verlab
How does this method perform if training without Reliability Map?
Hello. Thank you for your nice work and contribution of field.
I have some questions.
- How does this method perform if training without Reliability Map?
- Since the method detected keypoints with simple network(only 4 1*1 conv), did you validate some metric about keypoint detection, e.g. Repeatability?
I look forward to your reply and thank you for your excellent work!
Hi @weich37, thank you for your interest in our work!
Q1: For semi-dense matching, you can match the full (dense) descriptor map without filtering based on reliability (no need to train with a reliability map). This should provide the densest matching, but the inlier ratio (#correct/#total) may decrease, and the matching cost will increase.
Q2: We did not check keypoint repeatability because it does not always correlate well with downstream task performance. Instead, we directly evaluated the keypoints/descriptors on downstream tasks such as camera pose, homography estimation, and visual localization.
