My own datasets #17
Comments
|
To use your own data, you need to (1) perform SfM using COLMAP, and (2) convert the COLMAP outputs to MVS format, and (3) use TransMVSNet to perform the final reconstruction. More instruction can be found here. |
|
Thank you for your reply. I've used colmap to get sparse point cloud information from my dataset and applied it to TransMVSNet, but it didn't work well. I wonder if I should make my own training set to retrain a model，so I would like to ask how TransMVSNet's training set is made.
谭宇璇
***@***.***
…------------------ 原始邮件 ------------------
发件人: "megvii-research/TransMVSNet" ***@***.***>;
发送时间: 2022年11月8日(星期二) 上午10:04
***@***.***>;
***@***.******@***.***>;
主题: Re: [megvii-research/TransMVSNet] My own datasets (Issue #17)
To use your own data, you need to (1) perform SfM using COLMAP, and (2) convert the COLMAP outputs to MVS format, and (3) use TransMVSNet to perform the final reconstruction. More instruction can be found here.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
|
Hi, |
|
Thank you very much for your reply. The model_ Bld.ckpt has been used according to your suggestions, but the final effect is still not ideal. Attached is the test results of my dataset. I hope I could receive your advice.
谭宇璇
***@***.***
…------------------ 原始邮件 ------------------
发件人: "megvii-research/TransMVSNet" ***@***.***>;
发送时间: 2022年11月8日(星期二) 中午11:18
***@***.***>;
***@***.******@***.***>;
主题: Re: [megvii-research/TransMVSNet] My own datasets (Issue #17)
Hi,
(1) Would you mind showing some depth results or pcd results of your own data? So that we can find where the problem is.
(2) To retrain the model, you need a relative large dataset with high-quality GT depth, which might be hard to get.
(3) We didn't generate training data, and we use the the existing training dataset DTU and BlendedMVS. If you use TransMVSNet in outdoor scene, you can use the bld-finstuned ckpt.
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
从QQ邮箱发来的超大附件
scan4.rar (195.53M, 2022年12月09日 09:34 到期)进入下载页面：http://mail.qq.com/cgi-bin/ftnExs_download?k=5f34656154efa4c36f6a84424062511f58000352520656094c060402544f5755555248590500551d580c07595453540658030455666b634302550b55481002426109&t=exs_ftn_download&code=a4eafbc0
|
|
After checking your results, I think the data format you used might have problems. For example, have you given the depth_min value and depth_max value in cam.txt with proper format? Which dataloader is used in your experiment? I recomend checking the cam.txt and the dataloader to verify if the depth range be read properly. You can send me an e-mail to add my wechat for more information. |
|
@DingYikang 您好，我用colmap稀疏重建后跑MVSNet，得到的深度范围一直是6-8，最后估计的深度图也不对，重建点云很少，效果很差，这个深度范围应该怎么修改呢？ |
hi，建议您检查所使用的dataloader是否正确，如果使用colmap2mvsnet.py进行转换，建议使用general_eval.py，深度图不对的原因可能是在读取cams.txt时d_min，d_max读取错误，您可以对照检查自己的数据和公开数据集格式的区别，并修改dataloader进行适配。 |
|
@DingYikang 你好，感谢你的回复，我在使用tnt数据的时候，用colmap2mvsnet转换出来跑，也是没有能够得到一个较好的结果，深度图很乱，生成的点云很小，cams.txt里面深度范围是3-40，但是我看transmvsnet提供的tnt数据txt深度值都是0.几，是不是这个原因导致了深度图输出错误了呢？如果是的话应该如何使用colmap去转换数据呢？感谢您的回复！ |
如果您重新用colmap2mvsnet跑了tnt的数据，在选择dataloader时应该将tnt_eval.py替换为general_eval.py，猜测是因为这个原因导致读取cam.txt中的内容出错了 |
|
@DingYikang 你好，我的txt文件如下，我修改了colmap2mvsnet.py保存深度最小值和深度最大值，下图左图是从repo中下载的处理后的cam.txt文件，右图是用colmap2mvsnet转换出来的文件，深度差别很大，最后导致深度图和分割图就不对，点云文件很小，这个是什么原因呢？应该如何修改呢？感谢您的回复！ |
Could you please tell me how to create my own datasets if I want to reconstruct my school?
The text was updated successfully, but these errors were encountered: