/
pose_estimation.py
324 lines (249 loc) · 14.5 KB
/
pose_estimation.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
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
import os
from gaussian_splatting.utils.loss_utils import l1_loss, ssim
from gaussian_splatting.gaussian_renderer import render
from gaussian_splatting.scene import DiffGaussianModel
from argparse import ArgumentParser
from gaussian_splatting.arguments import ModelParams, PipelineParams
from gaussian_splatting.render import *
from gaussian_splatting.scene.cameras import Camera
import torch
import numpy as np
import matplotlib.pyplot as plt
from utils_pose_est import DefectDataset, pose_retrieval_loftr, camera_transf
classnames = ["01Gorilla", "02Unicorn", "03Mallard", "04Turtle", "05Whale", "06Bird", "07Owl", "08Sabertooth",
"09Swan", "10Sheep", "11Pig", "12Zalika", "13Pheonix", "14Elephant", "15Parrot", "16Cat", "17Scorpion",
"18Obesobeso", "19Bear", "20Puppy"]
def main_pose_estimation(cur_class, model_dir_location, k=150, verbose=False, data_dir=None):
model_dir = model_dir_location
data_dir = "MAD-Sim/" if data_dir is None else data_dir
trainset = DefectDataset(data_dir, cur_class, "train", True, True)
train_imgs = torch.cat([a[0][None,...] for a in trainset], dim=0)
train_poses = np.concatenate([np.array(a["transform_matrix"])[None,...] for a in trainset.camera_transforms["frames"]])
train_imgs = torch.movedim(torch.nn.functional.interpolate(train_imgs, (400,400)), 1, 3).numpy()
testset = DefectDataset(data_dir, cur_class, "test", True, True)
camera_angle_x = trainset.camera_angle
# Set up command line argument parser
eval_args = ["-w", "--eval", "-m", model_dir]
parser = ArgumentParser(description="Testing script parameters")
model = ModelParams(parser, sentinel=True)
pipeline = PipelineParams(parser)
parser.add_argument("--iteration", default=-1, type=int)
parser.add_argument("--skip_train", action="store_true")
parser.add_argument("--skip_test", action="store_true")
parser.add_argument("--quiet", action="store_true")
args = get_combined_args(parser, my_cmdline=eval_args)
dataset = model.extract(args)
pipeline = pipeline.extract(args)
bg_color = [1,1,1]
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
if verbose:
save_to = os.path.join(model_dir_location, "3dgs_imgs")
os.makedirs(save_to, exist_ok=True)
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
normal_images = list()
reference_images = list()
all_labels = list()
gt_masks = list()
times = list()
print("STARTING POSE ESTIMATION")
for i in tqdm(range(len(testset))):
cur_path = testset.images[i].split("/")
filename = f"{cur_path[-2]}_{cur_path[-1]}.png"
set_entry = testset[i]
all_labels.append(set_entry[1])
gt_masks.append(set_entry[2].cpu().numpy())
obs_img = torch.movedim(torch.nn.functional.interpolate(set_entry[0][None,...], (400, 400)).squeeze(), 0, 2)
c2w_init = pose_retrieval_loftr(train_imgs * 255, obs_img.numpy() * 255, train_poses)
start.record()
c2w_init[:3, 1:3] *= -1
w2c = np.linalg.inv(c2w_init)
R = np.transpose(w2c[:3,:3]) # R is stored transposed due to 'glm' in CUDA code
T = w2c[:3, 3]
c2w_init[:3,:3] = R
c2w_init[:3,3] = T
c2w_init = torch.from_numpy(c2w_init).type(torch.float).to("cuda")
cam_transf = camera_transf().to("cuda")
cam_transf.train()
optimizer = torch.optim.Adam(cam_transf.parameters(), lr=0.001, betas=(0.9, 0.999))
gaussians = DiffGaussianModel(dataset.sh_degree, c2w_init, cam_transf)
gaussians.load_ply(os.path.join(model_dir,
"point_cloud",
"iteration_" + str(30000),
"point_cloud.ply"))
init_image = None
for iters in range(k):
optimizer.zero_grad()
cur_view = Camera(colmap_id=123, R=c2w_init[:3,:3].cpu().numpy(), T=c2w_init[:3,3].cpu().numpy(),
FoVx=camera_angle_x, FoVy=camera_angle_x,
image=set_entry[0], gt_alpha_mask=None, image_name="aha", uid=123)
rendering = render(cur_view, gaussians, pipeline, background)["render"]
if init_image is None:
init_image = torch.clone(rendering).cpu().detach()
gt_image = set_entry[0].to("cuda")
loss = 0.8 * l1_loss(rendering, gt_image) + 0.2 * (1 - ssim(rendering, gt_image))
loss.backward()
optimizer.step()
new_lrate = 0.01 * (0.8 ** ((iters + 1) / 100))
for param_group in optimizer.param_groups:
param_group['lr'] = new_lrate
if iters == k - 1:
if verbose:
cur_save = os.path.join(save_to, filename.split(".")[0])
os.makedirs(cur_save, exist_ok=True)
torchvision.utils.save_image(set_entry[0].cpu().detach(), os.path.join(cur_save, "gt.png"))
torchvision.utils.save_image(init_image, os.path.join(cur_save, "first_pose.png"))
torchvision.utils.save_image(rendering.cpu().detach(), os.path.join(cur_save, "result.png"))
fig, axs = plt.subplots(2,2, figsize=(10, 6.4))
axs[0, 0].set_title("original image"), axs[0, 1].set_title("first pose")
axs[1,0].set_title(f"iteration {iters}"), axs[1,1].set_title(f"diff")
axs[0,0].imshow(torch.movedim(set_entry[0], 0, 2).cpu().detach())
axs[0,1].imshow(torch.movedim(init_image, 0, 2))
axs[1,0].imshow(torch.movedim(rendering, 0, 2).cpu().detach())
axs[1,1].imshow(torch.movedim(torch.abs(rendering - set_entry[0].to("cuda")), 0, 2).sum(dim=2).cpu().detach())
fig.savefig(os.path.join(save_to, filename))
plt.close(fig)
normal_images.append(set_entry[0].cpu().detach())
reference_images.append(rendering.cpu().detach())
end.record()
torch.cuda.synchronize()
times.append(start.elapsed_time(end))
print("all labels:", all_labels, len(all_labels))
for i in range(5):
print(f"norm: {normal_images[i].shape}. ref: {reference_images[i].shape}. gt: {gt_masks[i].shape}")
assert len(normal_images) == len(reference_images) == len(testset), f"Wrongly sized sets!" \
f"{len(normal_images)}. {len(reference_images)}. {len(testset)}"
assert len(normal_images) == len(gt_masks), f"Wrongly sized sets! {len(normal_images)}. {len(gt_masks)}"
return normal_images, reference_images, all_labels, gt_masks, times
def evaluate_pose_estimation(cur_class, model_dir_location, k=150, verbose=False):
model_dir = model_dir_location
from utils_pose_est import DatasetPose, pose_retrieval_loftr, camera_transf, matrix_to_quaternion
import numpy as np
import wandb
if cur_class in classnames:
data_dir = "MAD-Sim_Subsets/0.8_0/"
else:
data_dir = "nerf_synthetic/"
trainset = DatasetPose(data_dir, cur_class, "train", True)
train_imgs = torch.cat([a[0][None,...] for a in trainset], dim=0)
train_poses = np.concatenate([np.array(a["transform_matrix"])[None,...] for a in trainset.camera_transforms["frames"]])
train_imgs = torch.movedim(torch.nn.functional.interpolate(train_imgs, (400,400)), 1, 3).numpy()
testset = DatasetPose(data_dir, cur_class, "test", True)
camera_angle_x = trainset.camera_angle
all_labels = list()
# Set up command line argument parser
eval_args = ["-w", "--eval", "-m", model_dir]
parser = ArgumentParser(description="Testing script parameters")
model = ModelParams(parser, sentinel=True)
pipeline = PipelineParams(parser)
parser.add_argument("--iteration", default=-1, type=int)
parser.add_argument("--skip_train", action="store_true")
parser.add_argument("--skip_test", action="store_true")
parser.add_argument("--quiet", action="store_true")
args = get_combined_args(parser, my_cmdline=eval_args)
dataset = model.extract(args)
pipeline = pipeline.extract(args)
bg_color = [1,1,1]
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
if verbose:
save_to = os.path.join(model_dir, "img_results")
os.makedirs(save_to, exist_ok=True)
translation_errors = list()
quat_differences = list()
quat_other1 = list()
quat_other2 = list()
coarse_t_error = list()
coarse_rot_error = list()
print("STARTING POSE ESTIMATION")
for i in tqdm(range(len(testset))):
cur_path = testset.images[i].split("/")
filename = f"{cur_path[-2]}_{cur_path[-1]}"
set_entry = testset[i]
gt_pose = set_entry[2]
gt_trans = gt_pose[4:]
gt_quat = torch.nn.functional.normalize(gt_pose[:4], dim=0)
all_labels.append(set_entry[1])
# grab initial pose
obs_img = torch.movedim(torch.nn.functional.interpolate(set_entry[0][None,...], (400, 400)).squeeze(), 0, 2)
