number of views used for training

[ankan2709]

Hi, thank you for the amazing work.

I have a question about how to change the number of views used for training a particular scene or object. thank you.

[airalcorn2]

Can you provide more information?

[ankan2709]

In the pixel nerf paper the authors talk about pixel nerf can be used for sparse inputs containing 1, 3 or 5 input views. I wanted to know if the code has an option to decide how many images or views to train for a particular scene?

For each object in the "car" folder there are 50 images. How many images per object or scene are used for training?

[airalcorn2]

My implementation only conditions on a single view, but it should be straightforward to extend the code to condition on multiple views using Section 4.2 and Figure 18 as a guide. Essentially, you independently run pixelNeRF partially for each input view (i.e., using the relative transformations for the target view relative to each input view), take the average of the representations obtained halfway through the MLP, and then pass that averaged representation through the final layers of the MLP.

[ankan2709]

Thanks for your inputs. I am going to try this out.