-
Notifications
You must be signed in to change notification settings - Fork 3
/
preprocess_scene_s2_for_train.py
209 lines (175 loc) · 12.4 KB
/
preprocess_scene_s2_for_train.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
"""
Preprocess script for EgoBody for second stage of EgoHMR (i.e., the scene-conditioned diffusion model for local body pose).
Scene vertices around the ground truth body in a 2x2m cube will be cropped and saved.
The preprocessed scene vertices are used for training of the EgoHMR diffusion model.
"""
import open3d as o3d
import numpy as np
import os
import pickle as pkl
from tqdm import tqdm
import pandas as pd
import smplx
import torch
import copy
import random
import math
import argparse
from utils.other_utils import *
from utils.geometry import *
parser = argparse.ArgumentParser(description='ProHMR training code')
parser.add_argument('--scene_verts_num_target', type=int, default=20000, help='numbers of scene vertices to crop')
parser.add_argument('--cube_size', type=int, default=2, help='cropped scene cube size')
parser.add_argument('--split', type=str, default='train', help='val/train/test')
parser.add_argument('--data_root', type=str, default='/mnt/ssd/egobody_release/', help='path to egobody data')
parser.add_argument('--save_root', type=str, default='/mnt/ssd/egobody_release/Egohmr_scene_preprocess_cube_s2_from_gt',
help='path to save preprocessed scene point cloud')
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if __name__ == '__main__':
################################ read dataset information ################################
df = pd.read_csv(os.path.join(args.data_root, 'data_info_release.csv'))
recording_name_list = list(df['recording_name'])
scene_name_list = list(df['scene_name'])
body_idx_fpv_list = list(df['body_idx_fpv'])
body_idx_fpv_dict = dict(zip(recording_name_list, body_idx_fpv_list))
scene_name_dict = dict(zip(recording_name_list, scene_name_list))
data = np.load(os.path.join(args.data_root, 'smpl_spin_npz/egocapture_{}_smpl.npz'.format(args.split)))
with open(os.path.join(args.data_root, 'transf_matrices_all_seqs.pkl'), 'rb') as fp:
transf_matrices = pkl.load(fp)
imgname_list = data['imgname'] # 'egocentric_color/...'
betas_list = data['shape']
global_orient_pv_list = data['global_orient_pv']
transl_pv_list = data['transl_pv']
body_pose_list = data['pose']
gender_list = data['gender']
fx_list = data['fx']
fy_list = data['fy']
cam_cx_list = data['cx']
cam_cy_list = data['cy']
[imgname_list, seqname_list, _] = zip(*[get_right_full_img_pth(x, args.data_root) for x in imgname_list]) # absolute dir
seqname_list = [os.path.basename(x) for x in seqname_list]
body_model_male = smplx.create('data/smpl', model_type='smpl', gender='male').to(device)
body_model_female = smplx.create('data/smpl', model_type='smpl', gender='female').to(device)
# additional transformation from opengl coord to opencv system: ego camera - opengl coord, kinect camera - opencv coord
add_trans = np.array([[1.0, 0, 0, 0],
[0, -1, 0, 0],
[0, 0, -1, 0],
[0, 0, 0, 1]])
orig_scene_mesh_dict = {}
################################## scene pcd processing ##################################
step = 1
cnt = 0
last_scene_name = ''
for i in tqdm(range(0, len(imgname_list), step)):
imgname = imgname_list[i]
recording_name = imgname.split('/')[-4]
holo_recording_time = imgname.split('/')[-3]
frame_name = imgname.split('/')[-1][0:-4]
frame_id = imgname.split('/')[-1][-15:-4]
scene_name = scene_name_dict[recording_name]
if cnt % 1 == 0 or last_scene_name != scene_name:
trans_kinect2holo, trans_holo2pv = get_transf_matrices_per_frame(transf_matrices, imgname_list[i], seqname_list[i])
######## read scene
if scene_name not in orig_scene_mesh_dict.keys():
