thangvubk / SoftGroup

[CVPR 2022 Oral] SoftGroup for Instance Segmentation on 3D Point Clouds
MIT License
339 stars 80 forks source link

Transform_train for custom dataset #187

Open PrarthanaShah opened 9 months ago

PrarthanaShah commented 9 months ago

I am trying to train a custom dataset. For that I am using custo,.py script and hence the 'type' in config.yaml is set to 'custom'. I would like to know that under what circumstances the crop method generates 0 valid_idxs? Also, what is the use of 'xyz = xyz_middle * self.voxel_cfg.scale' in transform_train method?

thangvubk commented 8 months ago

The custom dataset is the base dataset only. For specific custom dataset, you may need to write additional dataset class. xyz = xyz_middle * self.voxel_cfg.scale is to transform point scale to voxel scale before doing some transformations.