Unconvergence on aleks-teapot when training the SC-GS

[guanjunwu]

Hi, thanks for your great work! I'm trying to use SC-GS to train the interp/aleks-teapot dataset. That's my command:

python train_gui.py --source_path data/hypernerf/interp/aleks-teapot --model_path outputs/hypernerf/interp/aleks-teapot --deform_type node --node_num 512 --hyper_dim 8 --eval --gt_alpha_mask_as_scene_mask --local_frame --resolution 2 --W 540 --H 960

It seems like training cannot convergence. I'm wondering whether the author are also faced with similar problem? As I concerned, it may caused by large motion and error camera poses?

[yihua7]

Hi,

The problem is unusual, and I haven't encountered this phenomenon before. Even if the motion is significant, the static parts should remain unaffected and can be reconstructed accurately. Is this dataset synthetic or real? It could possibly be an issue with the camera, such as inaccurate parameters.

[TCQian]

Hi @yihua7 , i realise the synthetic view generation here might be wrong. The focals and image are reusing the last image variable instance in the former for loop. Do you mind have a look on it?