-
Notifications
You must be signed in to change notification settings - Fork 19
/
image_fitting.py
260 lines (202 loc) · 11.2 KB
/
image_fitting.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
import argparse
from configs.config import load_config
# General config
#from model_quat import train_manifold2 as train_manifold
from model.posendf import PoseNDF
import shutil
from data.data_splits import amass_splits
import ipdb
import torch
import numpy as np
from body_model import BodyModel
from exp_utils import renderer, PerspectiveCamera
from pytorch3d.transforms import axis_angle_to_quaternion, axis_angle_to_matrix, matrix_to_quaternion
from tqdm import tqdm
from pytorch3d.io import save_obj
import os
import cv2
class ImageFit(object):
def __init__(self, posendf, body_model, out_path='./experiment_results/motion_denoise', debug=False, device='cuda:0', batch_size=1, gender='male', use_joints_conf=True):
self.debug = debug
self.device = device
self.pose_prior = posendf
self.body_model = body_model
self.out_path = out_path
self.use_joints_conf = use_joints_conf
self.dtype=torch.float32
init_joints_idxs = [9, 12, 2, 5]
self.init_joints_idxs = torch.tensor(init_joints_idxs, device=self.device)
self.trans_estimation = 10.0
self.batch_size= batch_size
def get_loss_weights(self):
"""Set loss weights"""
loss_weight = {'temp': lambda cst, it: 10. ** 2 * cst * (1 + it),
'data': lambda cst, it: 10. ** 1 * cst / (1 + it),
'pose_pr': lambda cst, it: 10. ** 2 * cst / (1 + it)
}
return loss_weight
@staticmethod
def backward_step(loss_dict, weight_dict, it):
w_loss = dict()
for k in loss_dict:
w_loss[k] = weight_dict[k](loss_dict[k], it)
tot_loss = list(w_loss.values())
tot_loss = torch.stack(tot_loss).sum()
return tot_loss
@staticmethod
def visualize(vertices, faces, out_path, device, joints=None, render=False, init=False, ):
# save meshes and rendered results if needed
os.makedirs(out_path,exist_ok=True)
os.makedirs( os.path.join(out_path, 'meshes'), exist_ok=True)
if init:
[save_obj(os.path.join(out_path, 'meshes', 'init_{:04}.obj'.format(i) ), vertices[i], faces) for i in range(len(vertices))]
else:
[save_obj(os.path.join(out_path, 'meshes', 'results_{:04}.obj'.format(i) ),vertices[i], faces) for i in range(len(vertices))]
if render:
renderer(vertices, faces, out_path, device=device, init=init)
def camera_loss(self, camera, body_model_output, gt_joints):
projected_joints = camera(body_model_output.joints)
joint_error = torch.pow(
torch.index_select(gt_joints, 1, self.init_joints_idxs) -
torch.index_select(projected_joints, 1, self.init_joints_idxs),
2)
joint_loss = torch.sum(joint_error) * self.loss_weight['data'] ** 2
depth_loss = 0.0
if (self.depth_loss_weight.item() > 0 and self.trans_estimation is not
None):
depth_loss = self.depth_loss_weight ** 2 * torch.sum((
camera.translation[:, 2] - self.trans_estimation[:, 2]).pow(2))
return joint_loss + depth_loss
def projection_loss(self, camera, body_model_output, gt_joints):
projected_joints = camera(body_model_output.joints)
joint_error = torch.pow(gt_joints -projected_joints, 2)
joint_loss = torch.sum(joint_error) * self.loss_weight['data'] ** 2
return joint_loss
def optimize(self, image, keypoints, iterations=10, steps_per_iter=10):
# create camers with given focal length
focal_length = 5000.
camera = PerspectiveCamera(focal_length_x=focal_length, focal_length_y=focal_length)
camera = camera.to(device=self.device)
# create initial SMPL from mean pose and shape
betas = torch.zeros((self.batch_size,10)).to(device=self.device)
pose_body = torch.zeros((self.batch_size,69)).to(device=self.device)
smpl_init = self.body_model(betas=betas, pose_body=pose_body)
keypoint_data = torch.tensor(keypoints, dtype=self.dtype)
gt_joints = keypoint_data[:, :, :2].to(device=self.device, dtype=self.dtype)
if self.use_joints_conf:
joints_conf = keypoint_data[:, :, 2].reshape(1, -1).to(device=self.device, dtype=self.dtype)
# Step 1: The indices of the joints used for the initialization of the camera
camera.translation.requires_grad = True
smpl_init.global_orient = True
smpl_init.pose_body = False
smpl_init.betas = False
camera_opt_params = [camera.translation, smpl_init.global_orient]
optimizer = torch.optim.Adam(camera_opt_params, 0.02, betas=(0.9, 0.999))
for it in range(iterations):
loop = tqdm(range(steps_per_iter))
loop.set_description('Optimizing camera translation for torso joints')
for i in loop:
optimizer.zero_grad()
loss_dict = dict()
smpl_init = self.body_model(betas=self.betas, pose_body=smpl_init.body_pose, global_orient=smpl_init.global_orient)
# Get total loss for backward pass
tot_loss = self.camera_loss(camera, smpl_init, gt_joints)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.8f}'.format(k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
