An error about rendering images from initial pointcloud

[shenglunch]

Thanks for your work! I want to bring it within my own reach, so I use a simple case to test it. However, I find an error when rendering images from the initial pointcloud. Solving it is difficult for me.

In my case, I generate the initial pointcloud from only one RGB image and the corresponding depth map captured by the D435 camera, and the pose is a 4x4 identity matrix. After calling self.gaussians.create_from_pcd() in scene/__init__.py, I add self.save(0) to save the initial gaussian model. Next, I run train.py and render.py to use the initial gaussian model to render an image at the same pose, a 4x4 identity matrix. There is an obvious black band at the top, and the result is as follows.

For another data, the black band is at the bottom.

I train the gaussian model for about 500 iterations. The black band disappears. Why does the initial gaussian model appear with the black band? Is it possible that the initial Gaussian model contains potential errors?

My data is here.
data.zip

[AsherJingkongChen]

It may be your test case too simple to build a volumetric model.