Some questions about the source view

[Lemon-XQ]

Hi, nice work! I have read your source code and have some questions...

1. It seems that your pretrained model was trained on 3 source views setting, so if I want to use more source views, I need to retrain the model from scratch. Is it right? Or can I simply modify some codes? Or only retrain some modules?
2. The training time of your pretrained model? I do not find it in your paper...
3. About the source view selection, in another issue, you mentioned that the source views should be very close. But in my understanding, the source views are used to initialize the volume, so why not use more sparse views (eg. front+back+left+right)? I think it will cover more seen regions?

Waiting for your answer. Thanks for your work again!

[apchenstu]

Hi thanks, 1) there are a few ways to relax the 3 input views setting, like the attention model here, you may try if you are interested in. 2) 6 epochs on DTU dataset with a single RTX 2080Ti takes about 30 hours. 3)large baseline is not friendly to geometric reasoning since cost volume-based methods require doing the feature matching stuffs.

[Lemon-XQ]

Hi thanks, 1) there are a few ways to relax the 3 input views setting, like the attention model here, you may try if you are interested in. 2) 6 epochs on DTU dataset with a single RTX 2080Ti takes about 30 hours. 3)large baseline is not friendly to geometric reasoning since cost volume-based methods require doing the feature matching stuffs.

Thanks for you answer! I still have 2 questions:

1. In your mentioned attention model, the param input_ch_views is the dimension of view direction, which should be 3, not the number of input views. Is it right?
2. I have tried to train a new multi-view model by using you default network(net_type='v0'), only modify some hard-coded parts (like here). Is it ok? Or your default network is only available for 3 views?