search

charlesq34 / pointnet2

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space
Other
3.06k stars 894 forks source link

The performance of semantic segmentation on S3DIS dataset #15

Open shuluoshu opened 6 years ago

shuluoshu commented 6 years ago

Hi, @charlesq34 , I trained the PointNet++ model by using the dataset S3DIS and the training code from pointnet , with the model "PointNet++", namely, the Pointnet2_sem_seg.py, however, I found no better performance than the original PointNet, so could you offer me the result (mean IoU and Accuracy) for comparison? thanks a lot!

charlesq34 commented 6 years ago

Hi @shuluoshu Have you tried to use XYZ information only? the RGB and X'Y'Z' features can be quite dominant in S3DIS dataset

shuluoshu commented 6 years ago

Hi, @charlesq34 , do you mean that by comparing the results of PointNet (XYZ instead of original XYZRGB) and PointNet++ only using the XYZ information? Does the PointNet++ model not support RGB information?

jefft255 commented 6 years ago

Hello,

After getting PointNet++ to work with S3DIS, I experienced the same thing. However, as Charles pointed out, PointNet uses XYZ RGB X'Y'Z' for segmentation, while, as of right now, PointNet++ semantic segmentation model (pointnet2_sem_seg.py) only uses XYZ. The fact that PointNet++ performs in the same ballpark as PointNet using so much less information is quite impressive in my opinion.

Having read the paper carefully, I do not see a fundamental difficulty in using other features in this model. It would be quite a bit of work to implement however.