Concatenate different feature levels to one level before passing data to the heads

[samsunglabs]

Concatenate different feature levels to one level before passing data to the heads

[filaPro]

Hi @FabianSchuetze ,

Don't quite understand your question. All tensors in fcaf3d are sparse and 3-dimensional (W x L x H x C). You can somehow use pooling or strided convolutions to remove one of dimensions and make the shape W x H x C. But why would you do this? To work with BEV projection? To make dense tensor from sparse you can use .dense() method.

[FabianSchuetze]

Thanks, @filaPro for your kind and detailed response.

Indeed, you are right: I would like to try a prediction from a kind-of-BEV. Making the predictions on different level is very interesting, and out of curiosity I would like to do a CenterPoint-type prediction from a single layer as ablation. I thought of upsamling the spare tensors to a common resolution and then doing a max pooling over them. Do you have an idea how I could to that.

I tried to call the dense method on the out matrix but this fails because some coordinates are negative and I have to provide a min_coordinate argument to the dense function. Do you have an idea which arguments I could use?