how to train on my own data

[wlsonia]

how to train on my own data,just have a txt file x,y,z

[burui11087]

Hi @wlsonia

Please refer to this file and convert you txt file to h5 file. xyz is enough and RGB information is not necessary.

Thanks.

[wlsonia]

thanks,I want to label my point cloud,can you recommend me an annotation software?thank you.

[tiger-bug]

Cloud Compare have some segmentation algorithms that may make annotating easier. What I've also done is projected my point cloud to 2D space and then used image segmentation techniques to make that sort of thing easier (especially if it's an aerial collect). Skimage has quite a few image segmentation routines that might make things easier for you.

[tiger-bug]

Also one thing I do is start with ground segmentation. that may make segmentation easier because you have the ground out of the way. @burui11087 would you be able to do a binary classification with PointCNN (ground vs non ground)? That might be a way to get used to deep learning with PointCNN. That's one thing I was considering trying with my data to get started and then going from there. CSF is a pretty straight forward way of segmenting the ground returns.

[sanghyunbai]

Hi, I have pts data prepared, I wish to understand exact part of code that I have to change. Please consult.

[SERPENT-H]

Hi I have txt files that contain x,y,z,r,g,b,i information,but when I run the prepare_sannet_data_cls.py file,an error occurred and I don't know how to solve it . **Traceback (most recent call last): File "prepare_scannet_cls_data.py", line 75, in main() File "prepare_scannet_cls_data.py", line 54, in main indices = np.random.choice(pt_num, sample_num, replace=(pt_num < sample_num)) File "mtrand.pyx", line 894, in numpy.random.mtrand.RandomState.choice ValueError: a must be greater than 0 unless no samples are taken ** I think it is points number < sample number,but points in my txt file are more than 3 million,so I have no idea what to do next