search

PruneTruong / DenseMatching

Dense matching library based on PyTorch
GNU Lesser General Public License v2.1
690 stars 82 forks source link

Why is the radius R equal to 1 in the confidence estimation? #3

Closed syqzdy closed 3 years ago

martin-danelljan commented 3 years ago

Hi. We used R=1 pixel for simplicity. If i remember correctly, larger theresholds gave similar results for eg Megadepth (@PruneTruong would know more). If you care more about robustness than small errors, a larger radius would be better.

PruneTruong commented 3 years ago

Yes, setting R=3 or R=5 gave similar results so we used R=1 everywhere for simplicity!