An efficient render repo based on pytorch3d.

Install

• Install the torch version you want, and find correponding pytorch3d in link.

pip install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
wget https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch3d/linux-64/pytorch3d-0.7.4-py38_cu113_pyt1100.tar.bz2;
pip install fvcore;
pip install iopath;
conda install --use-local pytorch3d-0.7.4-py38_cu113_pyt1100.tar.bz2;

• When you have runnable torch and pytorch3d, install this repo using pip:

pip install git+https://github.com/WenjiaWang0312/torch3d_render.git

Important tools

load, combine, trasnform and save meshes

from t3drender.mesh_utils import load_plys_as_meshes, save_meshes_as_plys, load_objs_as_meshes, save_meshes_as_objs, join_meshes_as_scene, join_meshes_as_batch, join_batch_meshes_as_scene
from t3drender.transforms.affine import transform_points

# load
device = torch.device('cuda')
meshes = load_plys_as_meshes(ply_path, device) # 1, V, 3
meshes = load_plys_as_meshes([ply_path1, ply_path2], device) # 2, V, 3
save_meshes_as_plys([ply_path1, ply_path2], meshes)

# combine
sphere_mesh = ico_sphere(5, device) # 1, V, 3
meshes = join_meshes_as_scene([sphere_mesh, sphere_mesh]) # 1, 2V, 3
batch_sphere = joint_meshes_as_batch([sphere_mesh, sphere_mesh]) # 2, V, 3
meshes = joint_batch_meshes_as_scene([batch_sphere, batch_sphere]) # 2, 2V, 3

# transform
T = torch.eye(4, 4)[None].to(device)
sphere_mesh = sphere_mesh.update_padded(transform_points(sphere_mesh.verts_padded(), T))

# save
save_meshes_as_objs('1.obj', sphere_mesh)
save_meshes_as_objs(['1.obj', '2.obj'], batch_sphere)

conversion

Cameras

Examples

Render Mesh

import torch
import numpy as np
from pytorch3d.utils import ico_sphere
from t3drender.render.render_functions import render_rgb
from pytorch3d.renderer.mesh.textures import TexturesVertex
import cv2
device = 0
# or
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# You need to konw where to put your meshes
meshes = ico_sphere(5, device)
moved_verts = meshes.verts_padded()
moved_verts[..., 2] += 2
meshes = meshes.update_padded(moved_verts)

# The default camera setting does not set extrinsic matrix, for advanced camera setting you need to feed R, T matrix into cameras
meshes.textures = TexturesVertex(torch.randn_like(meshes.verts_padded()))
image_tensors = render_rgb(meshes, device=device, resolution=(512, 512), fov=90, batch_size=30, verbose=True)
image = image_tensors[..., :3].detach().cpu().numpy() * 256
cv2.imwrite('sphere.png', image[0])

render multi gpu

import torch
import numpy as np
from pytorch3d.utils import ico_sphere
from t3drender.render.render_functions import render_rgb
from pytorch3d.renderer.mesh.textures import TexturesVertex
import cv2

device = [0, 1, 2]
# or
device = [torch.device('cuda:0'), torch.device('cuda:1'), torch.device('cuda:2')]

# You need to konw where to put your meshes
meshes = ico_sphere(5, device)
meshes = meshes.extend(100)
moved_verts = meshes.verts_padded()
moved_verts[..., 2] += 2
meshes = meshes.update_padded(moved_verts)

# The default camera setting does not set extrinsic matrix, for advanced camera setting you need to feed R, T matrix into cameras
meshes.textures = TexturesVertex(torch.randn_like(meshes.verts_padded()))
meshes = meshes.extend(100)
image_tensors = render_rgb(meshes, device=device, resolution=(512, 512), fov=90, batch_size=5, verbose=True)
image = image_tensors[..., :3].detach().cpu().numpy() * 256

Render Normal

import torch
import numpy as np
from pytorch3d.utils import ico_sphere
from t3drender.render.render_functions import render_normal
import cv2
device = 0

# You need to konw where to put your meshes
meshes = ico_sphere(5, device)
moved_verts = meshes.verts_padded()
moved_verts[..., 2] += 2
meshes = meshes.update_padded(moved_verts)

# The default camera setting does not set extrinsic matrix, for advanced camera setting you need to feed R, T matrix into cameras
image_tensors = render_normal(meshes, device=device, resolution=(512, 512), fov=90, batch_size=30, verbose=True)
image_tensors = image_tensors / 2 + 0.5
image = image_tensors[..., :3].detach().cpu().numpy() * 256
cv2.imwrite('sphere.png', image[0])

Render Depth

import torch
import numpy as np
from pytorch3d.utils import ico_sphere
from t3drender.render.render_functions import render_depth_orthographic
import cv2
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

meshes = ico_sphere(5, device)

# The default camera setting does not set extrinsic matrix, for advanced camera setting you need to feed R, T matrix into cameras
image_tensors = render_depth_orthographic(meshes, device=device, resolution=(512, 512), batch_size=30, verbose=True)
image_tensors = image_tensors / image_tensors.max()
image = image_tensors[..., :3].detach().cpu().numpy() * 256
cv2.imwrite('sphere.png', image[0])

Render PointCloud

Render Silhoutte

Render Segmentation Mask