scene_dir = os.path.join(os.path.join(args.data_root, 'scene_mesh'), '{}/{}.obj'.format(scene_name, scene_name))
orig_scene_mesh_dict[scene_name] = o3d.io.read_triangle_mesh(scene_dir, print_progress=True)
orig_scene_mesh_dict[scene_name].compute_vertex_normals()
cur_scene_mesh = copy.deepcopy(orig_scene_mesh_dict[scene_name])
scene_verts = np.asarray(cur_scene_mesh.vertices) # [n, 3]
########## read calibration files
calib_trans_dir = os.path.join(args.data_root, 'calibrations', recording_name)
cam2world_dir = os.path.join(calib_trans_dir,
'cal_trans/kinect12_to_world') # transformation from master kinect RGB camera to scene mesh
with open(os.path.join(cam2world_dir, scene_name + '.json'), 'r') as f:
trans_scene_to_main = np.array(json.load(f)['trans'])
trans_scene_to_main = np.linalg.inv(trans_scene_to_main)
###### get gt body in current pv coord
torch_param = {}
torch_param['transl'] = torch.tensor(transl_pv_list[[i]]).float().to(device)
torch_param['global_orient'] = torch.tensor(global_orient_pv_list[[i]]).float().to(device)
torch_param['betas'] = torch.tensor(betas_list[[i]]).float().to(device)
torch_param['body_pose'] = torch.tensor(body_pose_list[[i]]).float().to(device)
if gender_list[i] == 'm':
output = body_model_male(return_verts=True, **torch_param)
elif gender_list[i] == 'f':
output = body_model_female(return_verts=True, **torch_param)
body_verts = output.vertices.detach().cpu().numpy().squeeze() # [6890, 3]
######################## coord transform ########################
# for gt body: pv coord --> master kinect coord --> scene mesh space
gt_body_o3d_pv = o3d.geometry.TriangleMesh()
gt_body_o3d_pv.vertices = o3d.utility.Vector3dVector(body_verts) # [6890, 3]
gt_body_o3d_pv.triangles = o3d.utility.Vector3iVector(body_model_male.faces)
gt_body_o3d_pv.compute_vertex_normals()
gt_body_o3d_scene = copy.deepcopy(gt_body_o3d_pv)
gt_body_o3d_scene.transform(np.linalg.inv(add_trans))
gt_body_o3d_scene.transform(np.linalg.inv(trans_holo2pv))
gt_body_o3d_scene.transform(np.linalg.inv(trans_kinect2holo))
gt_body_o3d_scene.transform(np.linalg.inv(trans_scene_to_main))
body_verts_scene = np.asarray(gt_body_o3d_scene.vertices) # [6890, 3]
############# get body center
body_center = np.mean(body_verts_scene, axis=0) # [x, y, z]
####################### augment / crop scene mesh ########################
# random rotation / translation of a scene cube around the body
# note! in scene coord, y axis up
# rotate scene (pv coord) around body center
rot_angle = random.uniform(0, 2 * (math.pi))
scene_verts_aug = np.zeros(scene_verts.shape)
scene_verts_aug[:, 0] = (scene_verts[:, 0] - body_center[0]) * math.cos(rot_angle) - (scene_verts[:, 2] - body_center[2]) * math.sin(rot_angle) + body_center[0]
scene_verts_aug[:, 2] = (scene_verts[:, 0] - body_center[0]) * math.sin(rot_angle) + (scene_verts[:, 2] - body_center[2]) * math.cos(rot_angle) + body_center[2]
scene_verts_aug[:, 1] = scene_verts[:, 1]
# random shift the cude a bit as data augmentation, such that training does not overfit
random_shift = np.array([0.0, 0.0, 0.0])
body_verts_scene_auge = np.zeros(body_verts_scene.shape)
body_verts_scene_auge[:, 0] = (body_verts_scene[:, 0] - body_center[0]) * math.cos(rot_angle) - (body_verts_scene[:, 2] - body_center[2]) * math.sin(rot_angle) + body_center[0]
body_verts_scene_auge[:, 2] = (body_verts_scene[:, 0] - body_center[0]) * math.sin(rot_angle) + (body_verts_scene[:, 2] - body_center[2]) * math.cos(rot_angle) + body_center[2]
body_verts_scene_auge[:, 1] = body_verts_scene[:, 1]
body_auge_min_x = np.min(body_verts_scene_auge[:, 0])
body_auge_max_x = np.max(body_verts_scene_auge[:, 0])
body_auge_min_y = np.min(body_verts_scene_auge[:, 2])
