Bad result with custom camera poses

[otakudj]

What a great work! Its very convenient & efficient for prototyping. After test the lego for nerf, I caputure some images from a scene in a public UE project. Following the colmap, I get a quiet well result.

Then, I think if I can use the poses derived from UE? If ture, I can save the large part of the time used by colmap. However, the result is bad.

I follow the tips and try changing the factors including aabb_scale, scale, offset, I also scale the UE world coordinate to close the colmap2nerf.py result

But still can't get a comparable result with the quiet well result before. Here are the results of same view

ground truth	colmap pose	ue pose

Is there some critical problems in the images or the poses derived from 3D engine like UE?