point_cloud_blocks lost

[CBoLee]

When I train in chunks, like in the picture, it appears that there is no point cloud in the cell. The dataset I'm running is building-pixsfm and the .yaml file is building_c20_r4.yaml

[DekuLiuTesla]

Hi, @CBoLee , please check if cell4 is assigned with too few images (<50 images), or if OOM or other problems occur in training.

[CBoLee]

Hi, @CBoLee , please check if cell4 is assigned with too few images (<50 images), or if OOM or other problems occur in training.

Hi, I would like to know, I am following the readme and using the processed dataset, if it's the allocation of too few images, then how should I adjust the configuration?

[DekuLiuTesla]

Hi, in our experience, several untuned blocks won't significantly influence results. If you want to avoid this case, you can use fewer blocks with longer iteration, or increase the num_threshold here so as to enlarge the block.

[CBoLee]

，Thank you for your outstanding contribution, I still have a couple of questions here, firstly I'm getting certain blocks not existing after merge, and secondly I'm wondering what is the difference between the rough global 3DGS prior and the rendering used for each block optimization? I look at the code and see that all that is used is the train_large.py script

[DekuLiuTesla]

Hi, @CBoLee, for the lost of blocks, you may need to check if the block is successfully trained and loaded in merging. You may also need to check if the block_id is correctly mapped to the x and y index, and if the generated mask is desired. Besides, the rough global 3DGS is the initialization of each block. Your observation is correct!