Training on ScanNet

[sophiajw]

Hi all,

has anybody tried to apply the network on ScanNet data? I am currently trying to achieve similar performance as mentioned in PointNet++ paper, which is around 83% accuracy. However, when training with this model, I can hardly achieve values greater than 60%.

I tried using the same hyperparameters as in the original PointNet++ setting (i.e. changed lr to 1e-3).

Any suggestions or help would be appreciated.

Best, Sophia

[erikwijmans]

If I recall correctly, I was able to reproduce the results on ScanNet by setting the hyper-parameters to the same as in https://github.com/charlesq34/pointnet2/blob/master/scannet/train.py and adding the class balanced cross-entropy loss used there.

[daveredrum]

Hi there,

I'm facing the same problem of reproducing the performance in the PointNet++ paper. The best accuracy I had is 59%.

I use the code in https://github.com/charlesq34/pointnet2/blob/master/scannet/preprocessing/collect_scannet_scenes.py to preprocess the ScanNet data and the data loader in https://github.com/charlesq34/pointnet2/blob/master/scannet/scannet_dataset.py to fetch the data.

I'm wondering if anyone could share the code for training on ScanNet with this PointNet++ implementation? That'd be super helpful. :)

Best, Dave