渲染概述

1from typing import Union
2
3from pytorch3d.renderer import (
4    RasterizationSettings,
5    MeshRenderer,
6    MeshRasterizer,
7    SoftPhongShader
8)
9
10def create_renderer(
11    img_dims: Union[int, tuple[int, int]],
12    blur_radius: float,
13    faces_per_pixel: int
14) -> MeshRenderer:
15    """
16    创建一个MeshRenderer实例
17
18    参数:
19        img_dims: 要生成图像的尺寸
20        blur_radius: 多边形面片模糊的半径
21        faces_per_pixel: 每个像素参与贡献的面片数量
22    """
23    raster_settings_diff = RasterizationSettings(
24        image_size=img_dims,
25        blur_radius=blur_radius,
26        faces_per_pixel=faces_per_pixel,
27    )
28    return MeshRenderer(
29        rasterizer=MeshRasterizer(
30            cameras=cameras,
31            raster_settings=raster_settings_diff
32        ),
33        shader = SoftPhongShader(
34            cameras=cameras,
35            lights=lights
36        )
37    )
38
39
40# 图像分辨率
41img_dims = 256
42
43# 非可微情况
44ndif_renderer = create_renderer(img_dims, 0.0, 1)
45images_ndif = ndif_renderer(mesh)
46
47# 可微情况
48diff_renderer = create_renderer(img_dims, np.log(1. / 1e-4 - 1.) * 1e-4, 20)
49images_diff = diff_renderer(mesh)
50
51# 渲染图像
52compare_renders(images_ndif[0, ..., :3].numpy(), images_diff[0, ..., :3].numpy(),
53                title_1='非可微情况', title_2='可微情况')

1# 设置渲染器
2raster_settings_diff = RasterizationSettings(
3    image_size=256,
4    blur_radius=np.log(1. / 1e-4 - 1.) * 1e-4,
5    faces_per_pixel=20,
6)
7diff_renderer = MeshRenderer(
8    rasterizer=MeshRasterizer(
9        cameras=cameras,
10        raster_settings=raster_settings_diff
11    ),
12    shader=SoftSilhouetteShader()
13)
14
15# 渲染中性姿态下的图像
16images_neutral = diff_renderer(mesh)
17
18# 渲染偏移后的姿态
19T_offset = torch.full(T.shape, 0.5, requires_grad=True)
20cameras_shifted = FoVPerspectiveCameras(R=R, T=T + T_offset)
21images_shifted = diff_renderer(mesh, cameras=cameras_shifted)
22
23# 显示渲染结果
24compare_renders(images_neutral[0, ..., 3].numpy(), images_shifted[0, ..., 3].detach().numpy(),
25                title_1='中性姿态', title_2='偏移后')
26
27# 创建优化器
28optimizer = torch.optim.SGD([T_offset], lr=1.0, momentum=0.9)
29
30# 计算损失
31optimizer.zero_grad()
32loss = torch.mean((images_neutral[...,-1] - images_shifted[..., -1]) ** 2)
33loss.backward()
34print(T_offset.grad)
35optimizer.step()