Why indoor and outdoor data be trained separately?

magicleap / SuperGluePretrainedNetwork

SuperGlue: Learning Feature Matching with Graph Neural Networks (CVPR 2020, Oral)

Other

3.31k stars 672 forks source link

Why indoor and outdoor data be trained separately? #119

Open HLH13297997663 opened 2 years ago

HLH13297997663 commented 2 years ago

Hello, thank you for your art work. I have a question about the training data, why indoor and outdoor data be trained separately rather than together? Will it be more accurate to train a model with indoor and outdoor data together？

sarlinpe commented 1 year ago

The model trained on MegaDepth (outdoor) works well for indoor data too. We just wanted to specialize a model for ScanNet to push the accuracy.