search

filaPro / oneformer3d

[CVPR2024] OneFormer3D: One Transformer for Unified Point Cloud Segmentation
Other
348 stars 32 forks source link

S3DIS from the scrathch #60

Closed ImaneTopo closed 4 months ago

ImaneTopo commented 5 months ago

Hello, I want to train the model on the s3dis dataset from the scrath, is ti possible and how?

oneformer3d-contributor commented 5 months ago

Does #54 help?

ImaneTopo commented 5 months ago

Does #54 help?

I cannot work with the pretrained on the scannet, because I want to train the model on the s3dis from the scratch in order to work with it in another task because my original dataset format is similar to the s3dis ( X Y Z RGB).

oneformer3d-contributor commented 5 months ago

I think, all of our datasets are (x, y, z, rgb). So why not to start from pre-trained checkpoint? If you don't want to do it, just delete the load_from parameter from config file.

ImaneTopo commented 5 months ago

I think, all of our datasets are (x, y, z, rgb). So why not to start from pre-trained checkpoint? If you don't want to do it, just delete the load_from parameter from config file.

My task for the segmentation is related to agriculture (specifically fruits count) so I'm not working on indoor, for that reason I choose to train the model from the scratch with my own data prepared that it's format is simalar to S3DIS.

ImaneTopo commented 5 months ago

I think, all of our datasets are (x, y, z, rgb). So why not to start from pre-trained checkpoint? If you don't want to do it, just delete the load_from parameter from config file.

My task for the segmentation is related to agriculture (specifically fruits count) so I'm not working on indoor, for that reason I choose to train the model from the scratch with my own data prepared that it's format is simalar to S3DIS.

oneformer3d-contributor commented 5 months ago

just delete the load_from parameter from config file

Does this answers your question?

ImaneTopo commented 5 months ago

just delete the load_from parameter from config file

Does this answers your question?

Thank you for your response. So this change will not influence the precisions?

oneformer3d-contributor commented 5 months ago

Hard to say anything about your dataset, as we only tried for indoor environments. On S3DIS this pre-training brings quite some improvement in accuracy, because the dataset is very small.

Lizhinwafu commented 5 months ago

I think, all of our datasets are (x, y, z, rgb). So why not to start from pre-trained checkpoint? If you don't want to do it, just delete the load_from parameter from config file.

My task for the segmentation is related to agriculture (specifically fruits count) so I'm not working on indoor, for that reason I choose to train the model from the scratch with my own data prepared that it's format is simalar to S3DIS.

  • my task is for instance segmentation of fruits on trees.

We are similar dataset