pytorch_nicp编译

NICP学习：(117条消息) 激光slam理论与实践 - 第四章 前端配准（ICP）_ppipp1109的博客-CSDN博客

(117条消息) 激光SLAM:ICP升级版本NICP(Normal ICP)原理+法向量即曲率计算代码_月照银海似蛟龙的博客-CSDN博客

1.nicp pytorch版编译：  GitHub - wuhaozhe/pytorch-nicp: GPU Accelerated Non-rigid ICP for surface registration      

2.按照说明，先安装pyotorch3d: pytorch3d/INSTALL.md at main · facebookresearch/pytorch3d · GitHub ,执行以下语句：

conda create -n pytorch3d python=3.9

conda activate pytorch3d

conda install -c pytorch pytorch=1.9.1 torchvision cudatoolkit=10.2

conda install -c fvcore -c iopath -c conda-forge fvcore iopath

conda install -c bottler nvidiacub

conda install pytorch3d -c pytorch3d

3.安装依赖项：pip install -r requirements.txt

其中：sklearn>=0.24.2 应改成：scikit-learn>=0.24.2

环境要求open3d最低的版本为0.14.1，于是open3d==0.14.1

4.运行python demo_nicp.py

(1)报错：libc10_cuda.so: cannot open shared object file: No such file or directory

ImportError: libc10_cuda.so After installing using Intall.MD guide · Issue #1152 · facebookresearch/pytorch3d · GitHub，这个回答用的pytorch==1.10.0,删除环境安装conda install -c pytorch pytorch=1.10.0 torchvision cudatoolkit=11.1，还是报错

(2)报错找不到pytorch3d 模型，根据pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu113_pyt1110/download.html 安装pytorch3d,配置环境为：python=3.8,conda install -c pytorch pytorch=1.11.0 torchvision cudatoolkit=11.3，还是报错：libc10_cuda.so: cannot open shared object file

(3)ImportError: libc10_cuda.so: cannot open shared object file: No such file or directory · Issue #1060 · facebookresearch/pytorch3d · GitHub的回答，pytoch不应该是conda-forge，应该这样:

pytorch 1.6.0  py3.8_cuda10.2.89_cudnn7.6.5_0  pytorch

而我是:

pytorch   1.11.0    cpu_py38h39c826d_1    conda-forge

于是下载touch源码安装：下载

torch-1.11.0+cu113-cp38-cp38-linux_x86_64.whl和torchvision-0.12.0+cu113-cp38-cp38-linux_x86_64.whl

(4)安装后报错：libstdc++.so.6: version `GLIBCXX_3.4.26' not found

这种情况说明libstdc++.so.6连接的版本不够高。

a：查看libstdc++.so.6 运行库，注意大小写：strings /usr/lib64/libstdc++.so.6 |grep GLIBC 没有找到GLIBCXX_3.4.26

b: 直接：wget http://www.vuln.cn/wp-content/uploads/2019/08/libstdc.so_.6.0.26.zip

然后解压：unzip libstdc.so_.6.0.26.zip 

c: 先备份libstdc++.so.6 ：copy libstdc++.so.6 libstdc++.so.6.copy

d: 然后删除：rm libstdc++.so.6

e: 然后建立软连接: ln -s libstdc++.so.6.0.26 libstdc++.so.6

f: 再次查看：strings /usr/lib64/libstdc++.so.6 |grep GLIBC 

包含了GLIBCXX_3.4.26即可解决

(5) 报错：No module named 'face_alignment'，直接pip install face_alignment

5.成功：

Mesh2mesh:将一个mesh转换成需要的模型（标准模型）默认是bfm模型，需要*.obj,*.mtl以及uv贴图：扫描模型：216090个点，431544个面，通过nicp可转换成bfm模型：35709个点和70789个面

Mesh2pcl:将一个点云转换成bfm模型，需要点云上68个点的index信息

Demo_4d,通过一段视频流包含深度和彩色的图，生成bfm模型：

Denselandmark的3d点生成的bfm模型：     

Denselandmark生成的flam模型： 

 修改生成自己的模型：demo_nicp.py

device = torch.device('cuda:0')

pcls = io3d.load_ply_as_pointcloud('./test_data/base1.ply', device = device)

norm_pcls, norm_param = normalize_pcl(pcls)

pcl_lm_file = open('./test_data/base1_lmk.txt')

lm_list = []

for line in pcl_lm_file:

    line = int(line.strip())

    lm_list.append(line)

target_lm_index = torch.from_numpy(np.array(lm_list)).to(device)

lm_mask = (target_lm_index >= 0)

target_lm_index = target_lm_index.unsqueeze(0)

target_lm_index_m = target_lm_index[:, lm_mask]

#bfm_meshes, bfm_lm_index = load_bfm_model(torch.device('cuda:0'))

flame_meshes = io3d.load_obj_as_mesh('./test_data/base_flame.obj', device = device)

flame_meshes, flame_norm_param = normalize_mesh(flame_meshes)

dummy_render = render.create_dummy_render([1, 0, 0], device = device)

target_lm_index_flame, lm_mask_flame = get_mesh_landmark(flame_meshes, dummy_render)

#lm_mask_flame = torch.all(lm_mask, dim = 0)

flame_lm_index_m = target_lm_index_flame[:, lm_mask]

print(target_lm_index_m)

print(flame_meshes[flame_lm_index_m])

coarse_config = json.load(open('config/coarse_grain.json'))

registered_mesh = non_rigid_icp_mesh2pcl(flame_meshes, norm_pcls, flame_lm_index_m, target_lm_index_m, coarse_config)

io3d.save_meshes_as_objs(['test_data/final2.obj'], registered_mesh, save_textures = True)