Repository navigation
An error occurred when using init_geo for fixed pose in the LLFF data set (three-view sparse perspective condition) #70
Description
Hello, I tried to fix the existing colmap camera internal and external parameters on the llff data set, and used mast3r to infer the point cloud. I encountered the following situation in the room scene. There are no problems with the fortress, horns, and orchids scenes of llff. They can be aligned with the point cloud obtained by colmap. However, there are inaccuracies in several other scenes. Such an obvious error occurred in the room scene.
Activity
chenkangjie1123 commentedon Feb 27, 2025
loss = scene.compute_global_alignment(init="mst", niter=300, schedule=schedule, lr=lr)
# loss = scene.compute_global_alignment(init="known_poses", niter=300, schedule=schedule, lr=lr)
hello, i find the "known_poses" pattern can make this error, but the "mst" pattern doesnt..
chenkangjie1123 commentedon Feb 27, 2025
Hello, I found that the mst mode and the known pose mode have different reconstruction effects for different scenes (three-view sparse perspective condition). For example, the point clouds inferred using the mst mode in the orchids scene in the LLFF data set are overlapping, and it is normal to use the known_pose mode. It is OK to use the mst model for inference in the room scene. If you use the known_pose mode for inference, the disordered result above will appear.
[-]An error occurred when using init_geo for fixed pose in the LLFF data set[/-][+]An error occurred when using init_geo for fixed pose in the LLFF data set (three-view sparse perspective condition)[/+]zhanghaoyu816 commentedon Mar 19, 2025
Hi, Have you identified the cause of this issue? I want to know how to obtain dust3r point clouds that are aligned with known poses. From your experiments above, it appears that fixing camera poses may fail in some scenarios.