体素

1import torch
2
3
4n_batch = 1  # 批次大小
5width = 8    # 每个维度的宽度
6val = 0      # 初始值
7
8vgrid = torch.full([n_batch, width, width, width], val)
9print(vgrid.shape)

1import warnings
2import torch
3from pytorch3d.ops import cubify
4from pytorch3d.renderer import (
5    look_at_view_transform,
6    PerspectiveCameras
7)
8import matplotlib.pyplot as plt
9
10from renderer import render
11
12
13# 参数设置
14n_batch = 1   # 批次大小
15width = 3     # 每个维度的宽度
16val = 1.0     # 初始值
17thresh = 0.5  # 阈值
18
19# 创建一个仅包含单个非零元素的体素网格
20vgrid = torch.full([n_batch, width, width, width], 0.)
21vgrid[:, 1, 1, 1] = val
22
23# 应用cubify函数（当前版本会触发UserWarning）
24with warnings.catch_warnings():
25    warnings.simplefilter('ignore')
26    mesh = cubify(vgrid, thresh)
27
28print(mesh.verts_padded())
29
30# 渲染立方体图像
31R, T = look_at_view_transform(dist=10.0, elev=45, azim=45)
32cameras = PerspectiveCameras(focal_length=8.0, R=R, T=T)
33img = render(mesh, cameras)[0, ..., :3].numpy()
34plt.imshow(img)
35plt.savefig('./output/render.png')

1import torch
2from pytorch3d.structures import Volumes
3from pytorch3d.renderer import (
4    VolumeRenderer,
5    NDCMultinomialRaysampler,
6    EmissionAbsorptionRaymarcher
7)
8from pytorch3d.renderer import (
9    look_at_view_transform,
10    FoVPerspectiveCameras,
11)
12import matplotlib.pyplot as plt
13
14device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
15
16
17# 设置参数
18n_batch = 1        # 批处理中的体积数量
19render_size = 128  # 渲染图像尺寸
20volume_size = 64   # 体积尺寸（即每个维度上的体素数）
21min_depth = 0.1    # 相机到采样点的最小深度
22max_depth = 6.1    # 相机到采样点的最大深度
23volume_extent_world = max_depth - min_depth  # 世界坐标系下体积的空间范围
24
25# 创建用于体积渲染的光线采样器
26raysampler = NDCMultinomialRaysampler(
27    image_width=render_size,
28    image_height=render_size,
29    n_pts_per_ray=64,
30    min_depth=0.1,
31    max_depth=volume_extent_world,
32)
33
34# 创建光线行进器，用于对密度和特征应用发射-吸收算法
35raymarcher = EmissionAbsorptionRaymarcher()
36
37# 创建体积渲染器
38renderer = VolumeRenderer(raysampler=raysampler, raymarcher=raymarcher)
39
40# 创建张量存储体积密度
41densities = torch.full([n_batch, 1, *[volume_size] * 3], 0.1, device=device)
42
43# 创建沿y轴呈红色渐变的颜色张量
44colors = torch.stack([
45    torch.linspace(1, 0., volume_size, device=device),
46    torch.zeros([volume_size], device=device),
47    torch.zeros([volume_size], device=device)
48])
49
50# 将颜色张量扩展为体积的完整尺寸
51colors = colors.view(n_batch, 3, 1, volume_size, 1).expand(n_batch, 3, *[volume_size] * 3)
52
53# 创建相机视图变换
54R, T = look_at_view_transform(3.1, 0, 0)
55cameras = FoVPerspectiveCameras(device=device, R=R, T=T)
56
57# 创建Volumes实例
58volumes = Volumes(
59    densities=densities,
60    features=colors,
61    voxel_size=0.25 * (volume_extent_world / volume_size),
62)
63
64# 使用指定相机渲染该体积
65img = renderer(cameras=cameras, volumes=volumes)[0]
66
67# 显示并保存渲染结果
68plt.imshow(img.cpu().numpy()[0, ..., :3])
69plt.savefig('./output/render.png')