# Step 2: Optimize for all joints
smpl_init.global_orient = True
smpl_init.pose_body = True
smpl_init.betas = True
body_opt_params = [smpl_init.pose_body, smpl_init.global_orient, smpl_init.betas ]
optimizer = torch.optim.Adam(body_opt_params, 0.02, betas=(0.9, 0.999))
for it in range(iterations):
loop = tqdm(range(steps_per_iter))
loop.set_description('Optimizing full model')
for i in loop:
optimizer.zero_grad()
loss_dict = dict()
smpl_init = self.body_model(betas=self.betas, pose_body=smpl_init.body_pose, global_orient=smpl_init.global_orient)
# convert pose to quaternion and predict distance
pose_quat = axis_angle_to_quaternion(smpl_init.body_pose.view(-1, 23, 3)[:, :21])
loss_dict['pose_pr']= torch.mean(self.pose_prior(pose_quat, train=False)['dist_pred'])
# Get total loss for backward pass
loss_dict['data'] = self.projection_loss(camera, smpl_init, gt_joints)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.8f}'.format(k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
smpl_init = self.body_model(betas=betas, pose_body=pose_body)
self.visualize(smpl_init.vertices, smpl_init.faces, self.out_path, device=self.device, joints=smpl_init.Jtr, render=True, init=True)
init_joints = torch.from_numpy(smpl_init.Jtr.detach().cpu().numpy().astype(np.float32)).to(device=self.device)
# Optimizer
smpl_init.body_pose.requires_grad= True
smpl_init.betas.requires_grad = True
optimizer = torch.optim.Adam([smpl_init.body_pose], 0.02, betas=(0.9, 0.999))
# Get loss_weights
weight_dict = self.get_loss_weights()
for it in range(iterations):
loop = tqdm(range(steps_per_iter))
loop.set_description('Optimizing SMPL poses')
for i in loop:
optimizer.zero_grad()
loss_dict = dict()
# convert pose to quaternion and predict distance
pose_quat = axis_angle_to_quaternion(smpl_init.body_pose.view(-1, 23, 3)[:, :21])
loss_dict['pose_pr']= torch.mean(self.pose_prior(pose_quat, train=False)['dist_pred'])
# calculate temporal loss between mesh vertices
smpl_init = self.body_model(betas=self.betas, pose_body=smpl_init.body_pose)
# smpl_opt = self.smpl(betas=self.betas, pose_body=smpl_init.body_pose)
temp_term = smpl_init.vertices[:-1] - smpl_init.vertices[1:]
loss_dict['temp'] = torch.mean(torch.sqrt(torch.sum(temp_term*temp_term, dim=2)))
# calculate data term from inital noisy pose
if it > 0: #for nans
data_term = smpl_init.Jtr - init_joints
loss_dict['data'] = torch.mean(torch.sqrt(torch.sum(data_term*data_term, dim=2)))
# Get total loss for backward pass
tot_loss = self.backward_step(loss_dict, weight_dict, it)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.8f}'.format(k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
# ipdb.set_trace()
# create final results
smpl_init = self.body_model(betas=self.betas, pose_body=smpl_init.body_pose)
self.visualize(smpl_init.vertices, smpl_init.faces, self.out_path, device=self.device, joints=smpl_init.Jtr, render=True)
print('** Optimised pose **')
def main(opt, ckpt, image_folder, out_path):
batch_size = 120
### load the model
net = PoseNDF(opt)
device= 'cuda:0'
ckpt = torch.load(ckpt, map_location='cpu')['model_state_dict']
net.load_state_dict(ckpt)
net.eval()
net = net.to(device)
# load body model
bm_dir_path = '/BS/garvita/work/SMPL_data/models/smpl'
body_model = BodyModel(bm_path=bm_dir_path, model_type='smpl', batch_size=batch_size, num_betas=10).to(device=device)
# load image and keypoints
image = cv2.imread(os.path.join(image_folder, 'img.jpg'))
keypoints = torch.from_numpy(np.load(os.path.join(image_folder, 'kpts.npz'))['0'].astype(np.float32)).to(device=device)
# create Motion denoiser layer
imagefit = ImageFit(net, body_model=body_model, batch_size=1, out_path=out_path)
imagefit.optimize(image, keypoints)
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Interpolate using PoseNDF.'
)
parser.add_argument('--config', '-c', default='configs/amass.yaml', type=str, help='Path to config file.')
parser.add_argument('--ckpt_path', '-ckpt', default='/BS/humanpose/static00/pose_manifold/amass/test_lrelu_l1_1e-05_dist10.0_eik1.0/checkpoints/checkpoint_epoch_best.tar', type=str, help='Path to pretrained model.')
parser.add_argument('--image_folder', '-if', default='/BS/humanpose/static00/data/PoseNDF_exp/motion_denoise_data/SSM_synced/20161014_50033/punch_kick_sync_poses.npz', type=str, help='Path to image and its keypoint')
parser.add_argument('--outpath_folder', '-out', default='/BS/humanpose/static00/data/PoseNDF_exp/motion_denoise_results/SSM_synced/20161014_50033/punch_kick_sync_poses', type=str, help='Path to output')
args = parser.parse_args()
opt = load_config(args.config)
main(opt, args.ckpt_path, args.image_folder, args.outpath_folder)