search

charlesq34 / pointnet2

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

The part segmentation code under 'pointnet2/model' cannot reproduce your experiment result. #12

Open syb7573330 opened 6 years ago

syb7573330 commented 6 years ago

Hi, thanks for sharing your code. I am working on part segmentation task using Model40 dataset, but using the model defined under 'pointnet2_part_seg.py' and the training pipeline defined for pointnet part segmentation, I can only achieve less than 70% IoU. I guess something may go wrong, would you like to share your training code for pointnet++ part segmentation as well?

charlesq34 commented 6 years ago

Hi @syb7573330

sorry for the late response. I've added a new model which is the one we used to produce the results reported in the paper (the one with normal) at https://github.com/charlesq34/pointnet2/blob/master/models/pointnet2_part_seg_msg_one_hot.py

However, even for the previously uploaded model I can get: mIoU (for all shapes in all categories) 0.837830 without using normals.

I will try to organize and upload my training&evaluation code if you think it's helpful.

tejaskhot commented 6 years ago

Hi @charlesq34 Could you please upload your train-evaluation codes? Thanks!

charlesq34 commented 6 years ago

Hi @tejaskhot

the scannet part hasn't been organized yet (our major co-author did experiments for that part but he is working full-time towards SIGGRAPH ddl now). We will try to release it soon

suhangpro commented 6 years ago

Can you please comment on the difference between the two versions (the regular one and the one with "one_hot")? After Siggraph ddl is definitely fine :)

pxiangwu commented 6 years ago

Hi, for shape part segmentation experiment, I observe that the default split in your training code uses 'trainval'. And I wonder if you also used the split 'trainval' in your paper?

charlesq34 commented 6 years ago

Hi @astrolagrange

We didn't split trainval because the dataset is already pretty small (esp. for certain classes). We just select the model trained after say 200 epochs and evaluate that for the reported results.

Leerw commented 5 years ago

Can you please comment on the difference between the two versions (the regular one and the one with "one_hot")? After Siggraph ddl is definitely fine :)

@pxiangwu hi, have you got the answer?