search

facebookresearch / ContrastiveSceneContexts

Code for CVPR 2021 oral paper "Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts"
MIT License
221 stars 28 forks source link

Train split for limited scene reconstructions #37

Closed YilmazKadir closed 2 years ago

YilmazKadir commented 2 years ago

Do you use the first 1%, 5%, 10%, 20% of the data when you fine-tune the network on ScanNet semantic segmentation for limited reconstruction case, or do you randomly shuffle the data first? If so, do you avoid splitting subscenes that are generated in the same scene to train and validation?

Sekunde commented 2 years ago

We split out 1%,5%,10%,20% from the official train split, which already splits sub-scenes from val. If I remember correctly, we took the first 1%, 5%, 10%, 20% of the data; but you can verify that by checking e.g., if 1% scenes are all included in 5% scenes.

YilmazKadir commented 2 years ago

Thanks for quick reply. I think I have found the scenes that you used at ScanNet Data Efficient web page: http://kaldir.vc.in.tum.de/scannet_benchmark/data_efficient/documentation#submission-policy For those who also wonders here is the zip file that contains scene IDs: scenes.zip