search

meidachen / STPLS3D

🔥 Synthetic and real-world 2d/3d dataset for semantic and instance segmentation (BMVC 2022 Oral)
236 stars 20 forks source link

How to handle the seperated semantic instance segmentation results at the edge of different blocks? #5

Open JTGiser opened 2 years ago

JTGiser commented 2 years ago

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!

2022-06-13_15-46

meidachen commented 2 years ago

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 commented 2 years ago

Thanks for your prompt response, I will try it.

Wangcongyu620 commented 2 years ago

@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 commented 2 years ago

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 commented 10 months ago

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!

2022-06-13_15-46

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