aRI0U
Better accuracy with smaller block in S3DIS
Hi, I have implemented this repository with custom S3DIS data (xyz and normal only) I had tried with entire rooms, and blocks (5x5, 2x2, and 1x1)... The best accuracy I have got is with the smallest block... Is it possible that local encoder not work properly? It can related with the square distances? Thanks in advance.
I had got some improvements... I had mix some code from:
https://github.com/isl-org/Open3D-ML/tree/master/ml3d/torch/models
More improvements... I have tweaked dropout in the last mlp and the batch normalization parameters...
Thank you very much. i also find that the net cannot learn the features.....May be I should learn something from OPEN3D.
Hi @josegemez, I tried similar improvements with Open3D's code, even though the baseline results were better, but still I am only getting good scores for classes with larger point clouds only while inputting whole room as input. Except for few classes, for most of the other classes the model is learning nothing.
Could you please provide more specific details on the exact input samples during the training and your code improvements from the Open3D-ML's implementation. It would be really helpful, thanks in advance!
@josegemez Is it convenient to open source your code? I used the code of openml to merge and there is still no good result, thank you
可以跑通，但是根本学不到特征，我又对照其他各个版本改了下，还是很难收敛。
@jscksy 可以的，等晚上吧，我的代码也没有收敛，期待你可以把它收敛。
能加个好友吗，我改了一份代码是哪个自动驾驶版本的，也是遇到了问题，想发一份给你一起交流我的微信shendaln
可以啊，我后面跑出来的也挺奇怪的，我的邮箱 @.***
------------------ 原始邮件 ------------------ 发件人: "aRI0U/RandLA-Net-pytorch" @.>; 发送时间: 2023年2月8日(星期三) 晚上9:29 @.>; @.@.>; 主题: Re: [aRI0U/RandLA-Net-pytorch] Better accuracy with smaller block in S3DIS (Issue #42)
@jscksy 可以的，等晚上吧，我的代码也没有收敛，期待你可以把它收敛。
能加个好友吗，我改了一份代码是哪个自动驾驶版本的，也是遇到了问题，想发一份给你一起交流
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>
@jscksy 可以的，等晚上吧，我的代码也没有收敛，期待你可以把它收敛。
同学你好，请问可以发我一份代码吗，我最近也在复现这个，遇到很多bug，我的邮箱是[email protected]，万分感谢
@josegemez Dear josegemez, may you help me provide your modified code. Very bad results were gotten when using this code. Could you please help me, this is my email: [email protected] And thank you very much
