princeton-vl/DPVO

randomly selecting patch centroids vs. 2D keypoint locations

Closed this issue · 2 comments

luopengting commented

The paper mentions that Randomly selecting patch centroids works better than 2D keypoint locations produced via SIFT [22], ORB [29], Superpoint [9], or pixels with high image gradient, I have two questions:

  1. Do you use the classical method to select feature points and resize the location to the feature map size as the patch center centroids?
  2. Is the model retrained when the classical keypoint methods are used?

Looking forward to your reply. Thx a lot!

lahavlipson commented
  1. Yes, exactly.
  2. No, these ablations were performed with the model weights trained using random patch centroids.
luopengting commented
  1. Yes, exactly.
  2. No, these ablations were performed with the model weights trained using random patch centroids.

In my opinion, maybe it will be fairer to retrain the models. This paper helps a lot, thank you very much! 😄