Open Moreland-cas opened 11 months ago
Hi!
The PCK in pdcnet is on a different subset of megadepth and therefore not directly comparable.
As for the confidence estimate, yes it helps, but if you qualitatively look at the results you will see pdcnet has way more failure cases (not involving certainty).
If you want a network-to-network comparison of pdcnet and DKM/RoMa I suggest you use the training set and strategy of DKM with pdcnet architecture. I'm not sure exactly how big the gain would be, but if you look at some of our ablations I think we have some pretty big improvements.
The point of our papers are not to say that we are better than pdcnet. Rather we want to show that dense matching is competitive for two view geometry.
Ah, another point is that pdcnet does actually train on megadepth (just different strategy), so I think the conparison is pretty fair.
Thanks for your quick reply, I believe there are certainly some improvements to the architecture of DKM. There are two more questions that are bothering me, could you please help me out? 1) In the process of computing the fundamental matrix using OpenCV's cv2.findFundamentalMatrix, the algorithm typically uses a subset of points, such as 8-point correspondences, along with a RANSAC approach to filter out outliers. Given that the remaining points only contribute implicitly to the estimation of F, would it improve accuracy to re-estimate F using the inliers identified by the initial computation? 2) Is it possible to employ the dense warp in a differentiable manner to compute F, thereby enabling the training of all point correspondences directly?
Yes. Typically you would use a few LM optimization steps to improve. This is included in most opencv ransac, not sure about the most basic one. If using LO-RANSAC it would even be done for the scoring of the model using a local set.
Yes. There are works in this direction and Ive been meaning to try it out myself as well.
- Yes. Typically you would use a few LM optimization steps to improve. This is included in most opencv ransac, not sure about the most basic one. If using LO-RANSAC it would even be done for the scoring of the model using a local set.
- Yes. There are works in this direction and Ive been meaning to try it out myself as well.
Thanks again :)
Why does PDC-net perform reasonably well in pck but much worse in two-view geometry estimation? pdc-net pck:
Is it because the confidence it predicts is learned self-supervised? Also is it fair to compare DKM with pdc-net without retraining it fully supervised?
If I understand it wrong, please point it out, Thank you! : )