Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Issue with Saving and Visualizing Results as .pcd Files in PoinTr #164

Open
HANYUEZHAN opened this issue Dec 6, 2024 · 23 comments
Open

Issue with Saving and Visualizing Results as .pcd Files in PoinTr #164

HANYUEZHAN opened this issue Dec 6, 2024 · 23 comments

Comments

@HANYUEZHAN
Copy link

Dear Authors,
I hope this message finds you well. I have run the following command:

bash
python tools/inference.py
cfgs/PCN_models/AdaPoinTr.yaml ckpts/AdaPoinTr_PCN.pth
--pc_root demo/ \
--save_vis_img
--out_pc_root inference_result/
While the results are displayed in the terminal, I am unable to locate the saved file. I would like to save the result as a .pcd file and visualize the data. Could you please let me know how this issue is handled in the project? I attempted to modify main.py and parser.py, but encountered errors.

Thank you for your time and assistance.
@MaSiTou
Copy link

MaSiTou commented Jan 5, 2025

我能看见保存的文件啊，是.npy 文件和.jpg文件

@HANYUEZHAN
Copy link
Author

HANYUEZHAN commented Jan 7, 2025 via email

@MaSiTou
Copy link

MaSiTou commented Jan 7, 2025 via email

@HANYUEZHAN
Copy link
Author

HANYUEZHAN commented Jan 7, 2025 via email

@MaSiTou
Copy link

MaSiTou commented Jan 7, 2025 via email

@HANYUEZHAN
Copy link
Author

HANYUEZHAN commented Jan 8, 2025 via email

@MaSiTou
Copy link

MaSiTou commented Jan 8, 2025 via email

@HANYUEZHAN
Copy link
Author

HANYUEZHAN commented Jan 8, 2025 via email

@shf0127
Copy link

shf0127 commented Jan 14, 2025

你好，我没有处理shapenet的数据集，我使用的是我自己的数据集进行补全，所以这个问题我也回答不了要不你再去看看issue里面 在 2025-01-08 10:47:45，"MaSiTou" @.> 写道： 官方使用的shapenet55数据集已经完成了数据预处理，每个物体的点云都是大小统一，都经过了归一化，请问我还可以怎么预处理呢？您是如何操作的呢？
…
________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月8日 10:26 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 我用的是另一个模型进行推理，但是效果也很差；不过我建议你可以对数据进行一下预处理 在 2025-01-07 16:44:39，"MaSiTou" @.> 写道： 请问您试了官方的pointr_training_from_scratch_c55_best.pth权重了吗？我是用官方提供的这个权重进行inference,执行命令：python tools/inference.py cfgs/ShapeNet55_models/PoinTr.yaml pointr_training_from_scratch_c55_best.pth --pc_root demo/ --save_vis_img 发现无论推理官方的demo还是我给他随机一张点云，效果都挺差的。我把我的推理结果发给您您能帮我分析一下吗？
________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月7日 15:01 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 对，我现在是在研究点云补全，我放了自己的数据后问题确实很大，输出了一堆很奇怪的点 在 2025-01-07 11:10:12，"MaSiTou" @.> 写道： 我看作者在PoinTr中对数据的处理是对于训练数据进行归一化，然后将归一化后的数据作为了ground_truth，然后再将ground_truth进行crop 之后生成incomplete point clouds。作者在issues里面提到了使用自己的数据的话问题还是挺大的，我没记清楚是哪一个issues。另外，我想问一下，您也是在搞点云补全吗？我有一些关于这个PoinTr的inference的问题想和您讨论一下
________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月7日 11:04 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 哦哦谢谢您，我后面用PCN数据集进行测试的时候就找到了，那我还想请问一下，如果要用自己的点云数据进行补全和重建的话是不是要先对数据进行去噪、归一化和下采样 在 2025-01-05 10:01:09，"MaSiTou" @.> 写道： 我能看见保存的文件啊，是.npy 文件和.jpg文件 D Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> D Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6OKPESCCDTKNDHLEPT2JM723AVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZUGMYDEMRSGI. You are receiving this because you commented. D Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> D Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6OUHKE4VCEUPVWYX2L2JN3TZAVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZUGUZDCMRUHA. You are receiving this because you commented.Message ID: @.> ― Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> ― Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6IDQFAHS7CAJSH2Z532JSEGVAVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZWGU3TQOJXGA. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.>

你好请问你使用自己的数据集效果怎么样阿？最近也想用自己制作数据集进行，可以交流一下么

@HANYUEZHAN
Copy link
Author

HANYUEZHAN commented Jan 14, 2025 via email

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

@yuxumin
Copy link
Owner

yuxumin commented Jan 14, 2025

感谢讨论呀，PoinTr和AdaPoinTr这类的补全模型其实很多时候还是类似训练数据分布的其他数据，最近的一些研究我发现这种基于Chamfer Dis优化的模型在优化以后，会对点云的分布比较敏感（特别是用了DGCNN这种对局部区域很敏感的encoder的时候），所以一旦输入的点云的distribution出现了比较大的gap的时候，效果就容易比较差了。

上面我看到有一些reply也提到了normalization和预处理，这其实是一种让训练数据和inference数据的data distribution靠近的办法，但是更加本质的问题还是在方法层面的。真实场景或者是跨点云分布类型的例子确实表现的不是很好

