nakuramino / minimal-nerf Goto Github PK

We don't got no requirements.txt. Use this:

pip install pipreqs

pipreqs /path/to/project

I mean, how bomb would it be if I can render a new NeRF!! No Face here I come :333

From a cursory glance, it seems that an RGBA image is read as an RGB image. Seems problematic.

honestly, kind of stumped on this one, should I just plot everything using plotly?
reconstruct one image at a time and put them together to create a gif?
Not sure, but a few options I suppose

How about we implement a simple sampling scheme based on alpha values (from the original image) and use that to help training stabilize instead? Not research worthy but could be interesting enough for a class project. Then we can also do away with this stupid cropping bullshit i implemented too.

At first glance, it looks as if the inverse transform sampling does the right thing (it's more densely sampled towards the left:

But if you zoom in, you notice that it doesn't actually get sampled smoothly. The way I do it is currently clustered, which is not good:
.

To see it more clearly, red is the initial coarse samples:

When sampling for coarse and fine network, we sample coarse_ts and fine_ts. When concatenating the two, we have to sort the ts to be in the right order so that the calculated deltas make sense. However, sorting the ts can mess with the order of the density and rgbs. What to do what to do.

lol