Questions about init_from_known_poses

[hturki]

Thanks for the awesome work! I have a few questions about the init_from_known_poses method at

dust3r/dust3r/cloud_opt/init_im_poses.py

Line 24 in 01b2f1d

def init_from_known_poses(self, niter_PnP=10, min_conf_thr=3):

1. I don't understand the idea behind this statement:

   dust3r/dust3r/cloud_opt/init_im_poses.py

   Line 43 in 01b2f1d

   msk = self.conf_i[i_j] > min(min_conf_thr, self.conf_i[i_j].min() - 0.1)

Everything in self.conf_i[i_j] is larger than self.conf_i[i_j].min() - 0.1, so will this mask not always evaluate to true everywhere?

2. Also, when calculating these masks, since we are evaluating pred_j (ie: points in the second view in the reference frame of the first, would it not make more sense to use conf_j instead of conf_i?

3. (More of an comment than a question) The idea behind iterating between all edges, running opencv's pnp, and choosing the best depthmap based on the average confidence score as defined by

   dust3r/dust3r/cloud_opt/init_im_poses.py

   Line 54 in 01b2f1d

   score = float(self.conf_i[i_j].mean())

   seems a bit wasteful. Could one not first find the best pred_j for each image i as determined by average confidence, and then run pnp and get the scaled depthmaps only for those? That would reduce the number of opencv calls to pnp from num_edges to num_images