A Problem About Using SLAM Data as Gaussian Input

[yuyunlong2002]

I want to convert the map file generated by openvslam(*.msg) into colmap sparse point cloud file format (points3D.bin, images.bin, cameras.bin) as input for Gaussian Splatting, but the result generated by Gaussian is very bad.The point cloud generated by Gaussian Plating using SLAM results is very messy, especially on the Z-axis, the height is disordered.
This is the visualization result of my conversion of openvslam result to colmap format in GUI:

And this is the visualization result of colmap sparse reconstruction:

I believe that there are no errors in the position relationship between the camera and the point cloud in the conversion results.(Maybe I made some mistake^_^). From the two images above, it can be seen that colmap sparse reconstruction generates more point clouds. However, after downsampling the sparse point clouds generated by colmap, the results are also better than those generated by openvslam, so it should be possible to rule out the reason of the number of point clouds

This is the point cloud generated by me using the openvslam results for Gaussian generation:


And this is the point cloud I reconstructed using colmap for Gaussian generation:


It can be seen that the Colmap processing result is much better for GS than the openvslam result. I found that the Colmap result has 63 keyframes, while the openvslam result only has 54 keyframes. I have ruled out some possible reasons for the difference in effect, including the number of point clouds, point cloud color (Colmap sparse reconstruction result point cloud contains RGB information, openvslam result point cloud does not contain RGB information), error information (error mentioned in the Colamp Points3D file header), and TRACK information (track mentioned in the Colamp Points3D file header)
Can someone tell me which step went wrong?