Mismatch between labels and pointcloud when visualizing a custom dataset GT

[Yinxq]

I was trying to visualize the ground truth of a sementic segmentation dataset named Rellis-3D and encounted some problems:

I modified the following files according to theirs Semantic KITTI version:

ml3d/datasets/rellis.py
ml3d/vis/labellut.py
ml3d/datasets/_resources/rellis.yaml
ml3d/datasets/utils/rellis.yaml
ml3d/datasets/__init__.py
ml3d/configs/default_cfgs/rellis.yml

I think these files was customized properly, but in the visualizer, the labels and pointcloud were mismatched, for example, as the Figure shows: only "grass" and "fence" was visible, but I got pointcloud of "grass" and "person";

however, when i test the Semantic KITTI dataset, it works well.

Any help is greatly appreciated, thanks!

[ssheorey]

Hi @Yinxq the mapping between class label and class names for SemanticKITTI is here: https://github.com/isl-org/Open3D-ML/blob/5baf7223e9126e26feffc6f08d4b9ae728367dd1/ml3d/datasets/semantickitti.py#L112 You can modify this as appropriate for Rellis-3D.

[Yinxq]

Hi @Yinxq the mapping between class label and class names for SemanticKITTI is here:

https://github.com/isl-org/Open3D-ML/blob/5baf7223e9126e26feffc6f08d4b9ae728367dd1/ml3d/datasets/semantickitti.py#L112

You can modify this as appropriate for Rellis-3D.

Hi, @ssheorey the mapping between class label and class names for Rellis-3D is correct, the dataset file rellis.py is shown in the link: https://github.com/Yinxq/Open3D-ML/blob/master/ml3d/datasets/rellis.py#L108 I think my problem is similar to these issues: #225 #243 #409 , but #422 doesn't fix this problem in my case.

[sanskar107]

@Yinxq this must be a problem from your side as all datasets supported by Open3D-ML works just fine. Can you upload the files you've changed so that we can test it ?

[obeidahmad]

@sanskar107 I have this same problem when visualizing non dataset data. Aka when I load data manually and use Visualizer().visualize() instead of Visualizer().visualize_dataset()

[obeidahmad]

Loading my data in a CustomDataset() Class fixed the problem. It's still weird however that it doesn't well support the visualization of custom data like in the documentation