How can I train the model in a new real world images dataset

[sincerely1]

Your work is awesome. I want to train your work in a images dataset. I don't konw which config I should use? I think I should use nv3d config. But this config is train for video dataset. I don't know the code modify the code. Could you give me some advice?Thank you!

[Caoang327]

Thanks for it. The central part of training on own dataset is to write a dataset class like neural_3D_dataset_NDC.py in dataloader folder, which could read "rays", "rgbs" and "times" data using __get_item__() functions. You may also want to specify the "scene_bbox_min/max" and "self.near_far" attributes, which would be useful for rendering.

After that, you can call this class with "get_train_dataset" function in init.py file for dataloader.

Rember to adjust the models' settings based on your own data and results.

Feel free to ask any information. Thanks

[Akash-Kumbar]

Thanks for it. The central part of training on own dataset is to write a dataset class like neural_3D_dataset_NDC.py in dataloader folder, which could read "rays", "rgbs" and "times" data using get_item() functions. You may also want to specify the "scene_bbox_min/max" and "self.near_far" attributes, which would be useful for rendering.

After that, you can call this class with "get_train_dataset" function in init.py file for dataloader.

Rember to adjust the models' settings based on your own data and results.

Feel free to ask any information. Thanks

Hey, can you explain more about the near, far, and scene bbox min, and max attributes and how can I set this according to my dataset? Is it related to camera poses or the c2w matrix? Please explain more about it.

thank you

[Caoang327]

Hi:

The scene box min/max is the x,y,z boundary, in which we model the current scene. The near, far is the range we sample points along the rays.