Tao Lu

'blue' denotes 'far' in the right two images, but denotes 'near' in the left image. Are there some misaligned rules?

Hi, the feat bank works well when you have many views from different directions and distance to train the MLP fully. We set it to False by default because most...

Thanks. The viewer is a big problem for us now. Although we provide a Scaffold-GS Viewer in https://github.com/city-super/Scaffold-GS/tree/main/SIBR_viewers, the incompatibility between ours and original 3D-GS brings a lot of inconvenience....

Thanks for the suggestion. We will consider it.

Hi, could you clarify if the 'worse results' are in comparison to Scaffold's performance on lower resolution images or against 3D-GS in the same higher resolution?

Did you try to train for longer iterations? Since more details are presented in a higher resolution image, Scaffold-GS may need more growing operations.

Thanks. We choose direct concatenation because 1) reducing the total computation; 2) we find that a naive PE does not always contribute to a better results. Further explorations are welcome.

According to your previous issue https://github.com/city-super/Scaffold-GS/issues/29#issue-2178339409, I guess that you pose is very poor, making the learned gaussians fail to generalize to novel views.

Two ways: 1. you can refer to https://github.com/city-super/Scaffold-GS/blob/da97ef8257b46d51c432df0df8b62f7c3a3c1079/scene/dataset_readers.py#L307-L319 to give a rand init during loading data; 2. replace https://github.com/city-super/Scaffold-GS/blob/da97ef8257b46d51c432df0df8b62f7c3a3c1079/scene/gaussian_model.py#L237 with a rand init.

To do so, you can replace https://github.com/city-super/Scaffold-GS/blob/da97ef8257b46d51c432df0df8b62f7c3a3c1079/scene/dataset_readers.py#L191-L194 with ``` num_pts = 100_000 print(f"Generating random point cloud ({num_pts})...") xyz = np.random.random((num_pts, 3)) * 2.6 - 1.3 # adjust 2.6 and 1.3...