zzhang1987 / deep-graphical-feature-learning Goto Github PK

Hi @zzhang1987
This is great work. Thanks for sharing it.

I would like to know your comments on a few things:

1. The line gTruth[nIns:] = -100 looks wrong.
   I guess it should be gTruth[gTruth>=nIns] = -100. Could you confirm the same?

2. The paper/code tested only on matching scale-rotation-translation of the data but not perspective projections. May I know if such matching is tested on homography transformations, if so what are your observations. I would be glad if you could share any latest works in similar scenarios (maybe of others').

Thank you

Hi @zzhang1987 . Thanks for your sharing. I am doing the 2D image feature matching and 3D point cloud feature matching and I find your paper related and quite interesting. Here I have some questions regarding the paper.

Since you are using the PointNet as a feature extractor, have you tried some advanced network architecture designed to replace PointNet such as DGCNN, KPConv or DCP (which also deal with feature matching problem)? Actually I think most of the work on point cloud feature learning can be used in your case and they are also only using the point coordinate as input. Is there any specific reason why you are not comparing with these methods?

My second question is for experiment on PF-Pascal (sorry I am not familiar with this dataset so my understanding might be wrong), each image pair only have 6-18 manually annotated ground truth correspondence, then how large the k should be? In this situation, will the k-nn neighbor be too far away from the center point?

Thank you in advance!