c2w_init = pose_retrieval_loftr(train_imgs * 255, obs_img.numpy() * 255, train_poses) # my_init_poses[i] #
orig_init = torch.from_numpy(np.copy(c2w_init))
# prepare pose for input into gaussian splatting
c2w_init[:3, 1:3] *= -1
w2c = np.linalg.inv(c2w_init)
R = np.transpose(w2c[:3,:3]) # R is stored transposed due to 'glm' in CUDA code
T = w2c[:3, 3]
c2w_init[:3,:3] = R
c2w_init[:3,3] = T
c2w_init = torch.from_numpy(c2w_init).type(torch.float).to("cuda")
# set up learnable SE(3) transform and splatting model
cam_transf = camera_transf().to("cuda")
cam_transf.train()
gaussians = DiffGaussianModel(dataset.sh_degree, c2w_init, cam_transf)
gaussians.load_ply(os.path.join(model_dir,
"point_cloud",
"iteration_" + str(30000),
"point_cloud.ply"))
optimizer = torch.optim.Adam(cam_transf.parameters(), lr=0.001, betas=(0.9, 0.999))
init_image = None
# calculate errors of coarse pose estimation
coarse_t_error.append(torch.sqrt(torch.sum((gt_trans - orig_init[:3,3]) ** 2)).item())
init_quat = torch.nn.functional.normalize(matrix_to_quaternion(orig_init[:3,:3]), dim=0).type(torch.float)
coarse_rot_error.append(torch.arccos(torch.abs(torch.dot(gt_quat, init_quat))).item())
for iters in range(k):
optimizer.zero_grad()
cur_view = Camera(colmap_id=123, R=c2w_init[:3,:3].cpu().numpy(), T=c2w_init[:3,3].cpu().numpy(),
FoVx=camera_angle_x, FoVy=camera_angle_x,
image=set_entry[0], gt_alpha_mask=None, image_name="aha", uid=123)
rendering = render(cur_view, gaussians, pipeline, background)["render"]
if init_image is None:
init_image = torch.clone(rendering).cpu().detach()
gt_image = set_entry[0].to("cuda")
loss = 0.8 * l1_loss(rendering, gt_image) + 0.2 * (1 - ssim(rendering, gt_image))
loss.backward()
optimizer.step()
new_lrate = 0.01 * (0.8 ** ((iters + 1) / 100))
for param_group in optimizer.param_groups:
param_group['lr'] = new_lrate
if iters == k - 1:
# construct estimated SE(3) transform
estimation = torch.zeros((4,4), dtype=torch.float32)
estimation[:3,:3] = gaussians.R
estimation[:3,3] = gaussians.T
estimation[3,3] = 1.0
# apply my estimated SE(3) transform to the original initial pose
est_t = estimation.detach() @ orig_init.type(torch.float32)
trans_error = torch.sqrt(torch.sum((gt_trans - est_t[:3,3]) ** 2))
est_quat = torch.nn.functional.normalize(matrix_to_quaternion(est_t[:3,:3]), dim=0)
# measured as ph_1 from "Metrics for 3D Rotations: Comparison and Analysis"
# which ranges from [0,sqrt(2)]
norm_of_diff = min(torch.linalg.vector_norm(gt_quat - est_quat, ord=2),
torch.linalg.vector_norm(gt_quat + est_quat, ord=2))
# phi_2, this one is used for the SplatPose paper
quat_diff = torch.arccos(min(torch.abs(torch.dot(gt_quat, est_quat)), torch.tensor(1.0)))
# phi_3
quat_diff_2 = 1 - torch.abs(torch.dot(gt_quat, est_quat))
translation_errors.append(trans_error.item())
quat_differences.append(norm_of_diff)
quat_other1.append(quat_diff)
quat_other2.append(quat_diff_2)
if verbose:
cur_save = os.path.join(save_to, filename.split(".")[0])
os.makedirs(cur_save, exist_ok=True)
torchvision.utils.save_image(set_entry[0].cpu().detach(), os.path.join(cur_save, "gt.png"))
torchvision.utils.save_image(init_image, os.path.join(cur_save, "first_pose.png"))
torchvision.utils.save_image(rendering.cpu().detach(), os.path.join(cur_save, "result.png"))
continue
ret_dict = {
"translation_error" : translation_errors,
"rotation_error" : quat_differences,
"rotation_error_other1" : quat_other1,
"rotation_error_other2" : quat_other2,
"coarse_t_error" : coarse_t_error,
"coarse_rot_error" : coarse_rot_error
}
my_data = [[testset.images[i], coarse_t_error[i], coarse_rot_error[i], translation_errors[i], quat_differences[i], quat_other1[i], quat_other2[i]] for i in range(len(testset))]
columns = ["path", "coarse_t", "coarse_rot", "trans", "rot_1", "rot_2", "rot_3"]
cur_table = wandb.Table(data=my_data, columns=columns)
# wandb.log({"pose_est": cur_table})
assert all([len(a) == len(testset) for a in ret_dict.values()]), "Wrongly sized lists of errors!"
return ret_dict