Could you provide a dataloader.py for the Toronto3D dataset?

[whuhxb]

Hi @WeikaiTan

Could you provide a dataloader.py for the Toronto3D dataset? I want to see the specific data format to suit for the semantic segmentation baseline. I'm testing my algorithm on Toronto3D dataset, but can't load the data successfully. Or, could you please provide the code of RandLA-Net on Toronto3D as you shown in the table?

In addition, have you ever run the Open3D-ML framework for Toronto3D dataset? I have run RandLANet on Toronto3D dataset, but I only get half the mIoU with TF version as reported. The pytorch version always show can't pickle bugs due to the torch.multiprocessing module.

Thanks a lot in advance.

Xiaobing Han

[WeikaiTan]

I've uploaded my code for RandLA-Net at https://github.com/WeikaiTan/RandLA-Net.git
You may try it out.

[whuhxb]

@WeikaiTan
Hi, I have downloaded the code and run it now. But I find that the mIoU of RandLA-Net in training stage oscillates largely, for example from 40 to 4 and then to 50. It's so strange. Have you ever met this on Toronto3D dataset when running with RandLA-Net? Or, is this related with the grid size??

[WeikaiTan]

Yes, I experienced the same thing. First, the training and validation does not use the full point cloud, just random selection. Second, the initial learning rate is pretty large, so there will be some fluctuations in the early epochs. Third, the loss function is the cross entropy loss, so the validation accuracy is more stable compared to mIoU. It'll smooth out with more epochs trained.

[whuhxb]

Yes, I experienced the same thing. First, the training and validation does not use the full point cloud, just random selection. Second, the initial learning rate is pretty large, so there will be some fluctuations in the early epochs. Third, the loss function is the cross entropy loss, so the validation accuracy is more stable compared to mIoU. It'll smooth out with more epochs trained.

OK. Thank you for your kind reply.

[whuhxb]

@WeikaiTan
I have run the RandLA-Net code on Toronto3D dataset, but there is no mIoU results with the test phase. Have you ever met this? How to obtain the test result?

[WeikaiTan]

There is the test evaluation parameter in the main function

[whuhxb]

There is the test evaluation parameter in the main function

OK. Thanks. Have tried, and it works.

[whuhxb]

There is the test evaluation parameter in the main function

OK. Thanks. Have tried, and it works.

I just run RandLA-Net on Toronto3D with the code you provided, the mIoU is 72.9, much lower than the result you reported in the table mIoU 77. How to set the specific parameter to obtain the mIoU reported in the table? In helper_tool.py, use_intensity should be set True for without RGB training? For with RGB training, use_rgb and use_intensity should all be set True, right?

[WeikaiTan]

That result was reported by Hu, author of RanLANet, in his paper. I didn't get that good result either, but I didn't spend much time trying out different parameter settings. You may change the network settings in the helper_tool.py

[whuhxb]

That result was reported by Hu, author of RanLANet, in his paper. I didn't get that good result either, but I didn't spend much time trying out different parameter settings. You may change the network settings in the helper_tool.py

I have just check Hu's RandLA-Net paper, and haven't found the result of Toronto3D dataset, only with the introduction review of various 3D datasets but without results.

[whuhxb]

@WeikaiTan
Have you ever run Toronto3D dataset with SuperPointGraph code? Could you please share the dataloader code? Thanks a lot.

[WeikaiTan]

Hu's results can be found at https://doi.org/10.1109/TPAMI.2021.3083288. I'll test more on the configurations to see if I could get a similar result.

[whuhxb]

Hu's results can be found at https://doi.org/10.1109/TPAMI.2021.3083288. I'll test more on the configurations to see if I could get a similar result.

OK. I'm trying now, and waiting for your testing results. Thanks.