DGCNN and PointNet++ add expected segmentation results to examples

[AruniRC]

❓ Questions & Help

Can you please mention the expected mean IoU over all ShapeNet segmentation categories using the example code you provide for DGCNN and PointNet segmentation?

The current code example does not provide any indication of how close the results from this code is with respect to the reported numbers in the two papers. It would be really very helpful to have these as a sanity-check when using your models and codebase as a starting point for any project.

thank you.

[rusty1s]

This is a great idea! In addition, we might provide pre-trained models for those models. I will re-run the experiments and add the expected results when I have time.

[AruniRC]

Hi @rusty1s, a heads-up -- I am trying to reproduce the results for PointNet++ for ShapeNet segmentation, as a first step. A quick recap:

• Right now, I am getting IoU=79.09 , compared to 81.9 per-class IoU reported in the PointNet++ paper (Supplementary: C1 Table 4).
• Main deltas w.r.t. your sample code: training for 200 epochs with learning-rate drops at 100, 150 and 175 epochs.
• Pytorch-geometric's dataloader for ShapeNet does not provide normals, which are used by most papers in addition to the XYZ points (this could be a significant impediment to reproducing numbers)

Will let you know on this thread if I can resolve the remaining "deltas" and get comparable numbers to the published results.

[rusty1s]

Thank you for your investigations! I think we should add normals to the ShapeNet dataset which should bring huge gains IMO.

[Tofull]

normals or any additional n-dimensions (like intensity for lidar point clouds)

[tchaton]

Dear @AruniRC @Tofull @kolia @fadel,

I started to build a proper benchmark: https://github.com/nicolas-chaulet/deeppointcloud-benchmarks

Please, have a look :)

It is still under heavy dev.

Best to All,