search

naver / dust3r

DUSt3R: Geometric 3D Vision Made Easy
https://dust3r.europe.naverlabs.com/
Other
4.67k stars 519 forks source link

Q about Absolute pose estimation #136

Closed TwiceMao closed 2 days ago

TwiceMao commented 3 days ago

Thx for your awesome work! @yocabon @beanmilk @seyoung-hyun image

TwiceMao commented 3 days ago

Additionally, I have questions about "Global Optimization" and "Recovering Camera Parameters."

  1. In Equation (5) of the main paper, how is $\chi^v_i$ obtained?

I speculate that it involves first estimating the intrinsics for each image, then using the z-coordinates of the point map corresponding to each RGB image in the camera coordinate system as the depth. And the depth is utilized in the below equation as $D_{i,j}^{n}$

image

  1. Could you please provide the initial values for $P_e, \sigma_e$ and the scale factor?

  2. Could you please explain how you specifically perform global camera parameter optimization? Do you use Equation (5) from your main paper as the loss function and optimize it using the Adam optimizer?

yocabon commented 2 days ago

Hi, First question about visual localization, $I_B$ is the database images, so its pose, intrinsics and 2D-3D correspondences are known (part of the SfM map).

Second question, did you check the code of the global optimizer ? initialization: https://github.com/naver/dust3r/blob/main/dust3r/cloud_opt/init_im_poses.py#L123 optimization: https://github.com/naver/dust3r/blob/main/dust3r/cloud_opt/base_opt.py and https://github.com/naver/dust3r/blob/main/dust3r/cloud_opt/optimizer.py

TwiceMao commented 2 days ago

Thank you for your kind response~