The dataloader of SensatUrban with point-transformer?

[whuhxb]

Hi @meidachen

How to set the dataloader of SensatUrban dataset with point-transformer? As SensatUrban and STPLS3D datasets are both large-scale, could I refer to and follow the generate_blocks.py and stpls.py upon SensatUrban dataset with point-transformer code?

Thanks.

[RockyatASU]

Thanks for your interest in our datasets. Yes, you can follow the generate_blocks.py and stpls.py to create a new dataset class for the SensateUrban dataset. As the point cloud resolution of SensatUrban and STPLS3D datasets are different, you may need to change the default block size in generate_blocks.py.

[whuhxb]

@RockyatASU

OK Thanks a lot

[whuhxb]

Hi @RockyatASU

How to check or measure the resolution of each ply area in SensatUrban dataset? And how to set the voxel size for SensatUrban dataset? In addition, if I do not want to generate blocks for SensatUrban dataset, is it possible to follow the RandLA-Net dataloader setting for Point Transformer?

Thanks.

[RockyatASU]

Hi @whuhxb,

I think our STPLS3D has a similar point cloud resolution to the SensatUrban dataset so we recommend you try 0.3~0.5 voxel size for your experiments on SensatUrban. For details please refer to the author of SensatUrban.

You can refer to this link for the detailed parameters of point-transformer. Basically, point-transformer utilizes the voxelized point cloud to reduce the point cloud density during training but it still produces a semantic label for each point. Sure, you can feed one whole scene of SensatUrban into a point-transformer without generating subblocks and the setting of RandLA-Net may be a good starting point for your experiments.

[whuhxb]

Hi @RockyatASU，

Thank you for your kind and detailed reply. I will have a try.

In addition, for S3DIS dataset with area 5 as testing, Point Transformer uses 12346 areas as training and 5 area as evaluate to obtain the best model and then uses 5 as testing, where train val and test set should have no intersection. But I have seen several github codes on S3DIS dataset adopts this training setting. I have no idea why set like this.

I wonder to know whether Point Transformer for STPLS3D also adopts the same setting. WMSC is for evaluate and testing and other files are used for training. Or, how to split the train set, val set, and test set for STPLS3D dataset?

Train set, val set, and test set of SensatUrban dataset have no intersection.

Thanks.