Hi, great question! All graph convolution operations can be applied on any graph. After the graph convolution layers, we apply a global max pooling (see github.com/SoccerNet/sn-spotting/blob/main/Benchmarks/CALF_Calibration_GCN/src/model.py#L473) to aggregate the feature representation of each node into a single vector representation. This operation is independent on the number of nodes or edges and puts everything back to the same sized vector for the remaining of the network. I hope this helps.
In case of using GCN, it's not clear in the paper how would you handle the dynamic graph (Edges not being constant throughout the video chunk)
Hi, great question! All graph convolution operations can be applied on any graph. After the graph convolution layers, we apply a global max pooling (see github.com/SoccerNet/sn-spotting/blob/main/Benchmarks/CALF_Calibration_GCN/src/model.py#L473) to aggregate the feature representation of each node into a single vector representation. This operation is independent on the number of nodes or edges and puts everything back to the same sized vector for the remaining of the network. I hope this helps.