self.num_points suddenly decreased to 0 when training on my own dataset.

[xuxumiao777]

Hi, author!
Thanks for your fantastic work. I have encountered a problem when training on my own datasets. I have created tsv file for my own dataset. But the training process was terminated because self.num_points and num_splits suddenly decreased to zero(yes, I have monitered this parameter during the whole process). All datasets creashed at about 3000 iterations.

hope for your answer.

[KevinXu02]

Can you please try using splatfacto-w-light? That method should be using nerfstudio dataparser and might work for such datasets.
splatfacto-w was basically designed for phototourism datasets.

[xuxumiao777]

Can you please try using splatfacto-w-light? That method should be using nerfstudio dataparser and might work for such datasets. splatfacto-w was basically designed for phototourism datasets.

I have tried splatfacto-w-light, but it needs transforms.json. I have used colmap2nerf.py from nerf to generate a transforms.json. Honestly, it can run fluently without any error, but the result is a mess. I guess the coordinate system is not correct. Could you please provide a script, so that i can generate a correct transforms.json from colmap.
Thanks.

[KevinXu02]

ns-train splatfacto-w-light [OPTIONS] --data [PATH] [dataparser] colmap You can use colmap dataparser this way. And I think if you add --help in the end it will show the usage of colmap datapraser from nerfstudio.

[xuxumiao777]

ns-train splatfacto-w-light [OPTIONS] --data [PATH] [dataparser] colmap You can use colmap dataparser this way. And I think if you add --help in the end it will show the usage of colmap datapraser from nerfstudio.

tried_but_failed😂, here is my command:
ns-train splatfacto-w --data /gpfs/dataset/jingxigudao_tiankong/VID_20240804_111824/undistorted_colmap/dense --pipeline.datamanager.eval-image-indices None --viewer.make-share-url True colmap

[KevinXu02]

Please try splatfacto-w-light.

[xuxumiao777]

Please try splatfacto-w-light.

Sorry for that, I have now used splatfacto-w-light. It can run smoothly. But I can only see a mess in the viewer, which is different from the result by command splatfacto-w.

[KevinXu02]

Please consider enabling these options.

If you want to train datasets without nerf-w's train/test split or your own datasets, we provided a light-weight version of the method for general cases. To train with it, you can run the following command

ns-train splatfacto-w-light [OPTIONS] --data [PATH] [dataparser]
You can add these options to enhance the training process:

--pipeline.model.enable_bg_model True

Enables background modeling
Useful for scenes with distinct foreground and background
--pipeline.model.enable_alpha_loss True

Enables alpha loss for punishing gaussians from occupying sky areas.
Should be used with background modeling
--pipeline.model.enable_robust_mask True

Improves handling of transient objects (things that appear/disappear between views)

[xuxumiao777]

Please consider enabling these options.

If you want to train datasets without nerf-w's train/test split or your own datasets, we provided a light-weight version of the method for general cases. To train with it, you can run the following command

ns-train splatfacto-w-light [OPTIONS] --data [PATH] [dataparser]
You can add these options to enhance the training process:

--pipeline.model.enable_bg_model True

Enables background modeling
Useful for scenes with distinct foreground and background
--pipeline.model.enable_alpha_loss True

Enables alpha loss for punishing gaussians from occupying sky areas.
Should be used with background modeling
--pipeline.model.enable_robust_mask True

Improves handling of transient objects (things that appear/disappear between views)

use the following command and still a mess
ns-train splatfacto-w-light --data /gpfs/dataset/jingxigudao_tiankong/VID_20240804_111824/undistorted_colmap/dense --pipeline.datamanager.eval-image-indices None --viewer.make-share-url True --pipeline.model.enable_bg_model True --pipeline.model.enable_alpha_loss True --pipeline.model.enable_robust_mask True colmap
strange

[EliasDeSmijter]

Hi! Did you manage to find a solution? I am training splatfacto-w-light on a custom dataset (with a transforms.json) as well and have similar messy outputs when rendering images... Even when I train it on the nerfstudio poster dataset with
ns-train splatfacto-w-light --pipeline.model.enable_robust_mask True --pipeline.model.enable_bg_model True --pipeline.model.enable_alpha_loss True --data nerfstudio/poster --pipeline.model.random-init True I get a messy output

[KevinXu02]

Hi! Could you please try turn off these loss since they are mostly designed for outdoor scene. And splatfactow relies heavily on the sfm init. So maybe try ns-train splatfacto-w-light --data nerfstudio/poster and see if it works.

Hi! Did you manage to find a solution? I am training splatfacto-w-light on a custom dataset (with a transforms.json) as well and have similar messy outputs when rendering images... Even when I train it on the nerfstudio poster dataset with ns-train splatfacto-w-light --pipeline.model.enable_robust_mask True --pipeline.model.enable_bg_model True --pipeline.model.enable_alpha_loss True --data nerfstudio/poster --pipeline.model.random-init True I get a messy output

[EliasDeSmijter]

Hi, unfortunately this does not fix it.. Is the sfm init the most important aspect? When I train splatfacto-w-light with random-init on brandenburg the scene seems to converge which suggests it is not as important.