How to handle the seperated semantic instance segmentation results at the edge of different blocks?

[JTGiser]

Thanks for providing such a useful dataset! This is my instance segmentation predict result. I found the predict result is ok when the featue instance is inside every block, but at the edge of different blocks, the instance are predicted as seperated ones (for example the building). So how to handle it? Or can I input the whole point cloud file at once instead of dividing sepetated blocks?

thanks!

[meidachen]

Thanks for your interest in our datasets. But sorry I don't have a good answer for the label merging between different blocks, since I didn't try to do it myself. For merging the results, you may want to check this repo:
https://github.com/LiuShihHung/GICN/blob/29cd2975e539cf09e270159934d4ca3a7da01f0a/main_eval.py#L65
It seems like there is a block merging process there.

[JTGiser]

Thanks for your prompt response, I will try it.

[Wangcongyu620]

@JTGiser Have you solved the problem of merging labels？If you solve it, what method do you use? Look forward to your reply Thank you

[meidachen]

I'm only trying to provide information I got from other STPLS3D users here (we haven't had a chance to test it but hopefully it is helpful).

Special thanks to @ywyue, for providing the following information:
"We follow the block merging algorithm of SGPN basically, the code can be found here: https://github.com/laughtervv/SGPN/blob/master/utils/test_utils.py#L165 ."

@JTGiser , please provide more information on block merging if you solved it and would like to share, thanks.

[Greatest-Marx]

Thanks for providing such a useful dataset! This is my instance segmentation predict result. I found the predict result is ok when the featue instance is inside every block, but at the edge of different blocks, the instance are predicted as seperated ones (for example the building). So how to handle it? Or can I input the whole point cloud file at once instead of dividing sepetated blocks?

thanks!

Hi！Which model are you using, and is this a visualization of what you have tested and validated?