Feature Pyramid implementation

[aniket-gupta1]

Hi, Thank you for your amazing work!

I am trying to get features use PtV3 to generate feature pyramids for a pointcloud. So for example, if I have a network with 3 encoder and 3 decoder layers, I want to do something like this:

Can you please give some pointers on how can I do that?

[Gofinge]

Sorry for the late response. You can record a copy of the feature from the decoder during unspooling to prevent it from being decoded by the incoming decoder and then ensure it is also upsampled during the next unspooling layers.

Modifying the unpooling is enough to achieve this.

[aniket-gupta1]

But even after cloning, there is still the issue that the "pooling_parents" are deleted after every unpooling step. So even if I cloned the output of dec3 and used a separate unpooling layer, It will only increase the number of points by one step.

Basically, if N is the original number of points and Enc1 downsamples to N', Enc2 -> N'', Enc3 -> N'''. The output of Dec3 -> N'', so unpooling here will give me only N' points which can't be directly added to the output Dec1 which has N points

[Gofinge]

Hi, you can use "pooling_inverse" to map these embeddings to the original shape.