = body_fitting_loss_3d(body_pose, preserve_pose, betas, model_joints[:, self.smpl_index], camera_translation,
|
259 |
-
j3d[:, self.corr_index], self.pose_prior,
|
260 |
-
joints3d_conf=conf_3d,
|
261 |
-
joint_loss_weight=600.0,
|
262 |
-
use_collision=self.use_collision, model_vertices=model_vertices, model_faces=self.model_faces,
|
263 |
-
search_tree=search_tree, pen_distance=pen_distance, filter_faces=filter_faces)
|
264 |
|
265 |
-
vertices = smpl_output.vertices.detach()
|
266 |
-
joints = smpl_output.joints.detach()
|
267 |
pose = torch.cat([global_orient, body_pose], dim=-1).detach()
|
268 |
-
|
269 |
-
|
270 |
-
return vertices, joints, pose, betas, camera_translation, final_loss
|
|
|
8 |
)
|
9 |
from prior import MaxMixturePrior
|
10 |
from SMPLX.visualize_joint2smpl.joints2smpl.src import config
|
11 |
+
from tqdm import tqdm
|
12 |
|
13 |
@torch.no_grad()
|
14 |
def guess_init_3d(model_joints,
|
|
|
41 |
|
42 |
def __init__(self,
|
43 |
smplxmodel,
|
44 |
+
step_size=1.0,
|
|
|
45 |
num_iters=100,
|
|
|
|
|
46 |
joints_category="orig",
|
47 |
device=torch.device('cuda:0'),
|
48 |
):
|
49 |
|
50 |
# Store options
|
|
|
51 |
self.device = device
|
52 |
self.step_size = step_size
|
53 |
|
54 |
self.num_iters = num_iters
|
|
|
|
|
55 |
# GMM pose prior
|
56 |
self.pose_prior = MaxMixturePrior(prior_folder=config.GMM_MODEL_DIR,
|
57 |
num_gaussians=8,
|
58 |
dtype=torch.float32).to(device)
|
|
|
|
|
|
|
|
|
|
|
59 |
# reLoad SMPL-X model
|
60 |
self.smpl = smplxmodel
|
61 |
|
|
|
92 |
betas: SMPL beta parameters of optimized shape
|
93 |
camera_translation: Camera translation
|
94 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
# Split SMPL pose to body pose and global orientation
|
96 |
body_pose = init_pose[:, 3:].detach().clone()
|
97 |
global_orient = init_pose[:, :3].detach().clone()
|
|
|
110 |
# -------------Step 1: Optimize camera translation and body orientation--------
|
111 |
# Optimize only camera translation and body orientation
|
112 |
body_pose.requires_grad = False
|
113 |
+
betas.requires_grad = True
|
114 |
global_orient.requires_grad = True
|
115 |
camera_translation.requires_grad = True
|
116 |
|
117 |
+
camera_opt_params = [betas, global_orient, camera_translation]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
|
119 |
+
camera_optimizer = torch.optim.LBFGS(camera_opt_params, max_iter=10,
|
120 |
+
lr=self.step_size, line_search_fn='strong_wolfe')
|
121 |
+
cycle = tqdm(range(10))
|
122 |
+
for i in cycle:
|
123 |
+
def closure():
|
124 |
+
camera_optimizer.zero_grad()
|
125 |
smpl_output = self.smpl(global_orient=global_orient,
|
126 |
body_pose=body_pose,
|
127 |
betas=betas)
|
128 |
model_joints = smpl_output.joints
|
129 |
+
loss = camera_fitting_loss_3d(model_joints, camera_translation,
|
130 |
+
init_cam_t, j3d, self.joints_category)
|
131 |
|
|
|
|
|
|
|
132 |
loss.backward()
|
133 |
+
return loss
|
134 |
+
|
135 |
+
camera_optimizer.step(closure)
|
136 |
|
137 |
# Fix camera translation after optimizing camera
|
138 |
# --------Step 2: Optimize body joints --------------------------
|
|
|
140 |
body_pose.requires_grad = True
|
141 |
global_orient.requires_grad = True
|
142 |
camera_translation.requires_grad = True
|
143 |
+
betas.requires_grad = True
|
144 |
+
body_opt_params = [body_pose, betas, global_orient, camera_translation]
|
145 |
+
|
146 |
+
cycle = tqdm(range(self.num_iters))
|
147 |
+
body_optimizer = torch.optim.LBFGS(body_opt_params, max_iter=self.num_iters,
|
148 |
+
lr=self.step_size, line_search_fn='strong_wolfe')
|
149 |
+
for i in cycle:
|
150 |
+
def closure():
|
151 |
+
body_optimizer.zero_grad()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
smpl_output = self.smpl(global_orient=global_orient,
|
153 |
body_pose=body_pose,
|
154 |
betas=betas)
|
|
|
159 |
j3d[:, self.corr_index], self.pose_prior,
|
160 |
joints3d_conf=conf_3d,
|
161 |
joint_loss_weight=600.0,
|
162 |
+
pose_preserve_weight=5.0,
|
163 |
+
use_collision=False,
|
164 |
model_vertices=model_vertices, model_faces=self.model_faces,
