search

magicleap / SuperGluePretrainedNetwork

SuperGlue: Learning Feature Matching with Graph Neural Networks (CVPR 2020, Oral)
Other
3.31k stars 672 forks source link

indoor dataset #62

Closed bixiaopeng0 closed 3 years ago

bixiaopeng0 commented 3 years ago

Hi, In scannet dataset，How many images did you use for indoor training？do you use scannet_frames_25k or all scannet imgs?

sarlinpe commented 3 years ago

We use all ScanNet training images. From the paper:

Discarding pairs with too small or too large overlap, we select 230M training and 1500 test pairs.

bixiaopeng0 commented 3 years ago

thanks for your reply! do you use homographic method in one image to produce the gd matches? or use the different image pairs to produce the gd matches ?

sarlinpe commented 3 years ago

For ScanNet, we use pairs of images captured from different viewpoints. The matches are computed using the depth maps. Please close this issue if your problem is solved. Thank you!