Training 3DGS using parse_pcd.ply and transforms.json.

[CCtwelve]

Thank you for your work. I found that after using the camera parameters from transforms.json and converting the OpenCV coordinate system to OpenGL using the method provided, and then using parse_pcd.ply as the initial point cloud for GS training, the position of the person in the rendered image does not align with the image from the images path when viewed from certain angles.
Here is the translation:

"As shown in the figure:

• The first column is the rendered result of parse_pcd.ply.
• The second column is the images result.
• The third column is the alpha.
• The fourth column is the fmasks.
• The last column is the overlapping result of the alpha and fmasks."

[sipkode]

I think I'm suffering from this issue as well... I didn't bother rendering out a comparison like you did, but just tried to create a gaussian splat using the output data in Postshot, and got this result:

ballet_diffuman.mp4

Which looks like it's a result of misaligned frames. Were you able to make any progress on this>

[sipkode]

I was able to process some custom data without this issue, so it might be the case that the camera extrinsics provided in the dataset are incorrect, which would lead to misalignment...