search

meidachen / STPLS3D

🔥 Synthetic and real-world 2d/3d dataset for semantic and instance segmentation (BMVC 2022 Oral)
236 stars 20 forks source link

train / validation split #4

Closed ywyue closed 2 years ago

ywyue commented 2 years ago

Hi Thanks for the great work! For instance segmentation, I downloaded the training point clouds (1-25). Are we allowed to use all the data to train the model, or we must split it to training (all - 5, 10, 15, 20, 25) and validation set (5, 10, 15, 20, 25)?

meidachen commented 2 years ago

I appreciate your interest in our data. I would assume you are working on instance segmentation instead of semantic segmentation since you downloaded the 25 datasets. The result we reported in the paper was training (all - 5, 10, 15, 20, 25) and testing set (5, 10, 15, 20, 25). If you would like to compare with HAIS and PointGroup from our paper, it is recommended to do that. In addition, we are also providing an evaluation server on Codalab (https://codalab.lisn.upsaclay.fr/competitions/4646) and organizing an ECCV workshop (https://urban3dchallenge.github.io/), where we have an additional 3 datasets without releasing the label for competition (in that case, you can train the model with all 25 data sets as you like or use some of them as the validation sets). I hope this information helps.

ywyue commented 2 years ago

Thanks! It's very clear now.