Using Iphone images as input data

[ForMyCat]

If I want to use RGB images taken from Iphone as input data (no depth info), should I copy datasets_depth_iphone.py to create the datasets_depth_xyz.py? Or should I just use the original datasets_depth_iphone.py and nvs_rgbd_iphone.gin?

[ForMyCat]

Or can you briefly explain the pipeline for data preping if I want to use only RGB images as input?

[Wanggcong]

Thank you for you question.
If you do not use depth info, you can directly use train_llff_dtu.py like llff dataset.

More precisely, I do not know the question "if I want to use only RGB images as input". In our paper, if you do not have depth maps from depth sensors, you can use depth estimation to predict depth maps, such as LLFF and DTU. But on LLFF and DTU datasets, they split datasets by selecting every 8 images as a test set.

Moreover, different scripts about datasets_depth_xxx.py are very similar. We separate the scripts because Kinnect has some noises in depth maps, and some uses inverse depth maps while some use disparity maps.

[Wanggcong]

Basically, you can prepare your data as follows:

1. Capture a few RGB images. Prepare your dataset like the iPhone dataset, for example, training images are named with '_train.jpg', test images are named with '_test.jpg'

2. Calculate camera poses by COLMAP

3. Calcuate inverse depth maps by the DPT depth estimator.

4. Organize datasets as the iPhone dataset (the depth map provided by iPhone are also inverse depth maps).

5. Set variables in nvs_rgbd_iphone.gin, and use datasets_depth_iphone.py and train_iphone.py

[Wanggcong]

Do you have any questions about this issue?

[Wanggcong]

I closed this issue since there is no further question. You might re-open the issue and open a new issue.