body_auge_max_y = np.max(body_verts_scene_auge[:, 2])
random_shift[0] = random.uniform(max(-args.cube_size / 4, (body_auge_max_x - body_center[0]) - args.cube_size / 2),
min(args.cube_size / 4, args.cube_size / 2 - (body_center[0] - body_auge_min_x)))
random_shift[2] = random.uniform(max(-args.cube_size / 4, (body_auge_max_y - body_center[2]) - args.cube_size / 2),
min(args.cube_size / 4, args.cube_size / 2 - (body_center[2] - body_auge_min_y)))
min_x = body_center[0] - args.cube_size / 2 + random_shift[0]
max_x = body_center[0] + args.cube_size / 2 + random_shift[0]
min_y = body_center[2] - args.cube_size / 2 + random_shift[2]
max_y = body_center[2] + args.cube_size / 2 + random_shift[2]
# cropped scene verts in the cube
scene_verts_auge_crop = scene_verts_aug[np.where((scene_verts_aug[:, 0] >= min_x) & (scene_verts_aug[:, 0] <= max_x) &
(scene_verts_aug[:, 2] >= min_y) & (scene_verts_aug[:, 2] <= max_y))]
scene_verts_auge_crop_center = body_center + random_shift # define cube center
scene_verts_auge_crop_center[1] = np.min(scene_verts_auge_crop[:, 1]) + 1.0 # fix cube center 1m above ground in y axis
scene_verts_auge_crop = scene_verts_auge_crop
scene_verts_auge_crop = scene_verts_auge_crop[scene_verts_auge_crop[:, 1] <= np.min(scene_verts_auge_crop[:, 1]) + args.cube_size]
####################### dowmsample scene verts ########################
scene_pcd_auge_crop = o3d.geometry.PointCloud()
scene_pcd_auge_crop.points = o3d.utility.Vector3dVector(scene_verts_auge_crop)
n_verts = len(scene_verts_auge_crop)
if n_verts < args.scene_verts_num_target:
print('[ERROR]', 'scene vertex number', n_verts, '<', 'scene_verts_num_target', args.scene_verts_num_target)
exit()
downsample_rate = int(n_verts / args.scene_verts_num_target)
scene_pcd_auge_crop = scene_pcd_auge_crop.uniform_down_sample(every_k_points=downsample_rate)
scene_verts_auge_crop_downsample = np.asarray(scene_pcd_auge_crop.points)[0:args.scene_verts_num_target]
####################### reansform back to original scene space ##########################
scene_verts_crop_downsample = np.zeros(scene_verts_auge_crop_downsample.shape)
scene_verts_crop_downsample[:, 0] = (scene_verts_auge_crop_downsample[:, 0] - body_center[0]) * math.cos(-rot_angle) - (scene_verts_auge_crop_downsample[:, 2] - body_center[2]) * math.sin(-rot_angle) + body_center[0]
scene_verts_crop_downsample[:, 2] = (scene_verts_auge_crop_downsample[:, 0] - body_center[0]) * math.sin(-rot_angle) + (scene_verts_auge_crop_downsample[:, 2] - body_center[2]) * math.cos(-rot_angle) + body_center[2]
scene_verts_crop_downsample[:, 1] = scene_verts_auge_crop_downsample[:, 1]
# ####### visualize
# mesh_frame = o3d.geometry.TriangleMesh.create_coordinate_frame(size=1.0, origin=[0, 0, 0])
# scene_pcd_crop_downsample = o3d.geometry.PointCloud()
# scene_pcd_crop_downsample.points = o3d.utility.Vector3dVector(scene_verts_crop_downsample)
# # o3d.visualization.draw_geometries([scene_pcd_crop_downsample, mesh_frame])
# o3d.visualization.draw_geometries([scene_pcd_crop_downsample, mesh_frame, gt_body_o3d_scene])
####################### save pcds in scene coord system ##########################
# /mnt/ssd/proHMR_scene_preprocess/recording_name/2021-09-07-155421/timestamp_frame_xxxxx.obj
if not os.path.exists(os.path.join(args.save_root, args.split, recording_name, holo_recording_time)):
os.makedirs(os.path.join(args.save_root, args.split, recording_name, holo_recording_time))
save_path = os.path.join(args.save_root, args.split, recording_name, holo_recording_time, frame_name+'.npy')
np.save(save_path, scene_verts_crop_downsample)
cnt += 1
last_scene_name = scene_name
print('Completed scene point clouds preprocessing, results saved to {}.'.format(args.save_root))