Issue with saving and visualizing results.

[max-monastirsky]

Hi,

I have managed to run the first stage and save the quarry proposals.

Now I'm running the second stage with:
./tools/dist_test.sh ./projects/configs/voxformer/voxformer-S.py ./ckpts/voxformer-S/miou12.35_iou44.02_epoch_14.pth 1
And I'm having problems with saving and visualizing the results.

I had to uncomment the following line:
https://github.com/NVlabs/VoxFormer/blob/9d28f0ed857291f2b707e66757118c466e1f158d/tools/test.py#L242C21-L242C21

And I had to remove the ['bbox_results'] from mmcv.dump(outputs['bbox_results'], args.out) because it raised a KeyError.
Now I have managed to save the .pkl but have no clue on what should be next step.
How do I visualize the results? you said you used semantic-kitti-ap but it is not compatible with the .pkl file.
I also tried to add --show-dir and --show flags to the test script, but with no effect.

I would really appreciate help in visualizing the results.
Thank you.

[FeilongZzz]

Hi there. I am also working on this.

In my case, I saved the prediction result into pkl file. Then I checked semantic-kitti-api and if you can see in the dataset, .label file is the only additional file for visualize voxels. I then loaded the .pkl result and modified the array's shape and dtype(since the required semantic data is in uint16, but the output is uint64). The shape of each label should be 25625632 and dtype=uint16. After that you can save the labels as 000000.label ----- xxxxxx.label and put them in the prediction folder in the dataset, just like what api asked you to do.

I then run the script in Voxel Grids for Semantic Scene Completion and it worked for me. I simply ignored .occluded and .invalid files because they are not given by this method. Nothing went wrong in visulization.

I strongly advise you to use semantic-kitti-api, but if you have the time, I think the result can also be visualized by your own script using open3d or something else.. I tried to output the voxels as they are points in open3d, the result is not so good-looking, but you can say it's street in the 3d model. Using that api is really helpful.

Thank you.

[SZUshenyan]

Hi there. I am also working on this.

In my case, I saved the prediction result into pkl file. Then I checked semantic-kitti-api and if you can see in the dataset, .label file is the only additional file for visualize voxels. I then loaded the .pkl result and modified the array's shape and dtype(since the required semantic data is in uint16, but the output is uint64). The shape of each label should be 256_256_32 and dtype=uint16. After that you can save the labels as 000000.label ----- xxxxxx.label and put them in the prediction folder in the dataset, just like what api asked you to do.

I then run the script in Voxel Grids for Semantic Scene Completion and it worked for me. I simply ignored .occluded and .invalid files because they are not given by this method. Nothing went wrong in visulization.

I strongly advise you to use semantic-kitti-api, but if you have the time, I think the result can also be visualized by your own script using open3d or something else.. I tried to output the voxels as they are points in open3d, the result is not so good-looking, but you can say it's street in the 3d model. Using that api is really helpful.

Thank you.

Hello, I would like to ask how you load the pkl result and convert it into the data form required by semantic-kitti-api. I run the "./tools/my_dist_test. sh ./projects/configs/voxformer/voxformer-T.py /home/ZHX/Data/sy/VoxFormer/ckpts/VoxFormer-T/miou13.35_iou44.15_epoch_12.pth 1", I got a file called part_0.pkl which size is 20 g . If you can provide the conversion code, I will be appreciated, looking forward to your reply!

[Sxiansheng24]

Hello, thank you for providing the method. I'm facing the same issue and would appreciate it if you could provide the relevant code. I'm looking forward to your response.