前几天一篇曲面重建文章的审稿意见回来了,要求加近三年对比方法。在github上搜了一些项目,大部分的环境都很难配置成功。最后找了一个ICML2021年的点云重建项目[1]作为实验对比。
项目链接:mabaorui/NeuralPull-Pytorch
整体来说,该项目的配置比较容易,但是有几个比较蛋疼的地方需要注意:
1. 在安装pyhocon的时候报错
ERROR: Error [WinError 225] 无法成功完成操作,因为文件包含病毒或潜在的垃圾软件: while executing command python setup.py egg_info。
这里我是在pycharm里面使用setting内部的搜索实现安装的。
2. 解析npull.conf出错:
general {
base_exp_dir = ./outs/
recording = [
./,
./models
]
}
dataset {
data_dir = data/
np_data_name = carnew1w_norm.npz
}
train {
learning_rate = 0.001
maxiter = 40000
warm_up_end = 1000
eval_num_points = 100000
batch_size = 5000
save_freq = 5000
val_freq = 2500
report_freq = 1000
igr_weight = 0.1
mask_weight = 0.0
}
model {
sdf_network {
d_out = 1
d_in = 3
d_hidden = 256
n_layers = 8
skip_in = [4]
multires = 0
bias = 0.5
scale = 1.0
geometric_init = True
weight_norm = True
}
}
我查看了pyhocon包的说明,发现这个.conf的格式有一点问题。我按照说明将conf的格式改了一下,如下所示:
general: {
base_exp_dir : ./outs/
recording : [
./,
./models
]
}
dataset: {
data_dir : data/
np_data_name : carnew1w_norm.npz
}
train: {
learning_rate : 0.001
maxiter : 40000
warm_up_end : 1000
eval_num_points : 100000
batch_size : 5000
save_freq : 5000
val_freq : 2500
report_freq : 1000
igr_weight : 0.1
mask_weight : 0.0
}
model: {
sdf_network : {
d_out : 1 #
d_in : 3 #
d_hidden : 256 #
n_layers : 8
skip_in : [4]
multires : 0
bias : 0.5
scale : 1.0
geometric_init : True
weight_norm : True
}
}
这样程序就可以跑通了:
Reference
[1] B. Ma, Z. Han, et al. Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces:, 10.48550/arXiv.2011.13495[P]. 2020.