scene context from different encoders

hi @chrisdxie, @skanti,

- Does current inference code uses semi dense pointcloud only encoder, this part? >

- does the sample semi dense pointcloud from dataset comes with positional embedding?
what does this really mean, `positional embedding`: we append the `coordinates of the active sites` to their respective feature vectors fi ← cat(fi, ci).?

- your experiments also contain scene context obtained from scene images as well as semi dense pointcloud,
does this image encoder code exist in current code base? this part>

- any plan on providing updated training code?

thank you very much.