Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,762 Bytes
476e0f0 7760d2d 476e0f0 7760d2d 476e0f0 7760d2d 476e0f0 7760d2d 476e0f0 7760d2d 476e0f0 |
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 |
from typing import *
from torch import Tensor
import os
import numpy as np
from plyfile import PlyData, PlyElement
import torch
from diff_gaussian_rasterization import (
GaussianRasterizationSettings,
GaussianRasterizer,
)
class Camera:
def __init__(self,
C2W: Tensor, fxfycxcy: Tensor, h: int, w: int,
znear: float = 0.01, zfar: float = 100.,
):
self.fxfycxcy = fxfycxcy.clone().float()
self.C2W = C2W.clone().float()
self.W2C = self.C2W.inverse()
self.znear = znear
self.zfar = zfar
self.h = h
self.w = w
fx, fy, cx, cy = self.fxfycxcy[0], self.fxfycxcy[1], self.fxfycxcy[2], self.fxfycxcy[3]
self.tanfovX = 1 / (2 * fx) # `tanHalfFovX` actually
self.tanfovY = 1 / (2 * fy) # `tanHalfFovY` actually
self.fovX = 2 * torch.atan(self.tanfovX)
self.fovY = 2 * torch.atan(self.tanfovY)
self.shiftX = 2 * cx - 1
self.shiftY = 2 * cy - 1
def getProjectionMatrix(znear, zfar, fovX, fovY, shiftX, shiftY):
tanHalfFovY = torch.tan((fovY / 2))
tanHalfFovX = torch.tan((fovX / 2))
top = tanHalfFovY * znear
bottom = -top
right = tanHalfFovX * znear
left = -right
P = torch.zeros(4, 4, device=fovX.device)
z_sign = 1
P[0, 0] = 2 * znear / (right - left)
P[1, 1] = 2 * znear / (top - bottom)
P[0, 2] = (right + left) / (right - left) + shiftX
P[1, 2] = (top + bottom) / (top - bottom) + shiftY
P[3, 2] = z_sign
P[2, 2] = z_sign * zfar / (zfar - znear)
P[2, 3] = -(zfar * znear) / (zfar - znear)
return P
self.world_view_transform = self.W2C.transpose(0, 1)
self.projection_matrix = getProjectionMatrix(self.znear, self.zfar, self.fovX, self.fovY, self.shiftX, self.shiftY).transpose(0, 1)
self.full_proj_transform = self.world_view_transform @ self.projection_matrix
self.camera_center = self.C2W[:3, 3]
class GaussianModel:
def __init__(self):
self.xyz = None
self.rgb = None
self.scale = None
self.rotation = None
self.opacity = None
self.sh_degree = 0
def set_data(self, xyz: Tensor, rgb: Tensor, scale: Tensor, rotation: Tensor, opacity: Tensor):
self.xyz = xyz
self.rgb = rgb
self.scale = scale
self.rotation = rotation
self.opacity = opacity
return self
def to(self, device: torch.device = None, dtype: torch.dtype = None) -> "GaussianModel":
self.xyz = self.xyz.to(device, dtype)
self.rgb = self.rgb.to(device, dtype)
self.scale = self.scale.to(device, dtype)
self.rotation = self.rotation.to(device, dtype)
self.opacity = self.opacity.to(device, dtype)
return self
def save_ply(self, path: str, opacity_threshold: float = 0.):
os.makedirs(os.path.dirname(path), exist_ok=True)
xyz = self.xyz.detach().cpu().numpy()
f_dc = self.rgb.detach().cpu().numpy()
rgb = (f_dc * 255.).clip(0., 255.).astype(np.uint8)
opacity = self.opacity.detach().cpu().numpy()
scale = self.scale.detach().cpu().numpy()
rotation = self.rotation.detach().cpu().numpy()
# Filter out points with low opacity
mask = (opacity > opacity_threshold).squeeze()
xyz = xyz[mask]
f_dc = f_dc[mask]
opacity = opacity[mask]
scale = scale[mask]
rotation = rotation[mask]
rgb = rgb[mask]
dtype_full = [(attribute, "f4") for attribute in self._construct_list_of_attributes()]
