Rectangular box problem about S3DIS dataset visualization?

[MichaleWong]

❓ Questions & Help

it's the area6 of s3dis dataset, why it lools like a rectangular box? I check the range of x and y , it is between-0.5 and 0.5, but z is from 0.0 to 4.273, does any one know what's wrong with t that, by the way, I used open3d to visualize it

def writeO3D(dataset, file_path):
    points = dataset.data.pos.cpu().numpy()
    colors = dataset.data.x[:,:3].cpu().numpy()
    ply = o3d.geometry.PointCloud()
    ply.points = o3d.utility.Vector3dVector(points)
    ply.colors = o3d.utility.Vector3dVector(colors)
    o3d.io.write_point_cloud(file_path, ply)


test_dataset  = S3DIS(path,test_area=6, train=False)
gactest.writeO3D(test_dataset,"/home/michael/Data/S3DIS/area6.ply")

[tchaton]

Hey @MichaleWong,

Pytorch Geometric s3dis is the 1x1 square cylinder S3DIS dataset. It was prepared for PointNet2 originally. In TorchPoints3d, we re-implemented S3dis from raw data and support 3 cutting versions. 1x1 square cylinder, r radius sphere, r disk cylinder.

r disk cylinder seems to give best results.

You will find 3 versions there: https://github.com/nicolas-chaulet/torch-points3d

Best, Thomas Chaton

[aliasboink]

Hey @tchaton

Sorry if asking this here is improper, but this may help people with visualization as well later on and it seems proper to ask here.

Loading the data with S3DIS1x1 compared to the OriginalFused results in a different number of points. I am using Area 5 only. With Fused I get 78649818 points (the number of points presented in the study, so I wager this is just the raw truthful data) and with S3DIS1x1 I get about 4096*6852 points (which is pre-processed, I know, but I don't know how it is pre-processed!!!), which is about 50 million less. I want to visualize the predicted semantic segmentation labels of my model, but I can't make a one to one correspondence due to the Fused version having so many more points.

In this case, how do I go about visualizing the prediction from my model on the whole area that has been trained using the PyTorch Geometric S3DIS dataset? (ideally I want to load the fused area 5 fully and see the predictions on that, i.e., reduce it to the 28 million dimension and have the positions be spatially accurate, which is what I was trying to do by using the Fused "loader")

Best regards, Adrian