@yuxumin
Copy link
Owner

yuxumin commented Jan 14, 2025

如果是自己的数据集，最好是能通过supervise的方式做一些finetune吧，这样效果会好一些

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

@yuxumin
Copy link
Owner

yuxumin commented Jan 14, 2025

如果是在PCN ShapeNet内的数据肯定是没问题把，这类CAD来源的其实表现还是很鲁棒的。
跨点云分布是想表示训练数据和测试数据在data distribution上有gap。比如说一个是来自于CAD mesh的抽样，一个是来自于RGB-D相机或者是激光雷达扫描。

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

@HANYUEZHAN
Copy link
Author

如果是在PCN ShapeNet内的数据肯定是没问题把，这类CAD来源的其实表现还是很鲁棒的。 跨点云分布是想表示训练数据和测试数据在data distribution上有gap。比如说一个是来自于CAD mesh的抽样，一个是来自于RGB-D相机或者是激光雷达扫描。

是的是的，模型在PCN上的补全效果很好，但是我用自己的数据集，效果就有点不太好

@HANYUEZHAN
Copy link
Author

就如果使用数据集中的数据进行预测效果也不行么？
…
---Original--- From: @.> Date: Tue, Jan 14, 2025 19:35 PM To: @.>; Cc: @.@.>; Subject: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as.pcd Files in PoinTr (Issue #164) 效果不好嘞 在 2025-01-14 17:17:27，"shf0127" @.> 写道： 你好，我没有处理shapenet的数据集，我使用的是我自己的数据集进行补全，所以这个问题我也回答不了要不你再去看看issue里面 在 2025-01-08 10:47:45，"MaSiTou" @.> 写道： 官方使用的shapenet55数据集已经完成了数据预处理，每个物体的点云都是大小统一，都经过了归一化，请问我还可以怎么预处理呢？您是如何操作的呢？ … ________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月8日 10:26 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 我用的是另一个模型进行推理，但是效果也很差；不过我建议你可以对数据进行一下预处理 在 2025-01-07 16:44:39，"MaSiTou" @.> 写道： 请问您试了官方的pointr_training_from_scratch_c55_best.pth权重了吗？我是用官方提供的这个权重进行inference,执行命令：python tools/inference.py cfgs/ShapeNet55_models/PoinTr.yaml pointr_training_from_scratch_c55_best.pth --pc_root demo/ --save_vis_img 发现无论推理官方的demo还是我给他随机一张点云，效果都挺差的。我把我的推理结果发给您您能帮我分析一下吗？ ________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月7日 15:01 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 对，我现在是在研究点云补全，我放了自己的数据后问题确实很大，输出了一堆很奇怪的点 在 2025-01-07 11:10:12，"MaSiTou" @.> 写道： 我看作者在PoinTr中对数据的处理是对于训练数据进行归一化，然后将归一化后的数据作为了ground_truth，然后再将ground_truth进行crop 之后生成incomplete point clouds。作者在issues里面提到了使用自己的数据的话问题还是挺大的，我没记清楚是哪一个issues。另外，我想问一下，您也是在搞点云补全吗？我有一些关于这个PoinTr的inference的问题想和您讨论一下 ________________________________ 发件人: HANYUEZHAN @.> 发送时间: 2025年1月7日 11:04 收件人: yuxumin/PoinTr @.> 抄送: MaSiTou @.>; Comment @.> 主题: Re: [yuxumin/PoinTr] Issue with Saving and Visualizing Results as .pcd Files in PoinTr (Issue #164) 哦哦谢谢您，我后面用PCN数据集进行测试的时候就找到了，那我还想请问一下，如果要用自己的点云数据进行补全和重建的话是不是要先对数据进行去噪、归一化和下采样 在 2025-01-05 10:01:09，"MaSiTou" @.> 写道： 我能看见保存的文件啊，是.npy 文件和.jpg文件 D Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> D Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6OKPESCCDTKNDHLEPT2JM723AVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZUGMYDEMRSGI. You are receiving this because you commented. D Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> D Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6OUHKE4VCEUPVWYX2L2JN3TZAVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZUGUZDCMRUHA. You are receiving this because you commented.Message ID: @.> ― Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> ― Reply to this email directly, view it on GitHub<#164 (comment)>, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ASVUR6IDQFAHS7CAJSH2Z532JSEGVAVCNFSM6AAAAABTEZ7DQCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZWGU3TQOJXGA. You are receiving this because you commented.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> 你好请问你使用自己的数据集效果怎么样阿？最近也想用自己制作数据集进行，可以交流一下么 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.> — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

在PCN上效果很不错的

@yuxumin
Copy link
Owner

yuxumin commented Jan 14, 2025

啊，我就是这么想的啊。数据集中完整点云用SW转化的，部分点云用深度相机拍摄的。然后预测的时候用数据集中的部分点云。那这是不太行么？

感觉可以tuning一下，毕竟transfer的效果确实没有在设计model的时候考虑进去

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

@yuxumin
Copy link
Owner

yuxumin commented Jan 14, 2025

就是你可以在用到你的场景上之前，稍微做一下finetune，可以用预训练模型的权重作为weight去init

@shf0127
Copy link

shf0127 commented Jan 14, 2025 via email

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
4 participants
@yuxumin @MaSiTou @shf0127 @HANYUEZHAN