|
165 |
+
search_tree=None, pen_distance=None, filter_faces=None)
|
|
|
166 |
loss.backward()
|
167 |
+
return loss
|
168 |
|
169 |
+
body_optimizer.step(closure)
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
|
|
|
|
|
172 |
pose = torch.cat([global_orient, body_pose], dim=-1).detach()
|
173 |
+
return pose
|
|
|
|
SMPLX/visualize_joint2smpl/simplify_loc2rot.py
CHANGED
@@ -7,7 +7,6 @@ from SMPLX.visualize_joint2smpl.joints2smpl.src.smplify import SMPLify3D
|
|
7 |
from tqdm import tqdm
|
8 |
import argparse
|
9 |
|
10 |
-
|
11 |
class joints2smpl:
|
12 |
|
13 |
def __init__(self, num_frames, device, model_path=None, json_dict=None):
|
@@ -17,8 +16,9 @@ class joints2smpl:
|
|
17 |
self.batch_size = num_frames
|
18 |
self.num_joints = 22 # for HumanML3D
|
19 |
self.joint_category = "AMASS"
|
20 |
-
self.num_smplify_iters =
|
21 |
self.fix_foot = False
|
|
|
22 |
smplmodel = smplx.create(self.smpl_dir, model_type="smpl", gender="neutral", ext="pkl",
|
23 |
batch_size=self.batch_size).to(self.device)
|
24 |
|
@@ -33,7 +33,6 @@ class joints2smpl:
|
|
33 |
|
34 |
# # #-------------initialize SMPLify
|
35 |
self.smplify = SMPLify3D(smplxmodel=smplmodel,
|
36 |
-
batch_size=self.batch_size,
|
37 |
joints_category=self.joint_category,
|
38 |
num_iters=self.num_smplify_iters,
|
39 |
device=self.device)
|
@@ -92,18 +91,17 @@ class joints2smpl:
|
|
92 |
else:
|
93 |
print("Such category not settle down!")
|
94 |
|
95 |
-
|
96 |
-
new_opt_cam_t, new_opt_joint_loss = self.smplify(
|
97 |
pred_pose.detach(),
|
98 |
pred_betas.detach(),
|
99 |
pred_cam_t.detach(),
|
100 |
keypoints_3d,
|
101 |
conf_3d=confidence_input.to(self.device),
|
102 |
-
# seq_ind=idx
|
103 |
)
|
104 |
|
105 |
thetas = new_opt_pose.reshape(self.batch_size, 24 * 3)
|
106 |
vecs = thetas.detach().cpu().numpy()
|
|
|
107 |
return vecs, root_loc
|
108 |
|
109 |
|
|
|
7 |
from tqdm import tqdm
|
8 |
import argparse
|
9 |
|
|
|
10 |
class joints2smpl:
|
11 |
|
12 |
def __init__(self, num_frames, device, model_path=None, json_dict=None):
|
|
|
16 |
self.batch_size = num_frames
|
17 |
self.num_joints = 22 # for HumanML3D
|
18 |
self.joint_category = "AMASS"
|
19 |
+
self.num_smplify_iters = 15
|
20 |
self.fix_foot = False
|
21 |
+
|
22 |
smplmodel = smplx.create(self.smpl_dir, model_type="smpl", gender="neutral", ext="pkl",
|
23 |
batch_size=self.batch_size).to(self.device)
|
24 |
|
|
|
33 |
|
34 |
# # #-------------initialize SMPLify
|
35 |
self.smplify = SMPLify3D(smplxmodel=smplmodel,
|
|
|
36 |
joints_category=self.joint_category,
|
37 |
num_iters=self.num_smplify_iters,
|
38 |
device=self.device)
|
|
|
91 |
else:
|
92 |
print("Such category not settle down!")
|
93 |
|
94 |
+
new_opt_pose = self.smplify(
|
|
|
95 |
pred_pose.detach(),
|
96 |
pred_betas.detach(),
|
97 |
pred_cam_t.detach(),
|
98 |
keypoints_3d,
|
99 |
conf_3d=confidence_input.to(self.device),
|
|
|
100 |
)
|
101 |
|
102 |
thetas = new_opt_pose.reshape(self.batch_size, 24 * 3)
|
103 |
vecs = thetas.detach().cpu().numpy()
|
104 |
+
|
105 |
return vecs, root_loc
|
106 |
|
107 |
|
app.py
CHANGED
@@ -121,7 +121,7 @@ def t2m_demo():
|
|
121 |
|
122 |
with gr.Row():
|
123 |
condition = gr.Radio(['text', 'uncond'], value='text', label='Condition', info="If sythesize motion with prompt?")
|
124 |
-
out_size = gr.Number(value=
|
125 |
|
126 |
with gr.Row():
|
127 |
render_mode = gr.Radio(['joints','pyrender_fast', 'pyrender_slow'], value='joints', label='Render', info="If render results to 3D meshes? Pyrender need more time.")
|
|
|
121 |
|
122 |
with gr.Row():
|
123 |
condition = gr.Radio(['text', 'uncond'], value='text', label='Condition', info="If sythesize motion with prompt?")
|
124 |
+
out_size = gr.Number(value=256, label="Resolution", info="The resolution of output videos", minimum=128, maximum=2048, precision=0)
|
125 |
|
126 |
with gr.Row():
|
127 |
render_mode = gr.Radio(['joints','pyrender_fast', 'pyrender_slow'], value='joints', label='Render', info="If render results to 3D meshes? Pyrender need more time.")
|