-
Notifications
You must be signed in to change notification settings - Fork 0
/
render_video.py
247 lines (189 loc) · 10.2 KB
/
render_video.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
import imageio
# models
from models import *
from renderer import *
from data.ray_utils import get_rays
from scipy.spatial.transform import Rotation as R
from tqdm import tqdm
# pytorch-lightning
from data.ray_utils import ray_marcher
from data.llff import LLFFDataset
torch.cuda.set_device(0)
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
def decode_batch(batch):
rays = batch['rays'] # (B, 8)
rgbs = batch['rgbs'] # (B, 3)
return rays, rgbs
def unpreprocess(data, shape=(1, 1, 3, 1, 1)):
# to unnormalize image for visualization
# data N V C H W
device = data.device
mean = torch.tensor([-0.485 / 0.229, -0.456 / 0.224, -0.406 / 0.225]).view(*shape).to(device)
std = torch.tensor([1 / 0.229, 1 / 0.224, 1 / 0.225]).view(*shape).to(device)
return (data - mean) / std
def normalize(x):
return x / np.linalg.norm(x, axis=-1, keepdims=True)
def viewmatrix(z, up, pos):
vec2 = normalize(z)
vec1_avg = up
vec0 = normalize(np.cross(vec1_avg, vec2))
vec1 = normalize(np.cross(vec2, vec0))
m = np.eye(4)
m[:3] = np.stack([vec0, vec1, vec2, pos], 1)
return m
def ptstocam(pts, c2w):
tt = np.matmul(c2w[:3, :3].T, (pts - c2w[:3, 3])[..., np.newaxis])[..., 0]
return tt
def poses_avg(poses):
center = poses[:, :3, 3].mean(0)
vec2 = normalize(poses[:, :3, 2].sum(0))
up = poses[:, :3, 1].sum(0)
c2w = viewmatrix(vec2, up, center)
return c2w
def render_path_spiral(c2w, up, rads, focal, zdelta, zrate, N_rots=2, N=120):
render_poses = []
rads = np.array(list(rads) + [1.])
for theta in np.linspace(0., 2. * np.pi * N_rots, N + 1)[:-1]:
c = np.dot(c2w[:3, :4], np.array([np.cos(theta), -np.sin(theta), -np.sin(theta * zrate), 1.]) * rads)
z = normalize(c - np.dot(c2w[:3, :4], np.array([0, 0, -focal, 1.])))
render_poses.append(viewmatrix(z, up, c))
return render_poses
def get_spiral(c2ws_all, near_far, rads_scale=0.5, N_views=120):
# center pose
c2w = poses_avg(c2ws_all)
# Get average pose
up = normalize(c2ws_all[:, :3, 1].sum(0))
# Find a reasonable "focus depth" for this dataset
close_depth, inf_depth = near_far
dt = .75
mean_dz = 1. / (((1. - dt) / close_depth + dt / inf_depth))
focal = mean_dz
# Get radii for spiral path
zdelta = close_depth * .2
tt = c2ws_all[:, :3, 3] - c2w[:3, 3][None]
rads = np.percentile(np.abs(tt), 70, 0) * rads_scale
render_poses = render_path_spiral(c2w, up, rads, focal, zdelta, zrate=.5, N=N_views)
return np.stack(render_poses)
def position2angle(position):
''' nx3 '''
position = normalize(position)
theta = np.arccos(position[:, 2]) / np.pi * 180
phi = np.arctan2(position[:, 1], position[:, 0]) / np.pi * 180
return [theta, phi]
def pose_spherical_nerf(euler, radius=4.0):
c2ws_render = np.eye(4)
c2ws_render[:3, :3] = R.from_euler('xyz', euler, degrees=True).as_matrix()
c2ws_render[:3, 3] = c2ws_render[:3, :3] @ np.array([0.0, 0.0, -radius])
return c2ws_render
def nerf_video_path(c2ws, theta_range=10, phi_range=20, N_views=120):
rotvec = []
for i in range(c2ws.shape[0]):
r = R.from_matrix(c2ws[i, :3, :3])
euler_ange = r.as_euler('xyz', degrees=True).reshape(1, 3)
if i:
mask = np.abs(euler_ange - rotvec[0]) > 180
euler_ange[mask] += 360.0
rotvec.append(euler_ange)
rotvec = np.mean(np.stack(rotvec), axis=0)
render_poses = [pose_spherical_nerf(rotvec + np.array([angle, 0.0, -phi_range]), 4.0) for angle in
np.linspace(-theta_range, theta_range, N_views // 4, endpoint=False)]
render_poses += [pose_spherical_nerf(rotvec + np.array([theta_range, 0.0, angle]), 4.0) for angle in
np.linspace(-phi_range, phi_range, N_views // 4, endpoint=False)]
render_poses += [pose_spherical_nerf(rotvec + np.array([angle, 0.0, phi_range]), 4.0) for angle in
np.linspace(theta_range, -theta_range, N_views // 4, endpoint=False)]
render_poses += [pose_spherical_nerf(rotvec + np.array([-theta_range, 0.0, angle]), 4.0) for angle in
np.linspace(phi_range, -phi_range, N_views // 4, endpoint=False)]
render_poses = torch.from_numpy(np.stack(render_poses)).float().to(device)
return render_poses
def render_video(args):