dtype_full.extend([("red", "u1"), ("green", "u1"), ("blue", "u1")])
elements = np.empty(xyz.shape[0], dtype=dtype_full)
attributes = np.concatenate((xyz, f_dc, opacity, scale, rotation, rgb), axis=1)
elements[:] = list(map(tuple, attributes))
el = PlyElement.describe(elements, "vertex")
PlyData([el]).write(path)
def load_ply(self, path: str):
plydata = PlyData.read(path)
xyz = np.stack((
np.asarray(plydata.elements[0]["x"]),
np.asarray(plydata.elements[0]["y"]),
np.asarray(plydata.elements[0]["z"]),
), axis=1)
f_dc = np.stack((
np.asarray(plydata.elements[0]["f_dc_0"]),
np.asarray(plydata.elements[0]["f_dc_1"]),
np.asarray(plydata.elements[0]["f_dc_2"]),
), axis=1)
opacity = np.asarray(plydata.elements[0]["opacity"])[..., np.newaxis]
scale = np.stack((
np.asarray(plydata.elements[0]["scale_0"]),
np.asarray(plydata.elements[0]["scale_1"]),
np.asarray(plydata.elements[0]["scale_2"]),
), axis=1)
rotation = np.stack((
np.asarray(plydata.elements[0]["rot_0"]),
np.asarray(plydata.elements[0]["rot_1"]),
np.asarray(plydata.elements[0]["rot_2"]),
np.asarray(plydata.elements[0]["rot_3"]),
), axis=1)
self.xyz = torch.from_numpy(xyz).float()
self.rgb = torch.from_numpy(f_dc).float()
self.opacity = torch.from_numpy(opacity).float()
self.scale = torch.from_numpy(scale).float()
self.rotation = torch.from_numpy(rotation).float()
def _construct_list_of_attributes(self):
l = ["x", "y", "z"]
for i in range(self.rgb.shape[1]):
l.append(f"f_dc_{i}")
l.append("opacity")
for i in range(self.scale.shape[1]):
l.append(f"scale_{i}")
for i in range(self.rotation.shape[1]):
l.append(f"rot_{i}")
return l
def render(
pc: GaussianModel,
height: int,
width: int,
C2W: Tensor,
fxfycxcy: Tensor,
znear: float = 0.01,
zfar: float = 100.,
bg_color: Union[Tensor, Tuple[float, float, float]] = (1., 1., 1.),
scaling_modifier: float = 1.,
render_dn: bool = False,
):
viewpoint_camera = Camera(C2W, fxfycxcy, height, width, znear, zfar)
if not isinstance(bg_color, Tensor):
bg_color = torch.tensor(list(bg_color), dtype=torch.float32, device=C2W.device)
else:
bg_color = bg_color.to(C2W.device, dtype=torch.float32)
pc = pc.to(dtype=torch.float32)
subpixel_offset = torch.zeros((int(viewpoint_camera.h), int(viewpoint_camera.w), 2), dtype=torch.float32, device="cuda")
raster_settings = GaussianRasterizationSettings(
image_height=int(viewpoint_camera.h),
image_width=int(viewpoint_camera.w),
tanfovx=viewpoint_camera.tanfovX,
tanfovy=viewpoint_camera.tanfovY,
subpixel_offset=subpixel_offset,
kernel_size=0., # cf. Mip-Splatting; not used
bg=bg_color,
scale_modifier=scaling_modifier,
viewmatrix=viewpoint_camera.world_view_transform,
projmatrix=viewpoint_camera.full_proj_transform,
sh_degree=pc.sh_degree,
campos=viewpoint_camera.camera_center,
prefiltered=False,
debug=False,
)
alpha = normal = depth = None
rasterizer = GaussianRasterizer(raster_settings=raster_settings)
# Rasterize visible Gaussians to image, obtain their radii (on screen).
image, radii = rasterizer( # not used: radii, coord, mcoord, mdepth
means3D=pc.xyz,
means2D=torch.zeros_like(pc.xyz, dtype=torch.float32, device=pc.xyz.device),
shs=None,
colors_precomp=pc.rgb,
opacities=pc.opacity,
scales=pc.scale,
rotations=pc.rotation,
cov3D_precomp=None,
)
return {
"image": image,
"alpha": alpha,
"depth": depth,
"normal": normal,
}
|