Why not train good with white regions

[HungNgoCT]

Hi,

I captured a video length 1 minutes, then train with neus-facto. One frame in the video like this

The toy is with white sword. I trained 100k steps. Then I got the rendering and the mesh model as the images below. It lost a part of the white sword. This issue is also occurred on my experiments for other objects with white regions while results were normal with other colors.

Can any one help explains this? And how to overcome?

Thank you

[niujinshuchong]

Hi, could you try with a large batch size of training rays?

[blacksino]

Could you share your command?

[HungNgoCT]

Could you share your command? Hi, Sorry for long time respone. This is command:

1. ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data --data path_to_my_data

2. ns-extract-mesh --load-config path_to_config_file --output-path path_to_output_path. From observation, it is also affected to other experiments.

However, I trained up to 300k steps, the result is better

My question is that: why are the white regions are converged lately?

[HungNgoCT]

Hi, could you try with a large batch size of training rays?

Hi @niujinshuchong ,

I have not yet tried with large batch size of training rays. However, when I tried with 300k steps (took 1/2 days), the result is better. I will try with large batch size of rays then let you know later. But my question is that why are the white regions converged lately? This also occurred on my other experiments.

[blacksino]

how do you remove backgroud floaters?

[HungNgoCT]

how do you remove backgroud floaters?

Hi @blacksino ,

I don't sure correct or not, I trained more steps to get less floaters. Also, adjust bias. Some floaters were manually removed. But there are several parameters I have not understood yet actually.

[hanjoonwon]

1. ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data

Hi your mesh quality is very good Currently I'm not getting the quality I want, can you share the training command?

[HungNgoCT]

1. ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data

Hi your mesh quality is very good Currently I'm not getting the quality I want, can you share the training command?

Hi @hanjoonwon ,

Your quote is my training command ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data

[hanjoonwon]

1. ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data

Hi your mesh quality is very good Currently I'm not getting the quality I want, can you share the training command?

Hi @hanjoonwon ,

Your quote is my training command ns-train neus-facto --pipeline.model.sdf-field.inside-outside False --pipeline.model.background-model mlp --trainer.max-num-iterations 100000 --vis viewer --pipeline.model.sdf-field.bias 0.3 --pipeline.model.num-neus-samples-per-ray 96 sdfstudio-data

@HungNgoCT Thank you for reply :) I guess my problem is with my imagesets, I can't get a good quality mesh.

Would it be possible for you to advise me on the train command to extract the maximum quality from my image dataset when you have the time? Below is the link to my images. I would greatly appreciate your assistance

https://drive.google.com/drive/folders/1Q9zSM8sCsQ4n-6QKl4GEujuznx-abxvh?usp=drive_link

When I execute the command below with neus, I still get some results, but when I change the method, it comes out with very bad quality. --pipeline.datamanager.train-num-rays-per-batch 2048 --pipeline.model.sdf-field.inside-outside False
--pipeline.model.sdf-field.hidden-dim 256 --pipeline.model.sdf-field.num-layers 2
--pipeline.model.sdf-field.num-layers-color 2 --pipeline.model.sdf-field.use-grid-feature False
--pipeline.model.sdf-field.bias 0.3 --pipeline.model.sdf-field.beta-init 0.3 --pipeline.model.background-model mlp
--pipeline.model.sdf-field.use-appearance-embedding True --trainer.steps-per-eval-image 5000
--trainer.max-num-iterations 200000 --pipeline.model.near-plane 0.05 --pipeline.model.far-plane 2. --vis wandb
--pipeline.model.overwrite-near-far-plane True --pipeline.model.mono-normal-loss-mult 0.01 --viewer.websocket-port 7008 --pipeline.model.mono-depth-loss-mult 0.0 \