strange result for npm3d

[amiltonwong]

Hi, @aboulch ,

I had trained npm3d through 99 epochs, and got 91.64 iou. For testing, the visualization of the prediction outputs look fine. I attached ajaccio_2_predict.png, ajaccio_57_predict.png, dijon_9_predict.png here for reference. However, when I submitted the prediction output to NPM3D benchmark, I got low result mIOU = 30.7, link. It seems that there's strange problem in classifying on Trash can, Barrier, Pedestrian (only get IOU = 0.x). And the prediction on Car looks good in visualization but only got IOU = 19.5 in benchmark. Is there something conflicted with the original labeling during the preparation step?

THX!

[aboulch]

Hello, Label indices should be OK but I will look into that. Trash can, Barrier, Pedestrian are difficult classes. Please note that in the paper, results in the paper are obtained with a fusion process that I did not released yet for this dataset. It may explain the low results on these classes.

[Dario-Mantegazza]

Hello @aboulch, first thanks for the interesting paper and approach. I'm trying to use ConvPoint on a custom dataset that is similar to Semantic3D or npm3D and I would like to know if there are any news about the release of the fusion model. Thanks

[aboulch]

Hello, I have added pre-trained models on NPM3D and released the fusion code. The pre-trained models come with no guarantee as I changed affiliation and could not retreive all data from my previous company.