search

darglein / ADOP

MIT License
2.02k stars 196 forks source link

Image to Point Cloud Alignment Dataset and Experiments #9

Open bzvbzv opened 2 years ago

bzvbzv commented 2 years ago

Thank you for the great project. In the paper you mentioned the network has an application of Image to Point Cloud Alignment. Can you share the Dataset and Experiments of that part for evaluation?

PS: There is a small typo, In Table II, NBPG should be NPBG. In my opinion LiDaR usually write as lidar, LIDAR or LiDAR.

darglein commented 2 years ago

Hi, Thanks for the type corrections :)

I don't think I can release the dataset we used for image to point cloud alignment. The experiment were all default settings of ADOP with fix_poses = False in the config. With that we just trained it normally and exported the poses in the end.

A small side note is that for the image alignment to work well the initial pose guess must be decent and each 3D point must be seen by multiple (5+) cameras.

Gatsby23 commented 2 years ago

Dear Dr: Thank you for your reply. However, if I want to use the algorithm on the outdoor scene. What should I be careful about? I notice that the each 3D point must be seen by 5+ cameras, however ,the edge of the point cloud maybe can only be seen about 1 or 2 cameras. Do you think it will be a big problem from the rendering process?