-
Notifications
You must be signed in to change notification settings - Fork 366
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.
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
What is the result of PointCNN in semantic3d dataset? #13
Comments
We are working on this dataset. The code for processing semantic3d at the moment probably is not a good solution. It just serves as an exemplar to demonstrate that semantic3d could be processed in a very similar way as that for ScanNet and S3DIS, though it won't lead to the best performance. Compared with ScanNet and S3DIS, semantic3d is extremely dense (and yes, it takes longer to train/eval), and we are working on a way to better leverage all the training data, which is likely doable through a proper preprocessing, without touch the PointCNN core code. So, please stay tuned ;-) |
How long does it need to train the cls/seg network? What is your device? |
I find that it may be overfitting using 1024 epoch in S3DIS. Do you have such problem? |
Hi, @LZDSJTU |
any update on semantic segmentation? I'm happy to chase preprocessing ideas. Are people seeing improvement over pointnet? I have that up and running and am decently happy with results. I'm debating trying pointnet++ or pointcnn. |
Semantic3D is large in terms of point number, but rather small in terms of diverse scene. Due to this reason, we observed strong overfitting on Semantic3d. Nevertheless, we managed to achieve results that would reach 2nd place on the leader board. We believe the results can be further improved if we try some more data augmentation tricks, so we decided not to release our results on the leader board. Then we joined Alibaba and do not have enough time to play on this dataset (we have larger datasets to play with ;-). In short, we believe PointCNN should perform well on Semantic3D, if you have time/patience in fighting against the overfitting. |
got it. Thanks for the thoughts.
…On Wed, Sep 26, 2018 at 6:29 PM Yangyan Li ***@***.***> wrote:
Semantic3D is large in terms of point number, but rather small in terms of
diverse scene. Due to this reason, we observed strong overfitting on
Semantic3d. Nevertheless, we managed to achieve results that would reach
2nd place on the leader board. We believe the results can be further
improved if we try some more data augmentation tricks, so we decided not to
release our results on the leader board. Then we joined Alibaba and do not
have enough time to play on this dataset (we have larger datasets to play
with ;-).
In short, we believe PointCNN should perform well on Semantic3D, if you
have time/patience in fighting against the overfitting.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#13 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ABJwrMRcL8LqAQfEtNW4Xni6Swakukvdks5ufCn1gaJpZM4ScKwK>
.
--
Ben Weinstein, Ph.D.
Postdoctoral Fellow
University of Florida
http://benweinstein.weebly.com/
|
希望了解PointCNN在semantic3D数据集上的结果?以及训练这个数据集需要的时间?谢谢!
The text was updated successfully, but these errors were encountered: