A way to use the 3dgs Viewer

[m1ingshen]

I made some minor modifications to your code so that it can be used with the viewer provided by 3DGS. Here's my approach:

1. change the codes in function construct_list_of_attributes from scene/gaussian_model.py like this:

l = ['x', 'y', 'z', 'nx', 'ny', 'nz']
        # All channels except the 3 DC
        for i in range(self._features_dc.shape[1]*self._features_dc.shape[2]):
            l.append('f_dc_{}'.format(i))
        for i in range(self._features_rest.shape[1]*self._features_rest.shape[2]):
            l.append('f_rest_{}'.format(i))
        l.append('opacity')
        for i in range(self._scaling.shape[1] + 1):
            l.append('scale_{}'.format(i))
        ......

By this way, the scales can have 3 dim, the same as 3dgs.
2. change the codes in fuction save_ply from scene/gaussian_model.py like this:

mkdir_p(os.path.dirname(path))

        xyz = self._xyz.detach().cpu().numpy()
        normals = np.zeros_like(xyz)
        f_dc = self._features_dc.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy()
        f_rest = self._features_rest.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy()
        opacities = self._opacity.detach().cpu().numpy()
        scale = self._scaling.detach().cpu().numpy()
        rotation = self._rotation.detach().cpu().numpy()
        one_line = np.full((scale.shape[0],1), 0.0000000001)
        one_line = np.log(one_line)
        scale = np.concatenate((scale, one_line), axis=1)
        ......

By this way, the 2D gaussian is changed into 3D gaussian with a zeros_like dimension.

[hbb1]

Yes, have you observed severe perspective-distortion artifacts? Can you post some results here? By the way, we have made progress in creating a viewer.

[m1ingshen]

These are two examples of garden training 7_000 iterations


I have only tested a few of models, maybe there are some distortions I haven't met.

[hbb1]

Thank you for reporting the results! As I mentioned in the README, this is a possible but not perfect solution. Also, the additional scale will cause ambiguities. Therefore, I think it is better to keep the current Gaussian Model. Let's keep the issue open for who would like a quick visualizer. Or let me refer people to your fork if you have.