search

loicland / superpoint_graph

Large-scale Point Cloud Semantic Segmentation with Superpoint Graphs
MIT License
745 stars 214 forks source link

S3DIS superpoints generated by SPG does not look right #268

Closed anhtuanhsgs closed 2 years ago

anhtuanhsgs commented 2 years ago

Hi,

I followed the instruction from S3DIS.md to generate the superpoints for the S3DIS dataset using the learning method. I ran the model using the pretrained checkpoints and here is the visualization of the superpoints when running the visualization code:

Area_6, hallway_1 image

Area_5, hallway_12 image

Area_1, conferenceRoom_1 image

It looks like shapes of the superpoints are not very complex. I expect the superpoitns of the ceiling, walls, and floors should be larger. The results here also look very different from the paper (figure 3 - https://arxiv.org/pdf/1711.09869.pdf). Here is the evaluation of the superpoints.

image

We also retrain the SPG model but get the same partitioning result. What could be the error we made here? Or this is actually the right oversegmentation of S3DIS generated by SPG?

Thank you!

anhtuanhsgs commented 2 years ago

I verified that the over-segmentation I generated above is the correct one. The generated segmentations usually have squared patterns as in the image while the quality is still high.