Spaces:
Sleeping
Sleeping
File size: 6,519 Bytes
ba90c74 |
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 |
import torch
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from pytorch3d.renderer import (
PerspectiveCameras,
look_at_view_transform
)
# customized
import sys
sys.path.append(".")
from lib.constants import VIEWPOINTS
# ---------------- UTILS ----------------------
def degree_to_radian(d):
return d * np.pi / 180
def radian_to_degree(r):
return 180 * r / np.pi
def xyz_to_polar(xyz):
""" assume y-axis is the up axis """
x, y, z = xyz
theta = 180 * np.arccos(z) / np.pi
phi = 180 * np.arccos(y) / np.pi
return theta, phi
def polar_to_xyz(theta, phi, dist):
""" assume y-axis is the up axis """
theta = degree_to_radian(theta)
phi = degree_to_radian(phi)
x = np.sin(phi) * np.sin(theta) * dist
y = np.cos(phi) * dist
z = np.sin(phi) * np.cos(theta) * dist
return [x, y, z]
# ---------------- VIEWPOINTS ----------------------
def filter_viewpoints(pre_viewpoints: dict, viewpoints: dict):
""" return the binary mask of viewpoints to be filtered """
filter_mask = [0 for _ in viewpoints.keys()]
for i, v in viewpoints.items():
x_v, y_v, z_v = polar_to_xyz(v["azim"], 90 - v["elev"], v["dist"])
for _, pv in pre_viewpoints.items():
x_pv, y_pv, z_pv = polar_to_xyz(pv["azim"], 90 - pv["elev"], pv["dist"])
sim = cosine_similarity(
np.array([[x_v, y_v, z_v]]),
np.array([[x_pv, y_pv, z_pv]])
)[0, 0]
if sim > 0.9:
filter_mask[i] = 1
return filter_mask
def init_viewpoints(mode, sample_space, init_dist, init_elev, principle_directions,
use_principle=True, use_shapenet=False, use_objaverse=False):
if mode == "predefined":
(
dist_list,
elev_list,
azim_list,
sector_list
) = init_predefined_viewpoints(sample_space, init_dist, init_elev)
elif mode == "hemisphere":
(
dist_list,
elev_list,
azim_list,
sector_list
) = init_hemisphere_viewpoints(sample_space, init_dist)
else:
raise NotImplementedError()
# punishments for views -> in case always selecting the same view
view_punishments = [1 for _ in range(len(dist_list))]
if use_principle:
(
dist_list,
elev_list,
azim_list,
sector_list,
view_punishments
) = init_principle_viewpoints(
principle_directions,
dist_list,
elev_list,
azim_list,
sector_list,
view_punishments,
use_shapenet,
use_objaverse
)
return dist_list, elev_list, azim_list, sector_list, view_punishments
def init_principle_viewpoints(
principle_directions,
dist_list,
elev_list,
azim_list,
sector_list,
view_punishments,
use_shapenet=False,
use_objaverse=False
):
if use_shapenet:
key = "shapenet"
pre_elev_list = [v for v in VIEWPOINTS[key]["elev"]]
pre_azim_list = [v for v in VIEWPOINTS[key]["azim"]]
pre_sector_list = [v for v in VIEWPOINTS[key]["sector"]]
num_principle = 10
pre_dist_list = [dist_list[0] for _ in range(num_principle)]
pre_view_punishments = [0 for _ in range(num_principle)]
elif use_objaverse:
key = "objaverse"
pre_elev_list = [v for v in VIEWPOINTS[key]["elev"]]
pre_azim_list = [v for v in VIEWPOINTS[key]["azim"]]
pre_sector_list = [v for v in VIEWPOINTS[key]["sector"]]
num_principle = 10
pre_dist_list = [dist_list[0] for _ in range(num_principle)]
pre_view_punishments = [0 for _ in range(num_principle)]
else:
num_principle = 6
pre_elev_list = [v for v in VIEWPOINTS[num_principle]["elev"]]
pre_azim_list = [v for v in VIEWPOINTS[num_principle]["azim"]]
pre_sector_list = [v for v in VIEWPOINTS[num_principle]["sector"]]
pre_dist_list = [dist_list[0] for _ in range(num_principle)]
pre_view_punishments = [0 for _ in range(num_principle)]
dist_list = pre_dist_list + dist_list
elev_list = pre_elev_list + elev_list
azim_list = pre_azim_list + azim_list
sector_list = pre_sector_list + sector_list
view_punishments = pre_view_punishments + view_punishments
return dist_list, elev_list, azim_list, sector_list, view_punishments
def init_predefined_viewpoints(sample_space, init_dist, init_elev):
viewpoints = VIEWPOINTS[sample_space]
assert sample_space == len(viewpoints["sector"])
dist_list = [init_dist for _ in range(sample_space)] # always the same dist
elev_list = [viewpoints["elev"][i] for i in range(sample_space)]
azim_list = [viewpoints["azim"][i] for i in range(sample_space)]
sector_list = [viewpoints["sector"][i] for i in range(sample_space)]
return dist_list, elev_list, azim_list, sector_list
def init_hemisphere_viewpoints(sample_space, init_dist):
"""
y is up-axis
"""
num_points = 2 * sample_space
ga = np.pi * (3. - np.sqrt(5.)) # golden angle in radians
flags = []
elev_list = [] # degree
azim_list = [] # degree
for i in range(num_points):
y = 1 - (i / float(num_points - 1)) * 2 # y goes from 1 to -1
# only take the north hemisphere
if y >= 0:
flags.append(True)
else:
flags.append(False)
theta = ga * i # golden angle increment
elev_list.append(radian_to_degree(np.arcsin(y)))
azim_list.append(radian_to_degree(theta))
radius = np.sqrt(1 - y * y) # radius at y
x = np.cos(theta) * radius
z = np.sin(theta) * radius
elev_list = [elev_list[i] for i in range(len(elev_list)) if flags[i]]
azim_list = [azim_list[i] for i in range(len(azim_list)) if flags[i]]
dist_list = [init_dist for _ in elev_list]
sector_list = ["good" for _ in elev_list] # HACK don't define sector names for now
return dist_list, elev_list, azim_list, sector_list
# ---------------- CAMERAS ----------------------
def init_camera(dist, elev, azim, image_size, device):
R, T = look_at_view_transform(dist, elev, azim)
image_size = torch.tensor([image_size, image_size]).unsqueeze(0)
cameras = PerspectiveCameras(R=R, T=T, device=device, image_size=image_size)
return cameras |