prediction convertion?

[bututoubaobei-0108]

Hello,

I try to run the code of testing, and use the model "mode_best.pth" in hugging face for s3dis segmentation: https://huggingface.co/Pointcept/PointTransformerV3/tree/main/s3dis-semseg-pt-v3m1-0-rpe/model
However, when I apply the predicted label on the coordinates of the point clouds, for example, on area5_hallway9, the obtained hallway form does not correspond to the one obtained when I apply the groud truth label on the same coordinates of the point clouds.

For example, in the point clouds with the predicted labels, the roof has holes, and the hallway has some obstacles, which do not appear in the point clouds with the ground truth labels.

However, when I do not apply the labels to the point clouds, and only visualize the point clouds, the form is different from both the point clouds with predicted labels and those with groud truth labels.

Do you know what the problem is?

groundtruth_2.mp4

prediction_2.mp4

without_label_2.mp4

[Gofinge]

Nice visualization, how about first check the number metric of testing is matched? The released weight is for PTv3 is with Pointcept v1.5.1. I remember there exist a small gap between current version.

[jtj01]

Hello, I would like to ask, how do you visualize it? I am using the same pre trained model as you. I am unable to visualize it, and I am not sure if the result I ran is incorrect or for some reason. I hope you can help me answer this, thank you very much!

[bututoubaobei-0108]

Nice visualization, how about first check the number metric of testing is matched? The released weight is for PTv3 is with Pointcept v1.5.1. I remember there exist a small gap between current version.

Thank you for the suggestion. I will check it out about the metric value. In addition, do you have the latest version of the model of ptv3 for dataset s3dis?

[bututoubaobei-0108]

Hello, I would like to ask, how do you visualize it? I am using the same pre trained model as you. I am unable to visualize it, and I am not sure if the result I ran is incorrect or for some reason. I hope you can help me answer this, thank you very much!

the output is the label of each point. You should create a file in which you define a color for each label. In this way, each point has its corresponding color, and you can visualize it. I visualize it based on open3d library

[jtj01]

Thank you very much for your reply. I ran it successfully! Thank you！

[Gofinge]

Nice visualization, how about first check the number metric of testing is matched? The released weight is for PTv3 is with Pointcept v1.5.1. I remember there exist a small gap between current version.

Thank you for the suggestion. I will check it out about the metric value. In addition, do you have the latest version of the model of ptv3 for dataset s3dis?

Hi, sorry for the late response, I just recovered from the busy schedule for the CVPR project. I will have a big update a few weeks later. A stronger self-supervised PTv3 and turned weight for S3DIS sem seg will be available in Jan of next year after necessary review process.