for i_scene, scene in enumerate([args.datadir.split('/')[-1]]):
if args.is_finetuned:
args.ckpt = f'./runs_fine_tuning/{scene}/ckpts/latest.tar'
args.video_name = 'DSft_'
else:
args.video_name = ''
args.use_viewdirs = True
args.feat_dim = 8 + 3 * 4
# create models
if i_scene == 0 or args.is_finetuned:
# Create nerf model
render_kwargs_train, render_kwargs_test, start, grad_vars = \
create_nerf_mvs(args, use_mvs=True, dir_embedder=False, pts_embedder=True)
filter_keys(render_kwargs_train)
# Create mvs model
MVSNet = render_kwargs_train['network_mvs']
render_kwargs_train.pop('network_mvs')
datatype = 'val'
pad = 24 # the padding value should be same as your finetuning ckpt
args.chunk = 5120
dataset = LLFFDataset(args, split=datatype)
save_dir = f'./results'
os.makedirs(save_dir, exist_ok=True)
MVSNet.train()
MVSNet = MVSNet.cuda()
with torch.no_grad():
c2ws_all = dataset.poses
if args.is_finetuned:
# large baseline
imgs_source, proj_mats, near_far_source, pose_source = dataset.read_source_views(device=device)
volume_feature = torch.load(args.ckpt)['volume']['feat_volume']
volume_feature = RefVolume(volume_feature.detach()).cuda()
pad *= args.imgScale_test
pair_idx = torch.load('configs/pairs.th')[f'{scene}_train']
c2ws_render = get_spiral(c2ws_all[pair_idx], near_far_source, rads_scale=0.6,
N_views=180) # you can enlarge the rads_scale if you want to render larger baseline
else:
# neighboring views with position distance
imgs_source, proj_mats, near_far_source, pose_source = dataset.read_source_views(device=device)
volume_feature, _, _ = MVSNet(imgs_source, proj_mats, near_far_source, pad=pad, lindisp=args.use_disp)
pad *= args.imgScale_test
pair_idx = torch.load('configs/pairs.th')[f'{scene}_train']
c2ws_render = get_spiral(c2ws_all[pair_idx], near_far_source, rads_scale=0.6,
N_views=180) # you can enlarge the rads_scale if you want to render larger baseline
c2ws_render = torch.from_numpy(np.stack(c2ws_render)).float().to(device)
imgs_source = unpreprocess(imgs_source)
try:
tqdm._instances.clear()
except Exception:
pass
frames = []
img_directions = dataset.directions.to(device)
for i, c2w in enumerate(tqdm(c2ws_render)):
torch.cuda.empty_cache()
rays_o, rays_d = get_rays(img_directions, c2w) # both (h*w, 3)
rays = torch.cat([rays_o, rays_d,
near_far_source[0] * torch.ones_like(rays_o[:, :1]),
near_far_source[1] * torch.ones_like(rays_o[:, :1])],
1).to(device) # (H*W, 3)
N_rays_all = rays.shape[0]
rgb_rays, depth_rays_preds = [], []
for chunk_idx in range(N_rays_all // args.chunk + int(N_rays_all % args.chunk > 0)):
xyz_coarse_sampled, rays_o, rays_d, z_vals = ray_marcher(
rays[chunk_idx * args.chunk:(chunk_idx + 1) * args.chunk],
N_samples=args.N_samples, lindisp=args.use_disp)
# Converting world coordinate to ndc coordinate
H, W = imgs_source.shape[-2:]
inv_scale = torch.tensor([W - 1, H - 1]).to(device)
w2c_ref, intrinsic_ref = pose_source['w2cs'][0], pose_source['intrinsics'][0].clone()
xyz_NDC = get_ndc_coordinate(w2c_ref, intrinsic_ref, xyz_coarse_sampled, inv_scale,
near=near_far_source[0], far=near_far_source[1], pad=pad,
lindisp=args.use_disp)
# rendering
rgb, disp, acc, depth_pred, alpha, extras = rendering(args, pose_source, xyz_coarse_sampled,
xyz_NDC, z_vals, rays_o, rays_d,
volume_feature, imgs_source,
**render_kwargs_train)
rgb, depth_pred = torch.clamp(rgb.cpu(), 0, 1.0).numpy(), depth_pred.cpu().numpy()
rgb_rays.append(rgb)
depth_rays_preds.append(depth_pred)
depth_rays_preds = np.concatenate(depth_rays_preds).reshape(H, W)
depth_rays_preds, _ = visualize_depth_numpy(depth_rays_preds, near_far_source)
rgb_rays = np.concatenate(rgb_rays).reshape(H, W, 3)
H_crop, W_crop = np.array(rgb_rays.shape[:2]) // 20
rgb_rays = rgb_rays[H_crop:-H_crop, W_crop:-W_crop]
depth_rays_preds = depth_rays_preds[H_crop:-H_crop, W_crop:-W_crop]
img_vis = np.concatenate((rgb_rays * 255, depth_rays_preds), axis=1)
frames.append(img_vis.astype('uint8'))
imageio.mimwrite(f'{save_dir}/{args.video_name}{scene}.mov', np.stack(frames), fps=30, quality=10)
os.system(f"ffmpeg -i {save_dir}/{args.video_name}{scene}.mov -vcodec h264 -acodec mp2 {save_dir}/{args.video_name}{scene}.mp4")
os.system(f"rm {save_dir}/{args.video_name}{scene}.mov")