tangjiapeng / SA-ConvONet

ICCV2021 Oral SA-ConvONet: Sign-Agnostic Optimization of Convolutional Occupancy Networks
MIT License
89 stars 5 forks source link

on my own data #13

Open zhao-you-fei opened 1 year ago

zhao-you-fei commented 1 year ago

您好,请问我在自己的数据集(室内场景)上没有取得很好地效果,可以修改那些参数来增强效果。

tangjiapeng commented 1 year ago

没有test time optimization 也结果很差吗?

zhao-you-fei commented 1 year ago

您好,我有些关于这个代码的问题想咨询学习一下,内容详情请见附件

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月11日(星期二) 下午5:06 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #13)

没有test time optimization 也结果很差吗?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

zhao-you-fei commented 1 year ago

您好,我有些关于这个代码的问题想咨询学习一下,内容详情请见附件。请您抽空解答下,拜托了。

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月11日(星期二) 下午5:06 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #13)

没有test time optimization 也结果很差吗?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

tangjiapeng commented 1 year ago

我这边看不到附件 不知道你的问题

zhao-you-fei commented 1 year ago

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月13日(星期四) 晚上10:41 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #13)

我这边看不到附件 不知道你的问题

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

zhao-you-fei commented 1 year ago

非常感谢您抽时间看我留下的问题 

问题一:我用代码提供的预训练好的模型迁移到我用的DTU数据集生成的点云上,没有训练新的模型,我想请教您如果想让模型看起来更好除了重新训练模型之外可以修改,config.yaml(以这个为例)中的哪些参数来改进,比如采样点的数量或者是大小一类

问题二:请问应该怎么评估自己的模型,这个带有法向信息的真值曲面如何获取

我看到您提供的四个数据集里边的内容是不相同的,然后分别用代码进行了预处理,这一步是在采样真值吗?

这个真值是.off形式的mesh还是.ply形式的点云。

如果是采样点云是否有代码获取这个采样点,因为提供的数据集好像不是都有真值

是否可用colmap获取的数据当做真值。

处理后的共通之处就是有个point.ply和pointcloud.npz这种文件。eval输入的数据好像只有points_iou.npz和pointcloud.npz,所以我不理解这个是怎么评价的。是评价的mesh随机采样点(真值)和输入点云数据之间的距离或者是差别嘛?

问题三:这个输出格式是体素还是面片

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月13日(星期四) 晚上10:41 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #13)

我这边看不到附件 不知道你的问题

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

zhao-you-fei commented 1 year ago

抱歉,我不是作者,我也有一堆问题想要请教,预处理我只做成功了M3D数据集其他shapenet、scannet及室内的那个都没搞成

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月21日(星期五) 上午10:23 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #13)

作者可以留个邮箱交流学习一下吗,关于这个网络对数据进行的预处理我有一些比较简单的问题想请教一下

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>