Is KPconv kinda implemented in this repo?

[shangguan9191]

Dear author,

I notice that part of accuracy result of Lightconv is aggragated from result of KPconv in training. So, if this part of result can be regard as the result of KP implementation?

[aboulch]

Hello, No in this repo, we only reimplemented the network architecture of KPConv for K-nearest neighbors conv. layers, not the convolutional layer described in KPConv.

[shangguan9191]

I am writing a paper comparing different algorithm. The Lightconv and Conv is the fastest algo among all models. I intent to compare KPconv as well. However, i am not a coder. Do you know how use implement also Kpconv for the original layer, in your code style. is it possible that you show some guidance and add the layer in your repo if there is not?

# Network
network:
    model: KPConvCls
    backend_conv: 
        layer: FKAConv(KPConv)
        kernel_separation: false
    backend_search: SearchQuantized

[aboulch]

Hello, You cannot do for now. I have coded only the spatial search using K-nearest neighbors (all search algorithm are based on KNN not radius). I do not know if I will have time to do so.